facebook-mcp — Facebook Pages MCP Server

📖 Documentation site →

A local-first TypeScript Model Context Protocol (MCP) server for the Meta Graph API, focused on Facebook Pages — publishing, reading & insights, and comment/message moderation — driven from an MCP client (Claude, VS Code, the Inspector…) and operated locally by a Page admin using their own Meta developer app. Least-privilege tokens, plan-and-apply write safety, and no telemetry.

🚧 Status: pre-1.0, in active development. Published on npm as @ivanbaev/facebook-mcp — and the version number is the warning.

All seven tool packages are implemented — see the tool table for the exact surface. Nothing has been verified against the live Graph API yet. Every test in this repository runs against fakes and recorded fixtures behind a network fence that throws on a real fetch, so "the tests pass" means "the code does what the fixtures say", not "Meta accepted it". A live smoke harness exists (scripts/smoke/) but has not been run against a real Page as part of any released state. Treat every capability below as implemented but unproven, expect breaking changes until 1.0, and read Known limitations before you rely on anything.

Contents: Features · Requirements · Setup · Client compatibility · Configure credentials · Permissions · Tools · How this compares · Known limitations · Not in scope · Your responsibilities · Roadmap · Security notes · Documentation · Support · Trademark · License

Features

Implementation status only — see the status note above for what "implemented" does and does not mean.

The default profile exposes a deliberately small, curated tool surface — each tool wraps one real capability rather than mirroring every Graph edge.

• Local-first, no telemetry. Only three Meta hosts are ever reached (graph, graph-video, rupload); nothing phones home.
• Least-privilege credentials. Works with a never-expiring System User token (Business Manager) or a long-lived Page token — no App Review required for an admin operating their own assets.
• Safe writes. Tiered plan-and-apply gating; irreversible and spend actions require out-of-band confirmation and are never bypassed by an env flag.
• Secret hygiene. Value-based redaction at a single choke-point across logs, errors, tool results, and the write journal.
• Layered architecture. Lint-enforced core ← api ← mcp ← tools layering, tools-as-data with a central package registry.

Requirements

• Node.js ≥ 22 (enforced by engines and a runtime guard in the launcher; the project targets the version in .nvmrc).
• A Meta (Facebook) developer app — a Business-type app at Standard Access is enough; no App Review is needed to operate your own assets.
• A Facebook Page (and, for the opt-in ads package, an ad account) you administer.

Setup

(a) Via npx

Configure credentials first (see below), then run the pre-flight check:

npx -y @ivanbaev/facebook-mcp doctor   # token, scopes, package matrix

Register it with your MCP client:

{
  "mcpServers": {
    "facebook": {
      "command": "npx",
      "args": ["-y", "@ivanbaev/facebook-mcp"]
    }
  }
}

Claude Code plugin (installs the server wired up):

/plugin marketplace add IvanBBaev/facebook-mcp
/plugin install facebook-mcp

Claude Desktop takes the .mcpb bundle attached to the latest release. It carries a build-provenance attestation, so you can prove it came out of this repository's workflow before you install it:

gh attestation verify facebook-mcp-0.7.0.mcpb --repo IvanBBaev/facebook-mcp

(b) From source

For development, or to run a commit that is not released yet:

git clone https://github.com/IvanBBaev/facebook-mcp.git
cd facebook-mcp
npm install
npm run build
node build/index.js doctor   # pre-flight: token, scopes, package matrix
node build/index.js          # or: ./bin/facebook-mcp.mjs

Point your MCP client at that command with an absolute path:

{
  "mcpServers": {
    "facebook": {
      "command": "node",
      "args": ["/absolute/path/to/facebook-mcp/build/index.js"],
      "env": { "FB_SYSTEM_TOKEN": "…", "FB_PAGE_ID": "…" }
    }
  }
}

Client compatibility

This is a standard stdio MCP server with no client-specific code, so it should work anywhere the protocol does. "Should" is the operative word: the table records the configuration each client needs and whether that path has actually been exercised — not a promise.

