Repository navigation
Releases
Releases · lalbers/renderfetch-mcp
renderfetch-mcp 2.0.0
A security hardening release with breaking configuration and output changes.
Read Upgrading from 1.0.0 before rolling it out.
Security
- Connection-level SSRF protection: every render gets a loopback-only egress proxy
that validates the complete DNS answer, connects to the checked numeric IP and
keeps Host/TLS validation on the original name. This closes DNS rebinding and
covers redirects and subresources that Playwright routing never sees. - Service workers, WebSockets, downloads, extra pages, QUIC and non-proxied
WebRTC UDP are blocked, and so are write-method requests. - Untrusted-output pipeline: consistent sanitizing across formats, Unicode
normalization before matching, screening of the full bounded document before
truncation, and page titles/URLs only inside the untrusted envelope. Errors no
longer reflect page content. - OAuth: a required
mcp:fetchscope, sessions bound to client and identity,
client-bound revocation, refresh-token families with replay detection and a
fixed absolute expiry, and restart-safe one-use consent. - The consent login limit counts only failed same-origin credential attempts, so
anonymous junk requests can no longer lock the owner out. - Dynamic client registrations that never obtain a token are pruned after
OAUTH_UNUSED_CLIENT_TTL, so/registercannot be filled up permanently. - Secret comparisons no longer leak the secret's length.
- Container deployments run Chromium's renderers in its sandbox, via
deploy/seccomp-chromium.json. The container smoke test verifies it. - Dependencies updated,
npm auditclean: MCP SDK 1.32.2, Playwright 1.64.1,
zod 4, jose 6, jsdom 30, express 5.3.
Added
cookie_bannertool option (hideby default) removes recognized consent
overlays from the output snapshot without clicking or changing consent.- Settings:
ALLOWED_ORIGINS,TRUST_PROXY,FETCH_ALLOWED_PORTS,
FETCH_TIMEOUT_MS,FETCH_MAX_REQUESTS,FETCH_MAX_HTML_BYTES,
FETCH_MAX_TRANSFER_BYTES,BROWSER_MAX_QUEUE,MCP_SESSION_TTL_MS,
MCP_MAX_SESSIONS,MCP_MAX_SESSIONS_PER_CLIENT,OAUTH_MAX_CLIENTS,
OAUTH_UNUSED_CLIENT_TTLandCLIENT_SECRET_TTL. - Real-browser security regression tests run in CI.
Changed
- Page text is returned in both
contentandstructuredContent.text. Some
clients, including the claude.ai connector, render onlystructuredContent. - Client secrets issued by dynamic registration live for one year
(CLIENT_SECRET_TTL) instead of the SDK's 30 days. - Client registrations are logged by
client_idonly, without client-supplied
names or redirect URIs. - Invalid configuration (short secrets, malformed booleans, out-of-range numbers)
fails at startup. An emptySTATIC_BEARER_TOKENdisables static bearer auth.
Removed
structuredContent.titleandstructuredContent.final_url. Read them from the
fenced text instead.
Upgrading from 1.0.0
- Node 22.22.2+, 24.15+ or 26+ is required (Node 24 recommended). Use
the committed lockfile. - Back up the OAuth database first: the refresh-token migration is one-way.
Rotations made under 1.0.0 cannot be recognized as replays afterwards. - Existing tokens without
mcp:fetchmust reconnect. Refresh requires
offline_access. - Configuration is validated at startup. The example values from 1.0.0's
.env.example(change-me…) are rejected.AUTH_PASSWORDneeds 12+
characters.STATIC_BEARER_TOKENneeds 32+ printable ASCII characters without
whitespace.PUBLIC_BASE_URLmust be an HTTPS origin (HTTP only for localhost).
Booleans must betrue/false(or1/0,yes/no,on/off), and numeric
settings have bounds; the startup error names any value out of range. - Clients registered under 1.0.0 keep their 30-day client secret. A connector may
need to be re-linked once when it expires; new registrations get one year. structuredContent.titleand.final_urlare gone; read the fenced text.- Behind a reverse proxy, set
TRUST_PROXYto its address. Otherwise all clients
share one rate-limit bucket and anyone can lock out the login. - Browser tools such as MCP Inspector need an explicit
ALLOWED_ORIGINSentry. - Nonstandard destination ports require
FETCH_ALLOWED_PORTS. Unsafe schemes and
URL credentials are rejected even with private access enabled. - JavaScript still runs, but service workers, WebSockets and write-method requests
are blocked. Some POST-backed applications render incompletely. - The code default of
CHROMIUM_NO_SANDBOXchanged fromtruetofalse. Add
deploy/seccomp-chromium.jsonto your container setup (see the README).
Without it Chromium cannot start sandboxed:/healthzstill reports ok, but
every fetch fails. 1.0.0's example.envsetCHROMIUM_NO_SANDBOX=true
explicitly: change it tofalseafter adding the profile, or Chromium keeps
running unsandboxed. Usetrueonly as a last resort. - Cookie cleanup defaults to
hide. Usecookie_banner: "off"for unmodified
consent text.
Full changelog: v1.0.0...v2.0.0
renderfetch-mcp 1.0.0
First public release: a headless-Chromium fetch_url MCP tool with readability
extraction to Markdown, a self-contained OAuth 2.1 server (dynamic client
registration, PKCE, refresh-token rotation, SQLite store), prompt-injection
screening, an SSRF guard and rootless-Podman deployment examples.