feat: oauth refresh-token grant + per-runner subdomain TLS plumbing
All checks were successful
Deploy to Production / deploy (push) Successful in 52s

OAUTH REFRESH-TOKEN
- oauth_tokens.subject column added (migration applied to prod DB): stores
  the JWT sub claim from the original authorization so refreshes can
  re-mint with the same identity without re-walking the (consumed) code.
- Authorization-code branch now writes subject AND uses a 30-day
  expires_at for the row (was 1h — same as access token, which killed
  refresh after 1h).
- New refresh_token grant branch:
    * looks up token by refresh-hash + expiry
    * client_id must match, client_secret verified if confidential
    * RFC 8707: requested resource must equal stored resource
    * OAuth 2.1 rotation: atomic UPDATE WHERE old_hash → new access JWT,
      new refresh token, extended expiry; loser of a race sees invalid_grant
- Access TTL (1h) and refresh TTL (30d) extracted as constants.

Clients no longer have to re-authorize hourly. Closes Zb-001.

PER-RUNNER SUBDOMAIN TLS (Z1-002)
Code path:
- New MCP_DOMAIN env (e.g. "mcp.buildmymcpserver.com") + RUNNER_MAP_DIR
  (default /var/runner-map) in generator config.
- deployContainer: writes /var/runner-map/<slug>.conf with content
  "slug.MCP_DOMAIN port;" and computes publicUrl as
  https://<slug>.<MCP_DOMAIN>. Falls back to http://host:port when
  MCP_DOMAIN is unset (zero behaviour change until host is configured).
- stopContainer (both api/lib/docker.ts and generator/lib/deploy.ts) now
  accepts an optional slug arg and removes the map fragment. Callers
  (DELETE /v1/servers/:id, admin template takedown) updated.

Infra path (one-time host setup — Marco runs as root):
- scripts/setup-runner-tls.sh:
    1. nginx vhost matching *.mcp.buildmymcpserver.com via regex →
       reads slug→port from /opt/buildmymcpserver/runner-map.combined
    2. systemd inotify service watches the map dir, combines fragments
       on any change, reloads nginx
    3. installs inotify-tools if missing, idempotent
- Prereqs documented at top: Cloudflare wildcard DNS proxied, Origin CA
  cert for *.mcp.buildmymcpserver.com, SSL mode Full (strict).
- After running: edit docker-compose.prod.yml to mount the map dir into
  api + generator, set MCP_DOMAIN in env, recreate containers.

Closes Zb-001 fully. Closes Z1-002 on the code side; one Marco-on-host
action away from closing it on the infra side.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Marco Sadjadi
2026-05-25 22:09:06 +02:00
parent e9827b1f77
commit 8c6f04f034
8 changed files with 361 additions and 10 deletions

View File

@@ -1,14 +1,43 @@
import { spawn } from 'node:child_process';
import fs from 'node:fs/promises';
import path from 'node:path';
/**
* Per-runner nginx map fragment cleanup. Mirrors the generator-side helper
* (apps/generator/src/lib/deploy.ts) — when MCP_DOMAIN is set, the host
* runs an inotify watcher over the map dir that reloads nginx on any
* change. We remove the fragment here so the slug stops serving 502 the
* moment the user deletes their server.
*
* No-op if MCP_DOMAIN isn't configured (legacy http://host:port URLs are
* still in use). Idempotent — missing files are fine.
*/
const MCP_DOMAIN = process.env.MCP_DOMAIN ?? '';
const RUNNER_MAP_DIR = process.env.RUNNER_MAP_DIR ?? '/var/runner-map';
async function removeRunnerMapEntry(slug: string): Promise<void> {
if (!MCP_DOMAIN || !slug) return;
try {
await fs.rm(path.join(RUNNER_MAP_DIR, `${slug}.conf`), { force: true });
} catch {
/* ignore — not critical */
}
}
/**
* Stop and remove a generated MCP container by container id.
* Resolves regardless of outcome — failures are logged but never blocking.
* Production: should be moved to a Coolify HTTP-API call.
* Also drops the slug's nginx map fragment so the public URL stops 502'ing
* the moment the container goes away.
*/
export async function stopContainer(containerId: string): Promise<{ ok: boolean; detail: string }> {
export async function stopContainer(
containerId: string,
slug?: string,
): Promise<{ ok: boolean; detail: string }> {
if (!containerId || containerId.length < 4) {
return { ok: false, detail: 'invalid_container_id' };
}
if (slug) await removeRunnerMapEntry(slug);
return await new Promise<{ ok: boolean; detail: string }>((resolve) => {
const child = spawn('docker', ['rm', '-f', containerId], {
stdio: ['ignore', 'pipe', 'pipe'],