What is verified, by automated test:

• stdio framing — a test spawns the built server as a real subprocess, runs the full initialize → tools/list → tools/call handshake over pipes, and asserts that stdout carries protocol frames only (all logging goes to stderr).
• Tool contracts — every tool's input/output schema, annotations, write tier and error mapping are unit-tested against fakes and recorded fixtures.
• The HTTP transport — loopback binding, bearer-token rejection and session handling, again against a real local server.

None of that involves an MCP client or the Meta API. If you get it working with a client, say so in an issue — that is how this table turns into ✅.

Platform notes

CI runs the full check on Linux (Node 22/24/26) and Windows (Node 22). macOS is the primary development platform but is not in CI. Paths for the env file, the write journal and FB_MEDIA_DIR follow XDG on POSIX and %APPDATA% on Windows.

Configure credentials

facebook-mcp authenticates with a token you already control — there is no App Review, no OAuth callback server, and no hosted component. Provide at least one of the token variables below; the most specific wins (FB_SYSTEM_TOKEN → FB_ACCESS_TOKEN → FB_PAGE_TOKEN):

• FB_SYSTEM_TOKEN — a never-expiring System User token (Business Manager). Recommended: it does not expire and is scoped to the assets you assign it.
• FB_ACCESS_TOKEN — a Meta user access token (a long-lived one preferred).
• FB_PAGE_TOKEN — a long-lived Page token, the no-Business-Manager fallback. It belongs to exactly one Page, so set FB_PAGE_ID to that Page alongside it (doctor / facebook_whoami reports it as the acting Page).

Grant only the permissions the packages you enable actually need (full list below), and set FB_APP_SECRET so appsecret_proof is attached to every call — that makes a stolen bare token unusable on its own. Settings are read from an env file at the XDG/%APPDATA% config path, or from real environment variables (which take precedence).

If you would rather not assemble the env file by hand, run node build/index.js setup-token — the guided flow exchanges and classifies a token and writes the env file for you (--page=<id>, --env-file=<path>, --force to overwrite, --no-write/--dry-run to rehearse; pass the token via FB_SETUP_TOKEN rather than on the command line). See docs/runbooks/onboarding.md for the full walkthrough.

Then run node build/index.js doctor before anything else: it inspects the token, reports type, expiry and granted scopes, and prints a per-package usable / partial / blocked matrix so you find a missing permission before a tool call does. It ends with a one-line verdict over everything it found.

The doctor is a report, so it exits 0 whatever it finds — a wrapper that runs it for the text keeps working. Add --strict to make it a gate instead:

UNKNOWN outranks WARN on purpose: "nothing was established" must not hide behind "degraded". A platform without POSIX permission bits raises no finding for the credential file — there is no chmod to recommend there, and a warning nothing can clear would make --strict permanently red on Windows.

node build/index.js --version (or -v) prints the server, Node and MCP SDK versions on one line and exits — it needs no credential, so it still answers on an install that cannot start:

facebook-mcp 0.7.0 (node v22.23.0, darwin arm64, sdk 1.30.0)

The bare server version is always the second field, so --version | awk '{print $2}' keeps working. The same three versions plus the pinned Graph API version are reported by facebook_whoami in its server object (name, version, apiVersion, sdkVersion). --help (or -h) prints the command summary to stdout and exits 0, again without reading any credential. doctor, setup-token, --version and --help are the only arguments that exit on their own; anything else starts the stdio server and waits on JSON-RPC. An argument doctor does not recognise (say --stric) never gates, but the report names it — option names only, never a value — so a mistyped --strict in CI is visible.

Environment variables

Provide at least one token; everything else is optional tuning. Variables marked Secret are never logged or returned by a tool.

¹ Provide at least one of FB_SYSTEM_TOKEN, FB_ACCESS_TOKEN or FB_PAGE_TOKEN. A full, commented template lives in .env.example.

