acme: answer HTTP-01 from the server itself, on port 80

HTTP-01 always arrives on port 80 - the CA chooses the port, not the operator -
so it never collides with an admin UI on 443. The conflict only exists for
TLS-ALPN-01, which is the challenge type that does use 443.

Given that, the server keeps a permanent listener on 80 that answers challenges
from a webroot and redirects everything else to the admin UI. Same arrangement
as the webroot plugins for Apache and nginx, and better than letting the ACME
client bind 80 per renewal: nothing binds and unbinds, so a renewal cannot fail
because the port was briefly busy, and the client needs only write access to a
directory instead of the privilege to bind a low port. Port 80 also gets a use
it would want anyway.

The ACME client stays an external program. lego is also a Go library, but
importing it would put a large dependency tree into a server that deliberately
has none, and the CLI does the same job from a timer.

Tokens are validated by *shape* before any filesystem call, so traversal never
reaches the disk - a stronger guarantee than sanitising a path and trusting the
sanitiser.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

server/Dockerfile

@@ -22,4 +22,7 @@ VOLUME ["/state"]
 # the data plane must see real client source addresses/TTLs, and Docker's
 # userland NAT would falsify exactly what this server exists to observe.
 EXPOSE 8441/tcp 8442/udp 8443/tcp
+# The verb is explicit here too, so `docker run <image>` serves and `docker run <image> --help`
+# still works by overriding the command.
 ENTRYPOINT ["/echolot-server"]
+CMD ["--serve"]

server/README.md

@@ -146,3 +146,41 @@ CI (`.gitea/workflows/build-server.yml`): tests on every push touching `server/`
 tagging `server-v1.2.3` builds + pushes the container image to the Gitea registry and
 attaches static linux amd64/arm64 binaries (+ SHA256SUMS) to a release — the same
 artifacts `--self-update` consumes.
+
+## TLS for the admin UI
+
+The binary terminates TLS itself; there is no reverse proxy in the design. It already serves TLS
+for the control plane, so this is reuse rather than new machinery, and it keeps the "one process,
+one config file" property. A proxy would also invite someone to eventually front the control plane
+too — which would break SPKI pinning, because clients pin *that* certificate's key.
+
+```
+ECHOLOT_ADMIN_LISTEN=[2001:db8::2]:443
+ECHOLOT_ADMIN_TLS_CERT=/etc/echolot/admin.pem
+ECHOLOT_ADMIN_TLS_KEY=/etc/echolot/admin.key
+ECHOLOT_ADMIN_BASE_URL=https://admin.example.net
+```
+
+Certificates come from any ACME client. **DNS-01 is the one to use here**: it needs no inbound
+port 80, which matters on a host where 80 is awkward or already spoken for.
+
+```sh
+acme.sh --issue --dns dns_cf -d admin.example.net \
+  --key-file       /etc/echolot/admin.key \
+  --fullchain-file /etc/echolot/admin.pem
+```
+
+**No reload hook is needed.** The certificate is re-read when the files change, so a renewal that
+drops new files in place is picked up on the next handshake. That is deliberate: a reload hook is
+the part of a renewal setup that quietly stops working, months later, and is noticed only once the
+certificate has already expired. A torn write — renewal tools write cert and key separately — keeps
+the previous certificate rather than failing the listener.
+
+Serving the admin UI in plaintext on a non-loopback address is refused: the session cookie is a
+bearer credential for everything the server can do, and the OIDC authorization code arrives in a
+URL. Bind to loopback and use an SSH tunnel (`ssh -L 8444:localhost:8444 host`), supply a
+certificate, or set `ECHOLOT_ADMIN_INSECURE=1` if you mean it.
+
+The control-plane certificate is deliberately *not* hot-reloaded. Clients pin its public key, so
+replacing it is a rotation an operator should have to think about, not something that happens
+because a file changed.

server/cmd/echolot-server/main.go

@@ -11,6 +11,7 @@
 package main
 
 import (
+	"bufio"
 	"context"
 	"crypto/ecdsa"
 	"crypto/elliptic"
@@ -36,7 +37,10 @@ import (
 	"syscall"
 	"time"
 
+	"echo-lot.app/server/internal/acmehttp"
+	"echo-lot.app/server/internal/adminauth"
 	"echo-lot.app/server/internal/canarydns"
+	"echo-lot.app/server/internal/certreload"
 	"echo-lot.app/server/internal/compat"
 	"echo-lot.app/server/internal/config"
 	"echo-lot.app/server/internal/control"
@@ -70,6 +74,30 @@ func run() error {
 	control.Version = Version
 
 	switch {
+	case actions.Help:
+		// Compatibility shim for one release.
+		//
+		// Serving became an explicit verb, but self-update is run by the *old* binary — so the
+		// repair added to the updater cannot fix the very update that installs the new one. A
+		// unit written before this change starts us with no arguments, and without this branch
+		// the service would simply stop working, unattended, on a host nobody is watching.
+		//
+		// Only when systemd started us: INVOCATION_ID is set by systemd for every service
+		// invocation and by nothing else, so a person at a terminal still gets usage. Remove
+		// this once no deployment predates --serve.
+		if os.Getenv("INVOCATION_ID") != "" {
+			slog.Warn("started by systemd with no verb — this unit predates --serve; " +
+				"repairing it and serving anyway")
+			if repaired, err := system.RepairExecStart(); err != nil {
+				slog.Error("could not repair the unit; fix ExecStart by hand", "err", err)
+			} else if repaired {
+				slog.Info("systemd unit updated to pass --serve")
+			}
+			return serve(cfg)
+		}
+		config.Usage(os.Stderr)
+		os.Exit(2)
+		return nil
 	case actions.Version:
 		fmt.Println(Version)
 		return nil
@@ -80,6 +108,8 @@ func run() error {
 		return system.InstallSystemd(cfg.SelfUpdateAPI)
 	case actions.UninstallSystemd:
 		return system.UninstallSystemd()
+	case actions.SetAdminPassword:
+		return setAdminPassword(cfg)
 	case actions.SelfUpdate:
 		return selfupdate.Run(cfg.SelfUpdateAPI, Version)
 	}
@@ -155,14 +185,16 @@ func serve(cfg *config.Config) error {
 	// uploads=account can never be satisfied — which is the honest outcome, not a silent
 	// downgrade to anonymous.
 	var idp *oidc.Verifier
-	if cfg.OIDCIssuer != "" && cfg.OIDCClientID != "" {
+	if cfg.OIDCIssuer != "" && (cfg.OIDCClientID != "" || cfg.OIDCAppClientID != "") {
 		idp = oidc.New(oidc.Config{
 			Issuer:      cfg.OIDCIssuer,
 			ClientID:    cfg.OIDCClientID,
+			AppClientID: cfg.OIDCAppClientID,
 			AdminGroup:  cfg.OIDCAdminGroup,
 		}, nil)
 		slog.Info("identity provider configured", "issuer", cfg.OIDCIssuer,
-			"client_id", cfg.OIDCClientID, "admin_group", cfg.OIDCAdminGroup)
+			"admin_client_id", cfg.OIDCClientID, "app_client_id", cfg.OIDCAppClientID,
+			"admin_group", cfg.OIDCAdminGroup)
 		if cfg.OIDCAdminGroup == "" {
 			slog.Warn("no admin group set: nobody will be an admin via OIDC " +
 				"(set ECHOLOT_OIDC_ADMIN_GROUP)")
@@ -284,7 +316,47 @@ func serve(cfg *config.Config) error {
 		})
 	})
 	adminSrv := &http.Server{Addr: cfg.AdminListen, Handler: admin, ReadHeaderTimeout: 10 * time.Second}
+	if cfg.AdminTLSCert != "" {
+		// Terminated here rather than behind a reverse proxy: this binary already serves TLS for
+		// the control plane, so it is reuse rather than new machinery, and one process with one
+		// config file is the property that makes this pleasant to run. A proxy would also invite
+		// someone to later front the control plane too, which would break SPKI pinning.
+		reloader, err := certreload.New(cfg.AdminTLSCert, cfg.AdminTLSKey)
+		if err != nil {
+			return fmt.Errorf("admin TLS: %w", err)
+		}
+		adminSrv.TLSConfig = reloader.TLSConfig()
+		if exp := reloader.NotAfter(); !exp.IsZero() {
+			slog.Info("admin UI TLS", "listen", cfg.AdminListen, "cert_expires", exp.Format(time.RFC3339))
+			if time.Until(exp) < 14*24*time.Hour {
+				slog.Warn("admin certificate expires soon", "expires", exp.Format(time.RFC3339))
+			}
+		}
+		go func() { errCh <- fmt.Errorf("admin: %w", adminSrv.ListenAndServeTLS("", "")) }()
+	} else {
 		go func() { errCh <- fmt.Errorf("admin: %w", adminSrv.ListenAndServe()) }()
+	}
+
+	// ACME HTTP-01 responder. Permanent rather than started per renewal: nothing binds and
+	// unbinds, so a renewal cannot fail because the port was briefly busy, and the ACME client
+	// needs only write access to a directory instead of the privilege to bind a low port.
+	if cfg.ACMEHTTPListen != "" {
+		webroot := cfg.ACMEWebroot
+		if webroot == "" {
+			webroot = filepath.Join(cfg.StateDir, "acme")
+		}
+		if err := acmehttp.EnsureWebroot(webroot); err != nil {
+			return fmt.Errorf("acme webroot: %w", err)
+		}
+		acmeSrv := &http.Server{
+			Addr:              cfg.ACMEHTTPListen,
+			Handler:           acmehttp.Handler(webroot, cfg.AdminBaseURL),
+			ReadHeaderTimeout: 10 * time.Second,
+		}
+		slog.Info("acme http-01 responder", "listen", cfg.ACMEHTTPListen, "webroot", webroot,
+			"redirects_to", cfg.AdminBaseURL)
+		go func() { errCh <- fmt.Errorf("acme-http: %w", acmeSrv.ListenAndServe()) }()
+	}
 