View File

@@ -19,6 +19,12 @@ import { config } from '../config.js';
const db = createDb();
// Access-token lifetime is short so revocation propagates within the hour.
// Refresh-token lifetime is long so legitimate clients don't have to
// re-authorize daily; rotation on every refresh limits exposure if one leaks.
const ACCESS_TOKEN_TTL_S = 3600; // 1 hour
const REFRESH_TOKEN_TTL_MS = 30 * 24 * 3600 * 1000; // 30 days
function sha256(input: string): string {
return crypto.createHash('sha256').update(input).digest('hex');
}
@@ -236,12 +242,13 @@ export async function oauthRoutes(app: FastifyInstance): Promise<void> {
return reply.code(400).send({ error: 'invalid_grant' });
}
const subject = row.code.userId ?? row.client.clientId;
const accessToken = await signAccessToken({
subject: row.code.userId ?? row.client.clientId,
subject,
audience: resource,
issuer: `${config.CONTROL_PLANE_PUBLIC_URL}/oauth`,
scope: row.code.scope ?? '',
ttlSeconds: 3600,
ttlSeconds: ACCESS_TOKEN_TTL_S,
});
const refreshToken = crypto.randomBytes(32).toString('base64url');
await db.insert(oauthTokens).values({
@@ -250,18 +257,98 @@ export async function oauthRoutes(app: FastifyInstance): Promise<void> {
refreshTokenHash: sha256(refreshToken),
scope: row.code.scope ?? null,
resource,
expiresAt: new Date(Date.now() + 3600 * 1000),
subject,
// expiresAt is the REFRESH-token lifetime — 30 days. Access-token
// expiry lives inside the JWT's `exp` claim (1h, set above).
expiresAt: new Date(Date.now() + REFRESH_TOKEN_TTL_MS),
});
return reply.send({
access_token: accessToken,
token_type: 'Bearer',
expires_in: 3600,
expires_in: ACCESS_TOKEN_TTL_S,
refresh_token: refreshToken,
scope: row.code.scope ?? '',
});
}
// ─── grant_type: refresh_token ─────────────────────────────────────
// OAuth 2.1 with rotation: every successful refresh issues a NEW refresh
// token and atomically invalidates the old one. If a stolen refresh token
// gets used after the legitimate client refreshed, the second use sees
// invalid_grant — that's how rotation surfaces token theft.
if (parsed.data.grant_type === 'refresh_token') {
const { refresh_token, client_id, client_secret, resource: requestedResource } =
parsed.data;
if (!refresh_token || !client_id) {
return reply.code(400).send({ error: 'invalid_request' });
}
const refreshHash = sha256(refresh_token);
const [row] = await db
.select({ token: oauthTokens, client: oauthClients })
.from(oauthTokens)
.innerJoin(oauthClients, eq(oauthClients.id, oauthTokens.clientDbId))
.where(
and(
eq(oauthTokens.refreshTokenHash, refreshHash),
gt(oauthTokens.expiresAt, new Date()),
),
)
.limit(1);
if (!row) return reply.code(400).send({ error: 'invalid_grant' });
if (row.client.clientId !== client_id) {
return reply.code(401).send({ error: 'invalid_client' });
}
if (row.client.clientSecretHash) {
if (!client_secret || sha256(client_secret) !== row.client.clientSecretHash) {
return reply.code(401).send({ error: 'invalid_client' });
}
}
// RFC 8707: requested resource must equal the stored one — refreshes
// don't allow audience changes (would be a downgrade/escalation vector).
if (requestedResource && requestedResource !== row.token.resource) {
return reply.code(400).send({ error: 'invalid_resource' });
}
const subject = row.token.subject ?? row.client.clientId;
const newAccessToken = await signAccessToken({
subject,
audience: row.token.resource ?? '',
issuer: `${config.CONTROL_PLANE_PUBLIC_URL}/oauth`,
scope: row.token.scope ?? '',
ttlSeconds: ACCESS_TOKEN_TTL_S,
});
const newRefreshToken = crypto.randomBytes(32).toString('base64url');
const newRefreshHash = sha256(newRefreshToken);
// Atomic rotation: UPDATE only succeeds if the row still has the OLD
// refresh-hash. Two parallel refreshes with the same token can't both
// win — the loser sees zero rows and gets invalid_grant.
const rotated = await db
.update(oauthTokens)
.set({
accessTokenHash: sha256(newAccessToken),
refreshTokenHash: newRefreshHash,
expiresAt: new Date(Date.now() + REFRESH_TOKEN_TTL_MS),
})
.where(
and(eq(oauthTokens.id, row.token.id), eq(oauthTokens.refreshTokenHash, refreshHash)),
)
.returning({ id: oauthTokens.id });
if (rotated.length === 0) {
return reply.code(400).send({ error: 'invalid_grant' });
}
return reply.send({
access_token: newAccessToken,
token_type: 'Bearer',
expires_in: ACCESS_TOKEN_TTL_S,
refresh_token: newRefreshToken,
scope: row.token.scope ?? '',
});
}
return reply.code(400).send({ error: 'unsupported_grant_type' });
});

View File

@@ -521,7 +521,7 @@ export async function serverRoutes(app: FastifyInstance): Promise<void> {
// otherwise it keeps serving traffic with the user's secrets baked in.
let containerStopped = false;
if (server.containerId) {
const result = await stopContainer(server.containerId);
const result = await stopContainer(server.containerId, server.slug);
containerStopped = result.ok;
if (!result.ok) {
app.log.warn(