Permissions you need to grant

Grant only what the packages you actually enable require. The scopes below are the ones this server asks for; node build/index.js doctor compares them against what your token really has and prints a per-package usable / partial / blocked matrix. business_management is deliberately not in this list — it is a setup-only permission that should never ride on a runtime token.

Tools

Packages are the unit of exposure: you enable and disable whole packages with FB_TOOL_PACKAGES, and FB_PACKAGES_READONLY drops a package's write tools while keeping its reads.

All three package variables share one namespace of packages and profiles (core, all, reader, publisher, moderator, ads), and a profile name wins over a same-spelled package. That matters most in FB_PACKAGES_DENY: FB_PACKAGES_DENY=core denies the six-package core profile, not the single core package — and the core package is then forced back on anyway, since its identity and rate-limit tools are how you diagnose the surface you just narrowed. A name none of them recognize fails at startup rather than being ignored.

37 tools in 7 packages. 6 packages are on by default; the rest are opt-in via FB_TOOL_PACKAGES. Tier read never mutates. reversible writes are gated by FB_WRITE_MODE; irreversible and spend additionally require a per-call apply plus the plan_id of a preview you just ran, and are never unlocked by an environment variable alone.

How this compares

Point-in-time survey (2026-07-21); the full write-up is in docs/analysis/02-market-and-positioning.md.

If you only need ads, use Meta's official server. It is free, needs no developer app of your own, is maintained by the vendor, and covers more ads surface than this project ever will. The ads package here exists for one reason: composing ads with Pages, moderation and messaging under a single token and config, with the same spend gating. That is why it is off by default — ads are not the reason to install this.

The reason to install this is the Pages side: multi-Page, video and Reels upload, scheduled-post lifecycle, insights that tell you when a metric was renamed or retired, and comment/message moderation — in a typed, tested, MIT-licensed server that runs entirely on your machine.

Known limitations

Current, factual, and deliberately unflattering:

• No live Graph API verification. Nothing here has been proven against Meta's servers. Fixtures encode what the API is documented to do; reality gets a vote.
• Published, but brand new. 0.7.0 is on npm with provenance, the .mcpb bundle is attached to the release and the MCP Registry listing is active — but no install path has been walked end-to-end by anyone except CI. See Client compatibility for what that means per client.
• Clients that cannot prompt need an operator token. irreversible and spend writes need a per-call apply plus a plan_id and an out-of-band confirmation the model cannot supply itself. Where the client advertises the MCP elicitation capability, the server asks the human operator through it; where it does not, the caller must pass the FB_CONFIRM_TOKEN value as the tool's confirm_token argument. With neither route, those tools return confirmation_denied.
• Upload progress depends on the client asking for it. Chunked video and Reel uploads emit notifications/progress per chunk — but only when the caller supplies a progressToken on the request, as the MCP spec requires. A client that does not send one gets no frames, and a long upload just looks slow.
• Insights metrics move under you. Meta retired and renamed a large batch of Page/post metrics across three waves (2024-09, 2025-11 and 2026-06-15). The insights tools classify a requested metric and tell you when one is renamed or gone instead of silently returning nothing — but the underlying data loss is Meta's, and no wrapper can undo it.
• Messaging is bound by Meta's 24-hour window. A conversation that has gone quiet for more than 24 hours cannot be replied to without a message tag, and this server does not paper over that.
• Single maintainer, pre-1.0. Interfaces may change without a deprecation period until 1.0.