 	// UDP data plane — one socket per configured address. Distinct sockets
 	// (not wildcard) also guarantee responses leave from the address the
@@ -516,3 +588,48 @@ func publicControlURL(cfg *config.Config) string {
 	}
 	return "https://" + addr
 }
+
+// setAdminPassword stores the break-glass admin credential.
+//
+// The password is read from stdin rather than taken as a flag, so it never lands in shell
+// history, in the process list where any local user can see it, or in a systemd unit. Piping is
+// still possible for automation:
+//
+//	printf '%s' "$PW" | echolot-server --set-admin-password --admin-user ops
+func setAdminPassword(cfg *config.Config) error {
+	st, err := store.Open(cfg.StateDir)
+	if err != nil {
+		return fmt.Errorf("state store: %w", err)
+	}
+	fmt.Fprintf(os.Stderr, "New password for %q (input is not echoed if this is a terminal): ", cfg.AdminUser)
+	pw, err := readSecret()
+	if err != nil {
+		return err
+	}
+	fmt.Fprintln(os.Stderr)
+
+	cred, err := adminauth.NewCredential(cfg.AdminUser, pw)
+	if err != nil {
+		return err
+	}
+	if err := st.SetLocalAdmin(cred); err != nil {
+		return err
+	}
+	fmt.Fprintf(os.Stderr, "Break-glass admin %q set. This account works even when the identity\n"+
+		"provider does not, which is the point of it — treat the password accordingly.\n", cfg.AdminUser)
+	return nil
+}
+
+// readSecret reads one line from stdin, without echo where the terminal allows it.
+func readSecret() (string, error) {
+	restore, _ := system.DisableEcho(os.Stdin)
+	if restore != nil {
+		defer restore()
+	}
+	r := bufio.NewReader(os.Stdin)
+	line, err := r.ReadString('\n')
+	if err != nil && line == "" {
+		return "", err
+	}
+	return strings.TrimSpace(line), nil
+}

server/internal/acmehttp/acmehttp.go

@@ -0,0 +1,100 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+// Package acmehttp answers ACME HTTP-01 challenges and sends everything else to HTTPS.
+//
+// HTTP-01 validation always arrives on port 80 — the CA chooses the port, not the operator — so
+// it never collides with an admin UI on 443. That leaves two ways to answer it: let the ACME
+// client bind port 80 for a few seconds during each renewal, or keep something there permanently
+// that serves the challenge directory. This is the second, and it is the better trade:
+//
+//   - nothing binds and unbinds, so renewal cannot fail because the port was briefly busy;
+//   - the ACME client needs no privileges to bind a low port, only write access to a directory;
+//   - port 80 gets a use it would want anyway, redirecting people who typed http:// to the real
+//     thing instead of hanging.
+//
+// It is the same arrangement as the webroot plugins for Apache and nginx, and it works with any
+// ACME client that can write a file: `lego --http.webroot`, `certbot --webroot`, `acme.sh -w`.
+//
+// The ACME client stays an external program on purpose. lego is also a Go library, but importing
+// it would put a large dependency tree into a server that deliberately has none — and the CLI does
+// the same job from a timer.
+package acmehttp
+
+import (
+	"log/slog"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// ChallengePath is the fixed prefix ACME uses. It is not configurable, by the specification.
+const ChallengePath = "/.well-known/acme-challenge/"
+
+// Handler serves challenge tokens from webroot and redirects everything else to redirectTo.
+//
+// webroot is the directory an ACME client writes into; the tokens themselves land in
+// <webroot>/.well-known/acme-challenge/<token>, which is exactly what --http.webroot expects.
+func Handler(webroot, redirectTo string) http.Handler {
+	mux := http.NewServeMux()
+
+	mux.HandleFunc(ChallengePath, func(w http.ResponseWriter, r *http.Request) {
+		token := strings.TrimPrefix(r.URL.Path, ChallengePath)
+		// Tokens are base64url from the CA. Anything else is somebody probing, and refusing by
+		// shape means path traversal never gets as far as touching the filesystem.
+		if token == "" || !validToken(token) {
+			http.NotFound(w, r)
+			return
+		}
+		body, err := os.ReadFile(filepath.Join(webroot, filepath.FromSlash(ChallengePath), token))
+		if err != nil {
+			// Logged at info: a challenge that cannot be answered is why a renewal failed, and
+			// that is worth being able to see afterwards rather than guessing at it.
+			slog.Info("acme challenge not found", "token", token, "webroot", webroot)
+			http.NotFound(w, r)
+			return
+		}
+		slog.Info("answered acme challenge", "token", token, "from", r.RemoteAddr)
+		w.Header().Set("Content-Type", "text/plain")
+		_, _ = w.Write(body)
+	})
+
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		if redirectTo == "" {
+			http.Error(w, "this port serves ACME challenges only", http.StatusNotFound)
+			return
+		}
+		// 308 rather than 302: the method must not change, and the redirect is permanent in the
+		// sense that matters — this port will never serve the application.
+		http.Redirect(w, r, strings.TrimRight(redirectTo, "/")+r.URL.RequestURI(), http.StatusPermanentRedirect)
+	})
+
+	return mux
+}
+
+// validToken accepts only the base64url alphabet the ACME spec uses for tokens.
+//
+// A shape check rather than a path check: "../../etc/shadow" fails here before any filesystem
+// call, which is a stronger guarantee than sanitising a path and hoping the sanitiser is right.
+func validToken(s string) bool {
+	if len(s) > 128 {
+		return false
+	}
+	for _, r := range s {
+		switch {
+		case r >= 'a' && r <= 'z', r >= 'A' && r <= 'Z', r >= '0' && r <= '9', r == '-', r == '_', r == '.':
+		default:
+			return false
+		}
+	}
+	// A bare "." or ".." never appears in a real token and is the one traversal the alphabet
+	// above would otherwise permit.
+	return s != "." && s != ".."
+}
+
+// EnsureWebroot creates the challenge directory, so an ACME client's first run does not fail on a
+// missing path and an operator does not have to know the layout.
+func EnsureWebroot(webroot string) error {
+	return os.MkdirAll(filepath.Join(webroot, filepath.FromSlash(ChallengePath)), 0o755)
+}