View File

@@ -559,12 +559,12 @@ export async function templateRoutes(app: FastifyInstance): Promise<void> {
let stoppedContainers = 0;
if (b.data.status === 'takedown') {
const forkedServers = await db
.select({ id: mcpServers.id, containerId: mcpServers.containerId })
.select({ id: mcpServers.id, containerId: mcpServers.containerId, slug: mcpServers.slug })
.from(mcpServers)
.where(eq(mcpServers.templateId, p.data.id));
for (const fork of forkedServers) {
if (fork.containerId) {
const result = await stopContainer(fork.containerId);
const result = await stopContainer(fork.containerId, fork.slug);
if (result.ok) stoppedContainers++;
else
app.log.warn(

View File

@@ -13,6 +13,15 @@ const Env = z.object({
OAUTH_ISSUER: z.string().optional(),
MODEL_GENERATE: z.string().default('glm-4.5'),
MODEL_FIX: z.string().default('claude-haiku-4-5-20251001'),
// When set (e.g. "mcp.buildmymcpserver.com"), each deployed runner gets a
// public URL of the form https://<slug>.<MCP_DOMAIN> instead of the legacy
// http://<RUNNER_HOST>:<port> form. Requires host-side nginx + DNS setup
// (see scripts/setup-runner-tls.sh). When unset, falls back to plain HTTP.
MCP_DOMAIN: z.string().optional(),
// Directory the generator drops per-runner map fragments into. A host-side
// inotify service combines them and reloads nginx. Mounted as a volume by
// docker-compose (see setup-runner-tls.sh).
RUNNER_MAP_DIR: z.string().default('/var/runner-map'),
});
export const config = Env.parse(process.env);

View File

@@ -1,7 +1,51 @@
import fs from 'node:fs/promises';
import net from 'node:net';
import path from 'node:path';
import { createDb, eq, isNotNull, mcpServers } from '@bmm/db';
import { config } from '../config.js';
/**
* Per-runner subdomain TLS support. When MCP_DOMAIN is set, the generator
* publishes each container under https://<slug>.<MCP_DOMAIN> via a host-side
* nginx that reads a slug→port map. The generator writes a tiny config
* fragment per server; a systemd inotify watcher combines them and reloads
* nginx. See scripts/setup-runner-tls.sh for the one-time host setup.
*
* If MCP_DOMAIN is unset, both the URL formatter and the map writer no-op
* and we fall back to the legacy http://host:port URL — zero behaviour
* change without the host-side infra in place.
*/
function runnerMapPath(slug: string): string {
return path.join(config.RUNNER_MAP_DIR, `${slug}.conf`);
}
async function writeRunnerMapEntry(slug: string, port: number): Promise<void> {
if (!config.MCP_DOMAIN) return;
const line = `${slug}.${config.MCP_DOMAIN} ${port};\n`;
try {
await fs.mkdir(config.RUNNER_MAP_DIR, { recursive: true });
await fs.writeFile(runnerMapPath(slug), line, 'utf8');
} catch (err) {
// Don't fail the deploy if the map dir isn't mounted yet — runner still
// serves on http://host:port and the user can manually proxy.
console.warn(`[runner-tls] could not write map entry for ${slug}:`, err);
}
}
async function removeRunnerMapEntry(slug: string): Promise<void> {
if (!config.MCP_DOMAIN) return;
try {
await fs.rm(runnerMapPath(slug), { force: true });
} catch {
// Idempotent — missing file is fine.
}
}
function computePublicUrl(slug: string, port: number): string {
if (config.MCP_DOMAIN) return `https://${slug}.${config.MCP_DOMAIN}`;
return `http://${config.RUNNER_HOST}:${port}`;
}
/**
* Container hardening flags applied on every runner deployment on Linux
* production hosts. Skipped only when explicitly disabled (dev/Windows
@@ -115,7 +159,10 @@ export async function deployContainer(input: DeployInput): Promise<DeployHandle>
return;
}
const containerId = out.trim().slice(0, 64);
const publicUrl = `http://${config.RUNNER_HOST}:${input.hostPort}`;
const publicUrl = computePublicUrl(input.slug, input.hostPort);
// Drop the nginx map fragment BEFORE persisting publicUrl so the
// user-visible URL is reachable by the time the wizard polls "live".
await writeRunnerMapEntry(input.slug, input.hostPort);
await db
.update(mcpServers)
.set({
@@ -133,10 +180,16 @@ export async function deployContainer(input: DeployInput): Promise<DeployHandle>
export async function stopContainer(
containerId: string,
slug?: string,
): Promise<{ ok: boolean; detail: string }> {
if (!containerId || containerId.length < 4) {
return { ok: false, detail: 'invalid_container_id' };
}
// Remove the nginx map fragment first so the slug stops serving 502 from
// the proxy as soon as the container goes down. Idempotent — called
// multiple times with the same slug is fine.
if (slug) await removeRunnerMapEntry(slug);
const { spawn } = await import('node:child_process');
return await new Promise<{ ok: boolean; detail: string }>((resolve) => {
const child = spawn('docker', ['rm', '-f', containerId], {