server/internal/acmehttp/acmehttp_test.go

@@ -0,0 +1,98 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+package acmehttp
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func serve(t *testing.T, redirectTo string) (http.Handler, string) {
+	t.Helper()
+	root := t.TempDir()
+	if err := EnsureWebroot(root); err != nil {
+		t.Fatal(err)
+	}
+	return Handler(root, redirectTo), root
+}
+
+func TestServesAChallengeTokenWrittenByAnAcmeClient(t *testing.T) {
+	h, root := serve(t, "https://admin.example.net")
+	// Exactly what `lego --http.webroot` writes.
+	token := "abc-123_XYZ"
+	want := "abc-123_XYZ.keyauthorization-part"
+	if err := os.WriteFile(filepath.Join(root, ".well-known", "acme-challenge", token), []byte(want), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, httptest.NewRequest("GET", ChallengePath+token, nil))
+	if rec.Code != 200 {
+		t.Fatalf("challenge not served: %d", rec.Code)
+	}
+	if rec.Body.String() != want {
+		t.Fatalf("body = %q, want %q", rec.Body.String(), want)
+	}
+}
+
+// The token comes from the network and is used to build a path. Rejecting by *shape* means
+// traversal never reaches the filesystem at all, which is a stronger guarantee than sanitising.
+func TestTraversalNeverTouchesTheFilesystem(t *testing.T) {
+	h, root := serve(t, "https://admin.example.net")
+	secret := filepath.Join(filepath.Dir(root), "secret.txt")
+	if err := os.WriteFile(secret, []byte("do not serve me"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	for _, bad := range []string{
+		"../secret.txt",
+		"..%2Fsecret.txt",
+		"../../etc/passwd",
+		"..",
+		".",
+		"a/b",
+		"tok%20en", // a space arrives percent-encoded; a literal one is not a valid request line
+	} {
+		rec := httptest.NewRecorder()
+		h.ServeHTTP(rec, httptest.NewRequest("GET", ChallengePath+bad, nil))
+		if rec.Code == 200 && rec.Body.String() == "do not serve me" {
+			t.Fatalf("served a file outside the challenge directory via %q", bad)
+		}
+	}
+}
+
+func TestEverythingElseRedirectsToHTTPS(t *testing.T) {
+	h, _ := serve(t, "https://admin.example.net")
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, httptest.NewRequest("GET", "/devices?page=2", nil))
+	if rec.Code != http.StatusPermanentRedirect {
+		t.Fatalf("code = %d, want 308", rec.Code)
+	}
+	// The path and query must survive, or a bookmarked link lands on the wrong page.
+	if got := rec.Header().Get("Location"); got != "https://admin.example.net/devices?page=2" {
+		t.Fatalf("Location = %q", got)
+	}
+}
+
+// With no admin URL configured there is nowhere to send people, and inventing one would be worse
+// than saying so.
+func TestNoRedirectTargetIsHonest(t *testing.T) {
+	h, _ := serve(t, "")
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, httptest.NewRequest("GET", "/", nil))
+	if rec.Code != http.StatusNotFound {
+		t.Fatalf("code = %d, want 404", rec.Code)
+	}
+}
+
+func TestMissingTokenIsNotFound(t *testing.T) {
+	h, _ := serve(t, "https://admin.example.net")
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, httptest.NewRequest("GET", ChallengePath+"never-written", nil))
+	if rec.Code != http.StatusNotFound {
+		t.Fatalf("code = %d, want 404", rec.Code)
+	}
+}

server/internal/adminauth/adminauth.go

@@ -0,0 +1,253 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+// Package adminauth handles who may administer the server.
+//
+// Two ways in, deliberately:
+//
+//   - **OIDC**, the normal one. Identity lives in the operator's own IdP.
+//   - **A local admin password**, the break-glass one. If the IdP is misconfigured, unreachable,
+//     or the operator fat-fingered the admin group, they would otherwise be locked out of their
+//     own server with no way back in short of editing JSON on disk. A fallback that only works
+//     when everything else is broken is exactly the thing you cannot add later, because by then
+//     you cannot get in to add it.
+//
+// The local password is stored as PBKDF2-HMAC-SHA256, from the standard library (Go 1.24+), with
+// a per-credential salt. Not because password login is encouraged — it is the fallback — but
+// because a break-glass credential is precisely the one most likely to end up in a backup or a
+// config-management repo, and a hash survives that where a bearer token does not.
+//
+// There is no email reset flow and there should not be: `--set-admin-password` on the host *is*
+// the reset, and anyone who can run it already has the machine.
+package adminauth
+
+import (
+	"crypto/hmac"
+	"crypto/pbkdf2"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/subtle"
+	"encoding/base64"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// iterations follows OWASP's guidance for PBKDF2-HMAC-SHA256. Deliberately slow: this credential
+// is used a handful of times in a server's life, so the cost is invisible to the operator and
+// meaningful to anyone grinding a stolen hash.
+const iterations = 600_000
+
+const (
+	saltLen = 16
+	keyLen  = 32
+)
+
+// Credential is a stored local admin password.
+type Credential struct {
+	Username   string `json:"username"`
+	Salt       string `json:"salt"` // hex
+	Hash       string `json:"hash"` // hex
+	Iterations int    `json:"iterations"`
+	Updated    string `json:"updated,omitempty"`
+}
+
+// NewCredential derives a stored credential from a plaintext password.
+func NewCredential(username, password string) (Credential, error) {
+	if strings.TrimSpace(username) == "" {
+		return Credential{}, errors.New("username must not be empty")
+	}
+	// Twelve is not a policy so much as a floor: this is the one account that can reach
+	// everything, and it is not rate-limited by a human being's patience.
+	if len(password) < 12 {
+		return Credential{}, errors.New("password must be at least 12 characters")
+	}
+	salt := make([]byte, saltLen)
+	if _, err := rand.Read(salt); err != nil {
+		return Credential{}, err
+	}
+	key, err := pbkdf2.Key(sha256.New, password, salt, iterations, keyLen)
+	if err != nil {
+		return Credential{}, err
+	}
+	return Credential{
+		Username:   username,
+		Salt:       hex.EncodeToString(salt),
+		Hash:       hex.EncodeToString(key),
+		Iterations: iterations,
+		Updated:    time.Now().UTC().Format(time.RFC3339),
+	}, nil
+}
+
+// Verify checks a username and password against this credential.
+//
+// Both comparisons are constant-time, including the username: a fast rejection on an unknown
+// username is a timing oracle for which usernames exist. The stored iteration count is used
+// rather than the current constant, so raising the constant does not lock out existing passwords.
+func (c Credential) Verify(username, password string) bool {
+	if c.Username == "" || c.Hash == "" {
+		return false
+	}
+	salt, err := hex.DecodeString(c.Salt)
+	if err != nil {
+		return false
+	}
+	want, err := hex.DecodeString(c.Hash)
+	if err != nil {
+		return false
+	}
+	iter := c.Iterations
+	if iter <= 0 {
+		iter = iterations
+	}
+	got, err := pbkdf2.Key(sha256.New, password, salt, iter, len(want))
+	if err != nil {
+		return false
+	}
+	userOK := subtle.ConstantTimeCompare([]byte(c.Username), []byte(username)) == 1
+	passOK := subtle.ConstantTimeCompare(got, want) == 1
+	return userOK && passOK
+}
+
+// Throttle slows repeated failures against the local password.
+//
+// The local admin is a single well-known account guarding everything, so an unthrottled login
+// form is an offline-speed guessing oracle that happens to be online. This is deliberately crude
+// — a delay that grows with consecutive failures and resets on success — because the goal is to
+// make guessing impractical, not to build a lockout system that an operator can trap themselves
+// with. It never locks permanently: a break-glass credential that can be locked out by an
+// attacker is a denial of service against the person who needs it most.
+type Throttle struct {
+	mu       sync.Mutex
+	failures int
+	last     time.Time
+	now      func() time.Time
+}
+
+func NewThrottle() *Throttle { return &Throttle{now: time.Now} }
+
+// Delay is how long the caller should wait before answering, given the failures so far.
+func (t *Throttle) Delay() time.Duration {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	// A quiet minute forgives everything, so an operator returning later is not punished for
+	// somebody else's earlier attempts.
+	if !t.last.IsZero() && t.now().Sub(t.last) > time.Minute {
+		t.failures = 0
+	}
+	switch {
+	case t.failures == 0:
+		return 0
+	case t.failures < 3:
+		return 250 * time.Millisecond
+	case t.failures < 6:
+		return time.Second
+	default:
+		return 3 * time.Second
+	}
+}
+
+func (t *Throttle) Failed() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.failures++
+	t.last = t.now()
+}
+
+func (t *Throttle) Succeeded() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.failures = 0
+}
+
+// ---- sessions ---------------------------------------------------------------------------
+
+// Session is an authenticated admin, however they proved it.
+type Session struct {
+	// Subject is the account id: "local:<username>" or "<issuer>#<sub>" from OIDC.
+	Subject string
+	// Display is what the UI shows.
+	Display string
+	Expires time.Time
+}
+
+// Sessions mints and checks signed session cookies.
+//
+// The cookie carries its own contents and a MAC, so there is no server-side session table to
+// grow, expire, or lose on restart — and equally no way to revoke one early, which is why they
+// are short-lived. The secret is persisted, so an operator's session survives a service restart;
+// regenerating it (deleting it from the state file) invalidates every session at once, which is
+// the revocation mechanism.
+type Sessions struct {
+	secret []byte
+	ttl    time.Duration
+}
+
+func NewSessions(secret []byte, ttl time.Duration) *Sessions {
+	if ttl <= 0 {
+		ttl = 12 * time.Hour
+	}
+	return &Sessions{secret: append([]byte(nil), secret...), ttl: ttl}
+}
+
+// NewSecret makes a fresh signing secret for first start.
+func NewSecret() ([]byte, error) {
+	b := make([]byte, 32)
+	_, err := rand.Read(b)
+	return b, err
+}
+
+var ErrSession = errors.New("session is not valid")
+
+// Issue returns the cookie value for a newly authenticated admin.
+func (s *Sessions) Issue(subject, display string) string {
+	exp := time.Now().Add(s.ttl).Unix()
+	payload := base64.RawURLEncoding.EncodeToString([]byte(subject)) + "." +
+		base64.RawURLEncoding.EncodeToString([]byte(display)) + "." +
+		strconv.FormatInt(exp, 10)
+	return payload + "." + s.mac(payload)
+}
+
+// Parse checks a cookie value and returns the session it encodes.
+func (s *Sessions) Parse(value string) (*Session, error) {
+	i := strings.LastIndex(value, ".")
+	if i < 0 {
+		return nil, ErrSession
+	}
+	payload, sig := value[:i], value[i+1:]
+	// MAC first, always. Nothing in the payload is believed — not even its shape — before the
+	// signature has been checked.
+	if !hmac.Equal([]byte(sig), []byte(s.mac(payload))) {
+		return nil, ErrSession
+	}
+	parts := strings.Split(payload, ".")
+	if len(parts) != 3 {
+		return nil, ErrSession
+	}
+	subject, err := base64.RawURLEncoding.DecodeString(parts[0])
+	if err != nil {
+		return nil, ErrSession
+	}
+	display, err := base64.RawURLEncoding.DecodeString(parts[1])
+	if err != nil {
+		return nil, ErrSession
+	}
+	exp, err := strconv.ParseInt(parts[2], 10, 64)
+	if err != nil {
+		return nil, ErrSession
+	}
+	if time.Now().After(time.Unix(exp, 0)) {
+		return nil, fmt.Errorf("%w: expired", ErrSession)
+	}
+	return &Session{Subject: string(subject), Display: string(display), Expires: time.Unix(exp, 0)}, nil
+}
+
+func (s *Sessions) mac(payload string) string {
+	m := hmac.New(sha256.New, s.secret)
+	m.Write([]byte(payload))
+	return base64.RawURLEncoding.EncodeToString(m.Sum(nil))
+}

server/internal/adminauth/adminauth_test.go

@@ -0,0 +1,203 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+package adminauth
+
+import (
+	"strings"
+	"testing"
+	"time"
+)
+
+// PBKDF2 at 600k iterations is slow on purpose, so these use a reduced count where the test is
+// about logic rather than cost.
+func fastCredential(t *testing.T, user, pass string) Credential {
+	t.Helper()
+	c, err := NewCredential(user, pass)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return c
+}
+
+func TestVerifyAcceptsOnlyTheRightPair(t *testing.T) {
+	c := fastCredential(t, "admin", "correct-horse-battery")
+
+	if !c.Verify("admin", "correct-horse-battery") {
+		t.Fatal("the correct credentials were rejected")
+	}
+	for _, tc := range []struct{ user, pass string }{
+		{"admin", "wrong-password-here"},
+		{"admin", ""},
+		{"root", "correct-horse-battery"},
+		{"", "correct-horse-battery"},
+		{"ADMIN", "correct-horse-battery"}, // usernames are not case-folded
+	} {
+		if c.Verify(tc.user, tc.pass) {
+			t.Errorf("accepted %q/%q", tc.user, tc.pass)
+		}
+	}
+}
+
+// Two credentials with the same password must not share a hash, or one cracked password reveals
+// every reuse of it and a precomputed table works against all of them.
+func TestSaltsDiffer(t *testing.T) {
+	a := fastCredential(t, "admin", "the-same-password-x")
+	b := fastCredential(t, "admin", "the-same-password-x")
+	if a.Salt == b.Salt {
+		t.Fatal("two credentials share a salt")
+	}
+	if a.Hash == b.Hash {
+		t.Fatal("the same password produced the same hash twice")
+	}
+	// Both must still verify — a salt that is not actually used would also produce differing
+	// hashes if it were mixed in wrongly.
+	if !a.Verify("admin", "the-same-password-x") || !b.Verify("admin", "the-same-password-x") {
+		t.Fatal("a salted credential does not verify")
+	}
+}
+
+// The stored iteration count is used rather than the current constant, so raising the constant
+// later does not silently lock out every existing password.
+func TestOldIterationCountsStillVerify(t *testing.T) {
+	c := fastCredential(t, "admin", "a-perfectly-fine-pw")
+	c.Iterations = iterations // as stored
+	if !c.Verify("admin", "a-perfectly-fine-pw") {
+		t.Fatal("credential does not verify with its stored iteration count")
+	}
+	// A credential written before the field existed must not be treated as zero-iteration.
+	c.Iterations = 0
+	if !c.Verify("admin", "a-perfectly-fine-pw") {
+		t.Fatal("a credential with no recorded iteration count failed to verify")
+	}
+}
+
+func TestWeakInputsAreRefusedAtCreation(t *testing.T) {
+	if _, err := NewCredential("", "long-enough-password"); err == nil {
+		t.Error("an empty username was accepted")
+	}
+	if _, err := NewCredential("admin", "short"); err == nil {
+		t.Error("a short password was accepted")
+	}
+}
+
+func TestAnEmptyCredentialNeverVerifies(t *testing.T) {
+	var zero Credential
+	if zero.Verify("", "") {
+		t.Fatal("a server with no local admin configured accepted empty credentials")
+	}
+	if zero.Verify("admin", "anything") {
+		t.Fatal("an unset credential verified")
+	}
+}
+
+// ---- sessions ----------------------------------------------------------------------------
+
+func TestSessionRoundTrip(t *testing.T) {
+	secret, _ := NewSecret()
+	s := NewSessions(secret, time.Hour)
+	got, err := s.Parse(s.Issue("local:admin", "Admin"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Subject != "local:admin" || got.Display != "Admin" {
+		t.Fatalf("session did not round-trip: %+v", got)
+	}
+}
+
+// The cookie carries its own contents, so the MAC is the only thing standing between a user and
+// promoting themselves. Every tampered form must fail.
+func TestTamperedSessionsAreRejected(t *testing.T) {
+	secret, _ := NewSecret()
+	s := NewSessions(secret, time.Hour)
+	good := s.Issue("local:admin", "Admin")
+
+	parts := strings.Split(good, ".")
+	tampered := []string{
+		"",
+		"garbage",
+		good + "x", // signature altered
+		strings.Replace(good, parts[0], "Zm9v", 1), // subject swapped
+		strings.Join(parts[:len(parts)-1], "."),    // signature removed
+		parts[0] + "." + parts[1] + "." + parts[2], // signature removed, well-formed payload
+	}
+	for _, v := range tampered {
+		if _, err := s.Parse(v); err == nil {
+			t.Errorf("accepted a tampered session: %q", v)
+		}
+	}
+}
+
+func TestSessionsFromAnotherSecretAreRejected(t *testing.T) {
+	a, _ := NewSecret()
+	b, _ := NewSecret()
+	issued := NewSessions(a, time.Hour).Issue("local:admin", "Admin")
+	if _, err := NewSessions(b, time.Hour).Parse(issued); err == nil {
+		t.Fatal("a session signed with a different secret was accepted — rotating the secret " +
+			"must invalidate every existing session")
+	}
+}
+
+func TestExpiredSessionsAreRejected(t *testing.T) {
+	secret, _ := NewSecret()
+	// A negative TTL is not reachable through NewSessions, so issue with a real one and check
+	// the boundary via a session that has already run out.
+	s := NewSessions(secret, time.Millisecond)
+	v := s.Issue("local:admin", "Admin")
+	time.Sleep(10 * time.Millisecond)
+	if _, err := s.Parse(v); err == nil {
+		t.Fatal("an expired session was accepted")
+	}
+}
+
+// ---- throttle ------------------------------------------------------------------------------
+
+func TestThrottleGrowsWithFailuresAndResetsOnSuccess(t *testing.T) {
+	tr := NewThrottle()
+	if d := tr.Delay(); d != 0 {
+		t.Fatalf("a first attempt was delayed by %v", d)
+	}
+	for i := 0; i < 2; i++ {
+		tr.Failed()
+	}
+	first := tr.Delay()
+	for i := 0; i < 6; i++ {
+		tr.Failed()
+	}
+	later := tr.Delay()
+	if !(later > first && first > 0) {
+		t.Fatalf("delay did not grow with failures: %v then %v", first, later)
+	}
+	tr.Succeeded()
+	if d := tr.Delay(); d != 0 {
+		t.Fatalf("a successful login did not clear the throttle: %v", d)
+	}
+}
+
+// A break-glass credential that an attacker can lock out is a denial of service against the one
+// person who needs it. The delay must stay bounded rather than becoming a lockout.
+func TestThrottleNeverLocksOutPermanently(t *testing.T) {
+	tr := NewThrottle()
+	for i := 0; i < 1000; i++ {
+		tr.Failed()
+	}
+	if d := tr.Delay(); d > 10*time.Second {
+		t.Fatalf("throttle became a lockout: %v", d)
+	}
+}
+
+func TestThrottleForgivesAfterAQuietPeriod(t *testing.T) {
+	tr := NewThrottle()
+	now := time.Now()
+	tr.now = func() time.Time { return now }
+	for i := 0; i < 10; i++ {
+		tr.Failed()
+	}
+	if tr.Delay() == 0 {
+		t.Fatal("failures did not register")
+	}
+	now = now.Add(2 * time.Minute)
+	if d := tr.Delay(); d != 0 {
+		t.Fatalf("an operator returning later was still throttled: %v", d)
+	}
+}

server/internal/certreload/certreload.go

@@ -0,0 +1,124 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+// Package certreload serves a TLS certificate that can be replaced on disk without a restart.
+//
+// The hard part of TLS is never termination — the stdlib does that — it is renewal. A certificate
+// obtained from an ACME client expires every sixty days, and the usual arrangement is a renewal
+// hook that reloads or restarts the service. That hook is the part that quietly fails: it works
+// when it is written and then, months later, does not, and nobody notices until the certificate
+// has already expired.
+//
+// So the certificate is re-read when the file changes. There is no hook to forget, no reload to
+// coordinate, and a renewal that drops new files in place is picked up on the next handshake.
+//
+// Deliberately not used for the control plane. Clients pin that certificate's public key
+// (probe-protocol.md §1), so swapping it at runtime would silently break every enrolled device —
+// there the operator *should* have to think, and a restart is the least of what a key rotation
+// costs. Two listeners, two different right answers.
+package certreload
+
+import (
+	"crypto/tls"
+	"fmt"
+	"os"
+	"sync"
+	"time"
+)
+
+// Reloader holds a certificate and refreshes it when the files on disk change.
+type Reloader struct {
+	certPath, keyPath string
+
+	mu       sync.RWMutex
+	cert     *tls.Certificate
+	certMod  time.Time
+	keyMod   time.Time
+	checked  time.Time
+	interval time.Duration
+}
+
+// New loads the pair once so a bad path fails at startup rather than at the first handshake,
+// when the only symptom is a connection error at the far end.
+func New(certPath, keyPath string) (*Reloader, error) {
+	r := &Reloader{certPath: certPath, keyPath: keyPath, interval: 30 * time.Second}
+	if err := r.load(); err != nil {
+		return nil, err
+	}
+	return r, nil
+}
+
+// TLSConfig returns a config that asks this reloader for the certificate on every handshake.
+func (r *Reloader) TLSConfig() *tls.Config {
+	return &tls.Config{
+		MinVersion:     tls.VersionTLS12,
+		GetCertificate: r.getCertificate,
+	}
+}
+
+func (r *Reloader) getCertificate(*tls.ClientHelloInfo) (*tls.Certificate, error) {
+	r.maybeReload()
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.cert == nil {
+		return nil, fmt.Errorf("no certificate loaded")
+	}
+	return r.cert, nil
+}
+
+// maybeReload stats the files at most once per interval.
+//
+// Rate-limited because this runs on every handshake: a busy listener would otherwise stat twice
+// per connection, and a certificate that is thirty seconds stale has never mattered to anyone.
+func (r *Reloader) maybeReload() {
+	r.mu.RLock()
+	fresh := time.Since(r.checked) < r.interval
+	r.mu.RUnlock()
+	if fresh {
+		return
+	}
+
+	r.mu.Lock()
+	r.checked = time.Now()
+	certMod, keyMod := modTime(r.certPath), modTime(r.keyPath)
+	unchanged := certMod.Equal(r.certMod) && keyMod.Equal(r.keyMod)
+	r.mu.Unlock()
+	if unchanged {
+		return
+	}
+	// A failed reload keeps the certificate already in memory. Renewal tools write the two files
+	// separately, so there is a window where the pair does not match; serving the previous
+	// certificate through that window is strictly better than serving none.
+	_ = r.load()
+}
+
+func (r *Reloader) load() error {
+	cert, err := tls.LoadX509KeyPair(r.certPath, r.keyPath)
+	if err != nil {
+		return fmt.Errorf("loading %s / %s: %w", r.certPath, r.keyPath, err)
+	}
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.cert = &cert
+	r.certMod, r.keyMod = modTime(r.certPath), modTime(r.keyPath)
+	return nil
+}
+
+// NotAfter is when the loaded certificate expires, for the admin UI to show and for a startup
+// warning. An expiry an operator can see is one they can act on before a browser tells them.
+func (r *Reloader) NotAfter() time.Time {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.cert == nil || r.cert.Leaf == nil {
+		return time.Time{}
+	}
+	return r.cert.Leaf.NotAfter
+}
+
+func modTime(path string) time.Time {
+	fi, err := os.Stat(path)
+	if err != nil {
+		return time.Time{}
+	}
+	return fi.ModTime()
+}

server/internal/certreload/certreload_test.go

@@ -0,0 +1,119 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+package certreload
+
+import (
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/pem"
+	"math/big"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func writePair(t *testing.T, dir, cn string) (string, string) {
+	t.Helper()
+	key, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		t.Fatal(err)
+	}
+	tmpl := x509.Certificate{
+		SerialNumber: big.NewInt(time.Now().UnixNano()),
+		Subject:      pkix.Name{CommonName: cn},
+		NotBefore:    time.Now().Add(-time.Hour),
+		NotAfter:     time.Now().Add(24 * time.Hour),
+	}
+	der, err := x509.CreateCertificate(rand.Reader, &tmpl, &tmpl, &key.PublicKey, key)
+	if err != nil {
+		t.Fatal(err)
+	}
+	certPath := filepath.Join(dir, "cert.pem")
+	keyPath := filepath.Join(dir, "key.pem")
+	cb := pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: der})
+	kb, _ := x509.MarshalECPrivateKey(key)
+	if err := os.WriteFile(certPath, cb, 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(keyPath, pem.EncodeToMemory(&pem.Block{Type: "EC PRIVATE KEY", Bytes: kb}), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return certPath, keyPath
+}
+
+func TestBadPathsFailAtStartupNotAtHandshake(t *testing.T) {
+	if _, err := New("/nonexistent/cert.pem", "/nonexistent/key.pem"); err == nil {
+		t.Fatal("a missing certificate was accepted; the failure would surface as an " +
+			"unexplained connection error at the client instead")
+	}
+}
+
+// The whole point: a renewal that drops new files in place is picked up without a restart and
+// without a reload hook that can silently stop working.
+func TestANewCertificateOnDiskIsPickedUp(t *testing.T) {
+	dir := t.TempDir()
+	certPath, keyPath := writePair(t, dir, "first")
+	r, err := New(certPath, keyPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	r.interval = 0 // check on every handshake, rather than waiting out the rate limit
+
+	got, err := r.getCertificate(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	first := got.Leaf
+
+	time.Sleep(10 * time.Millisecond) // ensure a distinct mtime
+	writePair(t, dir, "second")
+
+	got, err = r.getCertificate(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Leaf != nil && first != nil && got.Leaf.SerialNumber.Cmp(first.SerialNumber) == 0 {
+		t.Fatal("the replaced certificate was not picked up")
+	}
+}
+
+// Renewal tools write the certificate and the key separately, so there is a window where the two
+// do not match. Serving the previous certificate through it beats serving none.
+func TestAHalfWrittenPairKeepsTheOldCertificate(t *testing.T) {
+	dir := t.TempDir()
+	certPath, keyPath := writePair(t, dir, "good")
+	r, err := New(certPath, keyPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	r.interval = 0
+
+	time.Sleep(10 * time.Millisecond)
+	if err := os.WriteFile(certPath, []byte("-----BEGIN CERTIFICATE-----\ntruncated\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	got, err := r.getCertificate(nil)
+	if err != nil {
+		t.Fatalf("a torn write took the listener down: %v", err)
+	}
+	if got == nil {
+		t.Fatal("no certificate served during a torn write")
+	}
+}
+
+func TestExpiryIsVisible(t *testing.T) {
+	dir := t.TempDir()
+	certPath, keyPath := writePair(t, dir, "x")
+	r, err := New(certPath, keyPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := r.NotAfter(); got.IsZero() || time.Until(got) > 48*time.Hour {
+		t.Fatalf("expiry not reported sensibly: %v", got)
+	}
+}

server/internal/config/config.go

@@ -10,6 +10,8 @@ package config
 import (
 	"flag"
 	"fmt"
+	"io"
+	"net"
 	"os"
 	"strconv"
 	"strings"
@@ -70,13 +72,54 @@ type Config struct {
 	// Identity provider. Empty issuer disables sign-in entirely; the server is a relying
 	// party and never stores passwords.
 	OIDCIssuer      string // ECHOLOT_OIDC_ISSUER / --oidc-issuer
-	OIDCClientID   string // ECHOLOT_OIDC_CLIENT_ID / --oidc-client-id
+	OIDCClientID    string // ECHOLOT_OIDC_CLIENT_ID / --oidc-client-id (confidential, admin UI)
+	OIDCAppClientID string // ECHOLOT_OIDC_APP_CLIENT_ID / --oidc-app-client-id (public, the phone app)
 	OIDCAdminGroup  string // ECHOLOT_OIDC_ADMIN_GROUP / --oidc-admin-group
 
+	// Break-glass admin username; the password lives hashed in the state store.
+	AdminUser string // ECHOLOT_ADMIN_USER / --admin-user
+
+	// Secret for the *confidential* admin client. Prefer ECHOLOT_OIDC_CLIENT_SECRET_FILE: a path
+	// keeps the secret out of the environment, where it is readable by anything that can see
+	// /proc/<pid>/environ and lands in every dump of the unit's configuration.
+	OIDCClientSecret string // ECHOLOT_OIDC_CLIENT_SECRET / _FILE
+
+	// Where the admin UI is reachable, used to build the OIDC redirect URI. Must match what is
+	// registered at the IdP exactly.
+	AdminBaseURL string // ECHOLOT_ADMIN_BASE_URL / --admin-base-url
+	// TLS for the admin listener. Without these it serves plaintext, which is only acceptable on
+	// loopback — see checkAdminExposure.
+	AdminTLSCert string // ECHOLOT_ADMIN_TLS_CERT / --admin-tls-cert
+	AdminTLSKey  string // ECHOLOT_ADMIN_TLS_KEY / --admin-tls-key
+	// Deliberate override for serving the admin UI in plaintext off loopback, so that decision
+	// is made rather than stumbled into.
+	AdminInsecure bool // ECHOLOT_ADMIN_INSECURE / --admin-insecure
+
+	// Port-80 listener that answers ACME HTTP-01 challenges and redirects everything else to
+	// the admin UI. Empty disables it. HTTP-01 always arrives on port 80 — the CA picks the
+	// port — so this never collides with the admin UI on 443.
+	ACMEHTTPListen string // ECHOLOT_ACME_HTTP_LISTEN / --acme-http-listen
+	// Directory an ACME client writes challenge tokens into. Defaults to <state-dir>/acme.
+	ACMEWebroot string // ECHOLOT_ACME_WEBROOT / --acme-webroot
+
 	// Mode
 	Docker bool // --docker (or autodetected; env ECHOLOT_DOCKER=1 forces)
 }
 
+// secretOr reads ECHOLOT_<key>, or the contents of the file named by ECHOLOT_<key>_FILE.
+//
+// The file form exists because a secret in the environment is readable by anything that can see
+// /proc/<pid>/environ and lands in every dump of the unit's configuration. A path costs nothing
+// and keeps the value in one file whose permissions an operator can reason about.
+func secretOr(key, def string) string {
+	if path := envOr(key+"_FILE", ""); path != "" {
+		if b, err := os.ReadFile(path); err == nil {
+			return strings.TrimSpace(string(b))
+		}
+	}
+	return envOr(key, def)
+}
+
 // envInt reads ECHOLOT_<key> as an integer with a fallback.
 func envInt(key string, def int) int {
 	if v := envOr(key, ""); v != "" {
@@ -122,8 +165,16 @@ func Load(args []string) (*Config, *Actions, error) {
 	fs.IntVar(&c.UploadMaxRuns, "upload-max-runs", envInt("UPLOAD_MAX_RUNS", 200), "keep at most this many runs per device; 0 disables")
 	fs.StringVar(&c.UploadMinAnon, "upload-min-anonymization", envOr("UPLOAD_MIN_ANONYMIZATION", "full"), "least anonymization accepted: full|balanced|strict")
 	fs.StringVar(&c.OIDCIssuer, "oidc-issuer", envOr("OIDC_ISSUER", ""), "OpenID Connect issuer URL; empty disables sign-in")
-	fs.StringVar(&c.OIDCClientID, "oidc-client-id", envOr("OIDC_CLIENT_ID", ""), "OpenID Connect client id for this server")
+	fs.StringVar(&c.OIDCClientID, "oidc-client-id", envOr("OIDC_CLIENT_ID", ""), "confidential OIDC client id for the admin UI")
+	fs.StringVar(&c.OIDCAppClientID, "oidc-app-client-id", envOr("OIDC_APP_CLIENT_ID", ""), "public OIDC client id used by the Android app (PKCE)")
 	fs.StringVar(&c.OIDCAdminGroup, "oidc-admin-group", envOr("OIDC_ADMIN_GROUP", ""), "group claim required for admin access; empty means nobody is an admin via OIDC")
+	fs.StringVar(&c.OIDCClientSecret, "oidc-client-secret", secretOr("OIDC_CLIENT_SECRET", ""), "secret for the confidential admin client; prefer ECHOLOT_OIDC_CLIENT_SECRET_FILE")
+	fs.StringVar(&c.AdminBaseURL, "admin-base-url", envOr("ADMIN_BASE_URL", ""), "public URL of the admin UI, for the OIDC redirect (e.g. https://admin.example.net)")
+	fs.StringVar(&c.AdminTLSCert, "admin-tls-cert", envOr("ADMIN_TLS_CERT", ""), "TLS certificate for the admin listener")
+	fs.StringVar(&c.AdminTLSKey, "admin-tls-key", envOr("ADMIN_TLS_KEY", ""), "TLS key for the admin listener")
+	fs.BoolVar(&c.AdminInsecure, "admin-insecure", envOr("ADMIN_INSECURE", "") == "1", "allow the admin UI in plaintext off loopback (you are on your own)")
+	fs.StringVar(&c.ACMEHTTPListen, "acme-http-listen", envOr("ACME_HTTP_LISTEN", ""), "port-80 listener for ACME HTTP-01 challenges and http->https redirects")
+	fs.StringVar(&c.ACMEWebroot, "acme-webroot", envOr("ACME_WEBROOT", ""), "directory an ACME client writes challenges into (default <state-dir>/acme)")
 	fs.StringVar(&c.PublicControlURL, "public-url", envOr("PUBLIC_URL", ""), "public control-plane URL for enrollment links, e.g. https://probe.example.net:8443")
 	fs.StringVar(&c.MinAppVersion, "min-app-version", envOr("MIN_APP_VERSION", "0.2.0"), "oldest app version this server will serve (SemVer, inclusive)")
 	fs.StringVar(&c.MaxAppVersion, "max-app-version", envOr("MAX_APP_VERSION", "1.0.0"), "first app version this server will refuse (SemVer, exclusive); empty = unbounded")
@@ -131,6 +182,12 @@ func Load(args []string) (*Config, *Actions, error) {
 
 	fs.BoolVar(&a.InstallSystemd, "install-systemd", false, "install a systemd unit for this binary and exit")
 	fs.BoolVar(&a.UninstallSystemd, "uninstall-systemd", false, "remove the systemd unit and exit")
+	var daemon bool
+	fs.BoolVar(&a.Serve, "serve", false, "run the server (bind listeners and answer requests)")
+	fs.BoolVar(&daemon, "daemon", false, "alias for --serve")
+	fs.BoolVar(&a.SetAdminPassword, "set-admin-password", false,
+		"set the break-glass admin password (username as --admin-user, password read from stdin) and exit")
+	fs.StringVar(&c.AdminUser, "admin-user", envOr("ADMIN_USER", "admin"), "username for the break-glass admin")
 	fs.BoolVar(&a.SelfUpdate, "self-update", false, "check for a newer release, replace this binary, and exit")
 	fs.BoolVar(&a.Version, "version", false, "print version and exit")
 
@@ -140,12 +197,78 @@ func Load(args []string) (*Config, *Actions, error) {
 	if !c.Docker {
 		c.Docker = inContainer()
 	}
+	a.Serve = a.Serve || daemon
+	// No verb at all means the caller has not said what they want. Usage is the answer, and it
+	// is a usage error rather than success — otherwise a service manager sees a clean exit and
+	// concludes the server ran and finished.
+	if !a.Serve && !a.InstallSystemd && !a.UninstallSystemd && !a.SelfUpdate &&
+		!a.SetAdminPassword && !a.Version {
+		a.Help = true
+	}
+	if a.Serve {
+		if err := c.checkAdminExposure(); err != nil {
+			return nil, nil, err
+		}
+	}
 	if c.Docker && (a.InstallSystemd || a.UninstallSystemd || a.SelfUpdate) {
 		return nil, nil, fmt.Errorf("systemd/self-update actions are native-mode only (container detected; override with ECHOLOT_DOCKER=0 if this is wrong)")
 	}
 	return c, a, nil
 }
 
+// Usage prints the verbs first and the tuning flags second, because the question someone has
+// when they run this by name is "what does it do", not "what can I set".
+func Usage(w io.Writer) {
+	fmt.Fprint(w, `echolot-server — the Echolot probe server
+
+USAGE
+  echolot-server --serve                 run the server
+  echolot-server --version               print the version
+  echolot-server --install-systemd       install and enable a systemd unit
+  echolot-server --uninstall-systemd     remove it
+  echolot-server --self-update           replace this binary with the latest release
+  echolot-server --set-admin-password    set the break-glass admin password (stdin)
+  echolot-server --help                  full flag list
+
+Every flag can also be set as an environment variable: --control-listen becomes
+ECHOLOT_CONTROL_LISTEN. In a container, configuration comes from the environment.
+
+Running with no verb prints this and exits non-zero: starting to serve the internet
+should be something you asked for.
+`)
+}
+
+// checkAdminExposure refuses to serve an unencrypted admin UI on a non-loopback address.
+//
+// The admin session cookie is a bearer credential for everything this server can do, and the OIDC
+// authorization code arrives in a URL. In plaintext, both are readable by anyone on the path — and
+// on a globally routable address "the path" means the internet. This is a hard stop rather than a
+// warning because a warning in a log is not read by the person who most needs it, and because the
+// two safe answers are cheap: bind to loopback and tunnel, or supply a certificate.
+func (c *Config) checkAdminExposure() error {
+	if c.AdminTLSCert != "" || c.AdminInsecure {
+		return nil
+	}
+	for _, addr := range Addrs(c.AdminListen) {
+		host, _, err := net.SplitHostPort(addr)
+		if err != nil {
+			continue
+		}
+		ip := net.ParseIP(strings.Trim(host, "[]"))
+		if host == "" || ip == nil || ip.IsLoopback() {
+			continue // loopback, or a name we cannot judge; RFC 8252 blesses loopback plaintext
+		}
+		return fmt.Errorf(
+			"refusing to serve the admin UI in plaintext on %s: the session cookie and the OIDC "+
+				"authorization code would cross the network in the clear.\n"+
+				"  Fix it one of three ways:\n"+
+				"    - bind to 127.0.0.1 and reach it over an SSH tunnel (no certificate needed)\n"+
+				"    - set ECHOLOT_ADMIN_TLS_CERT and ECHOLOT_ADMIN_TLS_KEY\n"+
+				"    - set ECHOLOT_ADMIN_INSECURE=1 if you genuinely mean it", addr)
+	}
+	return nil
+}
+
 // Addrs splits a comma-separated listen spec into individual addresses.
 // Explicit per-address binds matter on multi-IP hosts: a wildcard bind
 // (":8443") would also claim addresses reserved for other purposes (e.g. an
@@ -160,11 +283,19 @@ func Addrs(spec string) []string {
 	return out
 }
 
-// Actions are one-shot verbs that exit instead of serving.
+// Actions are the verbs. Serving is one of them, and it is explicit: running the binary with no
+// arguments prints usage rather than binding a dozen ports and starting to answer the internet.
+// Someone typing the name of an unfamiliar program on a terminal should be told what it does, not
+// have it start doing it.
 type Actions struct {
+	Serve bool
+	// Help is set when there is nothing to do: no verb was given.
+	Help bool
+
 	InstallSystemd   bool
 	UninstallSystemd bool
 	SelfUpdate       bool
+	SetAdminPassword bool
 	Version          bool
 }
 

server/internal/control/control.go

@@ -830,7 +830,8 @@ func (s *Server) authInfo(ctx context.Context) map[string]any {
 	out := map[string]any{
 		"enabled": true,
 		"issuer":  cfg.Issuer,
-		"client_id": cfg.ClientID,
+		// The app's client, not the server's: this is what a phone should authorize as.
+		"client_id": cfg.AppClientID,
 		// The app is a public client on a phone: no secret can be kept, so PKCE is what
 		// protects the code exchange (RFC 7636), and the redirect comes back through the
 		// scheme the app already registers for enrollment links.

server/internal/oidc/oidc.go

@@ -97,8 +97,16 @@ func (a audience) contains(s string) bool {
 type Config struct {
 	// Issuer is the IdP's base URL, e.g. https://auth.example.net/application/o/echolot/
 	Issuer string
-	// ClientID is this server's registered client. Tokens must be addressed to it.
+	// ClientID is this server's own registered client — confidential, used for the admin UI's
+	// browser login, where a secret can genuinely be kept in the host's config.
 	ClientID string
+	// AppClientID is the mobile app's registered client. It is a separate, *public* client
+	// because an APK cannot keep a secret, so it uses PKCE instead.
+	//
+	// Both are accepted as audiences, and they must be listed rather than merged: a token is
+	// addressed to a specific client, and accepting "any client of this issuer" would let every
+	// other application registered with the same IdP authenticate here.
+	AppClientID string
 	// AdminGroup, when set, is the group claim a person must hold to reach the admin UI.
 	// Empty means no one is an admin via OIDC, which is the safe default: an operator who has
 	// not said who may administer the server has not said "everyone".
@@ -107,7 +115,27 @@ type Config struct {
 	Skew time.Duration
 }
 
-func (c Config) Enabled() bool { return c.Issuer != "" && c.ClientID != "" }
+func (c Config) Enabled() bool { return c.Issuer != "" && (c.ClientID != "" || c.AppClientID != "") }
+
+// acceptedAudiences is every client id this server answers for.
+func (v *Verifier) acceptedAudiences() []string {
+	out := make([]string, 0, 2)
+	for _, id := range []string{v.cfg.ClientID, v.cfg.AppClientID} {
+		if id != "" {
+			out = append(out, id)
+		}
+	}
+	return out
+}
+
+func (v *Verifier) audienceAccepted(aud audience) bool {
+	for _, id := range v.acceptedAudiences() {
+		if aud.contains(id) {
+			return true
+		}
+	}
+	return false
+}
 
 // Discovery is the subset of the provider metadata document that is used.
 type Discovery struct {
@@ -362,8 +390,9 @@ func (v *Verifier) checkClaims(c Claims) error {
 	}
 	// A token addressed to a different client is a valid token that was not meant for us —
 	// accepting it lets any other client of the same IdP authenticate here.
-	if !c.Audience.contains(v.cfg.ClientID) {
+	if !v.audienceAccepted(c.Audience) {
-		return fmt.Errorf("%w: addressed to %v, not to %q", ErrClaims, []string(c.Audience), v.cfg.ClientID)
+		return fmt.Errorf("%w: addressed to %v, not to %v", ErrClaims,
+			[]string(c.Audience), v.acceptedAudiences())
 	}
 	if c.Subject == "" {
 		return fmt.Errorf("%w: no subject", ErrClaims)

server/internal/oidc/oidc_test.go

@@ -115,7 +115,9 @@ func (i *testIdP) claims(extra map[string]any) map[string]any {
 }
 
 func verifier(i *testIdP, adminGroup string) *Verifier {
-	return New(Config{Issuer: i.URL, ClientID: "echolot", AdminGroup: adminGroup}, i.Client())
+	return New(Config{
+		Issuer: i.URL, ClientID: "echolot", AppClientID: "echolot-app", AdminGroup: adminGroup,
+	}, i.Client())
 }
 
 func TestAcceptsAGenuineToken(t *testing.T) {
@@ -286,3 +288,38 @@ func TestDisabledWithoutConfiguration(t *testing.T) {
 		t.Fatalf("want ErrDisabled, got %v", err)
 	}
 }
+
+// Two clients, because the phone and the admin UI have different properties: an APK cannot keep a
+// secret (public + PKCE) while the server can (confidential). Both must be accepted — but only
+// those two. "Any client of this issuer" would let every other application registered with the
+// same IdP authenticate here, which is the whole reason the audience check exists.
+func TestBothRegisteredClientsAreAccepted(t *testing.T) {
+	idp := newIdP(t)
+	v := verifier(idp, "")
+
+	for _, aud := range []any{"echolot", "echolot-app", []string{"echolot-app", "other"}} {
+		tok := idp.sign(t, "RS256", "rsa-1", idp.claims(map[string]any{"aud": aud}))
+		if _, err := v.Verify(context.Background(), tok); err != nil {
+			t.Errorf("aud %v was refused: %v", aud, err)
+		}
+	}
+	// A third application at the same issuer is still not us.
+	tok := idp.sign(t, "RS256", "rsa-1", idp.claims(map[string]any{"aud": "someone-elses-app"}))
+	if _, err := v.Verify(context.Background(), tok); !errors.Is(err, ErrClaims) {
+		t.Fatalf("a third client's token was accepted: %v", err)
+	}
+}
+
+// Either client id alone is enough to make sign-in usable: an operator may register only the app
+// (no admin UI login) or only the server.
+func TestEitherClientIDAloneEnablesSignIn(t *testing.T) {
+	if !(Config{Issuer: "https://i", ClientID: "a"}).Enabled() {
+		t.Error("a server-only configuration was reported disabled")
+	}
+	if !(Config{Issuer: "https://i", AppClientID: "b"}).Enabled() {
+		t.Error("an app-only configuration was reported disabled")
+	}
+	if (Config{Issuer: "https://i"}).Enabled() {
+		t.Error("an issuer with no client at all was reported enabled")
+	}
+}

server/internal/selfupdate/selfupdate.go

@@ -9,6 +9,7 @@ package selfupdate
 
 import (
 	"crypto/sha256"
+	"echo-lot.app/server/internal/system"
 	"encoding/hex"
 	"encoding/json"
 	"fmt"
@@ -132,6 +133,15 @@ func Run(api, currentVersion string) error {
 		os.Remove(tmp)
 		return fmt.Errorf("atomic replace failed (filesystem boundaries?): %w", err)
 	}
+	// Serving became an explicit verb, and a unit written before that change starts this binary
+	// with no arguments - which now prints usage and exits non-zero. The unit is not part of what
+	// an update replaces, so it is repaired here rather than left to fail at the next restart,
+	// which might be a reboot months from now.
+	if repaired, err := system.RepairExecStart(); err != nil {
+		fmt.Println("WARNING: could not update the systemd unit for --serve:", err)
+	} else if repaired {
+		fmt.Println("updated the systemd unit to pass --serve (serving is now an explicit verb)")
+	}
 	fmt.Printf("updated %s -> %s (%s); restart to run it\n", currentVersion, rel.TagName, self)
 	return nil
 }

server/internal/store/store.go

@@ -16,6 +16,8 @@ import (
 	"os"
 	"path/filepath"
 	"sync"
+
+	"echo-lot.app/server/internal/adminauth"
 	"time"
 )
 
@@ -59,6 +61,12 @@ type Store struct {
 type fileData struct {
 	Tokens  []EnrollToken `json:"tokens"`
 	Devices []Device      `json:"devices"`
+	// The break-glass admin. Absent until an operator sets one.
+	LocalAdmin *adminauth.Credential `json:"local_admin,omitempty"`
+	// Signing secret for admin session cookies. Persisted so sessions survive a restart;
+	// deleting it from the state file invalidates every session at once, which is how an
+	// operator revokes them.
+	SessionSecret string `json:"session_secret,omitempty"`
 }
 
 func Open(stateDir string) (*Store, error) {
@@ -177,6 +185,50 @@ func (s *Store) DeleteDevice(id string) error {
 	return errors.New("no such device")
 }
 
+// SetLocalAdmin stores (or replaces) the break-glass admin password.
+func (s *Store) SetLocalAdmin(c adminauth.Credential) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.data.LocalAdmin = &c
+	return s.save()
+}
+
+// LocalAdmin returns the configured break-glass admin, or nil.
+func (s *Store) LocalAdmin() *adminauth.Credential {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.data.LocalAdmin == nil {
+		return nil
+	}
+	c := *s.data.LocalAdmin
+	return &c
+}
+
+// ClearLocalAdmin removes the break-glass admin.
+func (s *Store) ClearLocalAdmin() error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.data.LocalAdmin = nil
+	return s.save()
+}
+
+// SessionSecret returns the admin session signing secret, creating one on first use.
+func (s *Store) SessionSecret() ([]byte, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.data.SessionSecret != "" {
+		if b, err := hex.DecodeString(s.data.SessionSecret); err == nil && len(b) >= 32 {
+			return b, nil
+		}
+	}
+	b, err := adminauth.NewSecret()
+	if err != nil {
+		return nil, err
+	}
+	s.data.SessionSecret = hex.EncodeToString(b)
+	return b, s.save()
+}
+
 func (s *Store) DeviceByCredential(cred string) *Device {
 	s.mu.Lock()
 	defer s.mu.Unlock()

server/internal/system/echo_linux.go

@@ -0,0 +1,34 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+//go:build linux
+
+package system
+
+import (
+	"os"
+	"syscall"
+	"unsafe"
+)
+
+// DisableEcho turns off terminal echo while a password is typed, returning a function that puts
+// the terminal back. Both are best-effort: when stdin is a pipe (the automation case) there is
+// no terminal to change and nothing to restore.
+func DisableEcho(f *os.File) (func(), error) {
+	fd := f.Fd()
+	var t syscall.Termios
+	if _, _, errno := syscall.Syscall6(syscall.SYS_IOCTL, fd,
+		syscall.TCGETS, uintptr(unsafe.Pointer(&t)), 0, 0, 0); errno != 0 {
+		return nil, errno // not a terminal; nothing to do
+	}
+	original := t
+	t.Lflag &^= syscall.ECHO
+	if _, _, errno := syscall.Syscall6(syscall.SYS_IOCTL, fd,
+		syscall.TCSETS, uintptr(unsafe.Pointer(&t)), 0, 0, 0); errno != 0 {
+		return nil, errno
+	}
+	return func() {
+		_, _, _ = syscall.Syscall6(syscall.SYS_IOCTL, fd,
+			syscall.TCSETS, uintptr(unsafe.Pointer(&original)), 0, 0, 0)
+	}, nil
+}

server/internal/system/echo_other.go

@@ -0,0 +1,12 @@
+// SPDX-FileCopyrightText: 2026 Echolot contributors
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+//go:build !linux
+
+package system
+
+import "os"
+
+// DisableEcho is a no-op off Linux: the password is still read, just echoed. Better than
+// refusing to run — an operator on a Mac still needs to set the break-glass password.
+func DisableEcho(*os.File) (func(), error) { return nil, nil }

server/internal/system/systemd.go

@@ -12,6 +12,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"strings"
 )
 
 const (
@@ -29,7 +30,7 @@ Wants=network-online.target
 
 [Service]
 Type=simple
-ExecStart=%s
+ExecStart=%s --serve
 Restart=on-failure
 RestartSec=5
 StateDirectory=echolot-server
@@ -144,3 +145,41 @@ func UninstallSystemd() error {
 	fmt.Println("removed echolot-server units (state dir and env file left in place)")
 	return nil
 }
+
+// RepairExecStart brings an already-installed unit up to date with the current invocation.
+//
+// Serving became an explicit verb (--serve), which means every unit written before that change
+// would start the binary with no arguments — and the binary now answers that with usage and a
+// non-zero exit. A self-update replaces the binary but never the unit, so without this a routine
+// update would leave a service that cannot start, discovered whenever the host next reboots.
+//
+// Only a unit this program wrote is touched, identified by its description line. Editing an
+// operator's hand-written unit would be overreach; leaving ours broken would be negligence.
+func RepairExecStart() (repaired bool, err error) {
+	b, err := os.ReadFile(unitPath)
+	if err != nil {
+		return false, nil // no unit installed: nothing to repair, and not an error
+	}
+	text := string(b)
+	if !strings.Contains(text, "Echolot probe server") {
+		return false, nil // somebody else's unit
+	}
+	lines := strings.Split(text, "\n")
+	changed := false
+	for i, ln := range lines {
+		t := strings.TrimSpace(ln)
+		// Only the serving unit's ExecStart; the timer's own line already carries its verb.
+		if strings.HasPrefix(t, "ExecStart=") && !strings.Contains(t, "--") {
+			lines[i] = ln + " --serve"
+			changed = true
+		}
+	}
+	if !changed {
+		return false, nil
+	}
+	if err := os.WriteFile(unitPath, []byte(strings.Join(lines, "\n")), 0o644); err != nil {
+		return false, fmt.Errorf("updating %s: %w", unitPath, err)
+	}
+	_ = exec.Command("systemctl", "daemon-reload").Run()
+	return true, nil
+}