Not in scope

Things people reasonably expect from a "Facebook MCP server" that this one does not do. Listed so you can rule it out in thirty seconds instead of after an install:

• Personal profiles and Groups. Publishing to a personal timeline has no API, and the Groups API was discontinued in April 2024. This server operates Pages — nothing else.
• Instagram, Threads, WhatsApp. Different products, different permission sets. Possible sibling servers one day; not this one.
• Webhooks / real-time ingestion. A local stdio server has no public URL to receive callbacks on. Reads are polled, by design.
• Boosting an organic post (/{page-id}/promotions). The organic→ads bridge is the most-requested crossover feature and is deliberately parked post-1.0 — it needs the ads package plus spend gating that has been thought through, not bolted on.
• Page Stories. The publishing endpoints exist but are unverified; Stories are not promised for 1.0 and stay out until a live run proves them.
• Events, Live video, album/photo-library reads, Page profile editing, organic post targeting. Real endpoints, outside the post / video / Reels / scheduling / moderation surface this server commits to.
• Multi-tenant or hosted operation, serving other businesses' ad accounts. The design targets one operator, Standard Access, own assets, own machine. Anything else needs App Review and Business Verification, which is a different product.

The full reasoning is in docs/analysis/01-goals-and-scope.md.

Your responsibilities as the operator

This server automates actions Meta attributes to you, under your app and your token. Automating them does not exempt you from the rules that govern them:

• Meta's Platform Terms and Developer Policies still apply — including the anti-spam and automated-behaviour rules. Bulk or repetitive posting, commenting, or messaging can get a Page restricted or an app disabled, and the fact that a model chose the timing is not a defence.
• Messenger automation must be disclosed. Meta expects users to know when they are talking to an automated system. If you wire the messaging tools to answer people, tell them.
• Content and consent are yours. The tools will publish whatever you point them at; deciding that you have the rights to it — and, for private replies and DMs, that the contact is expected — is not something a wrapper can do.
• This is not legal advice. It is the honest note that a capable automation tool comes with obligations, and this project would rather say so up front.

The security-side counterpart is in SECURITY.md.

Roadmap

Work is tracked publicly on the facebook-mcp roadmap board and grouped into release milestones:

v1.0.0 ships without the ads package — deliberately. The Pages surface is the product; ads is opt-in, off by default, and lands in 1.1.0. The code for it is already in the working tree, which is exactly why the milestone is about verification, not implementation: milestones close when a surface has been verified live against Meta's servers and released, not when the code lands. The binding version map is docs/analysis/10-v1-release-definition.md.

The design behind each item is written up in advance in docs/analysis/ — the roadmap is a consequence of that corpus, not a replacement for it.

Security notes

• Three-host fence. Only graph.facebook.com, graph-video.facebook.com and rupload.facebook.com are ever contacted — the allowlist is fixed in code and not user-configurable, so a redirected or mistyped host cannot silently receive a token.
• Plan-and-apply write gating. Writes default to plan (a non-mutating preview); apply executes. The exception is moderation, which defaults to apply because hiding a comment is high-volume, reversible work — and FB_WRITE_MODE overrides any package default outright, in either direction, so the mode is an operator decision rather than a per-tool promise. What the mode cannot touch: irreversible and spend actions always require an out-of-band confirmation, and anything that reaches a live audience (publishing a post or a Reel, publishing a scheduled post early, sending a DM) additionally requires a plan_id bound to a preview of that exact call. No env flag bypasses either.
• Single-choke-point redaction. Secret values are stripped at one place before anything reaches logs, errors, tool results or the write journal; Page access tokens are derived to a boolean and their values never enter a payload.
• No telemetry, local-first. The server logs only to stderr, collects nothing, and phones home nowhere. The http transport binds loopback (127.0.0.1) only and fails closed without FB_HTTP_TOKEN.

See SECURITY.md for the full model and vulnerability reporting.

Documentation

• Design corpus — the full pre-implementation analysis lives in docs/analysis/: goals & scope, market positioning, Graph API landscape, auth & security model, architecture, tool catalog, risks, roadmap, corner cases, the v1.0 release definition, and the parallel task-breakdown that drives development.
• Runbooks — operational procedures in docs/runbooks/.
• Documentation site — ivanbbaev.github.io/facebook-mcp.

The environment table, the package/tool tables and the permission tables above are generated from the code by npm run metadata; CI fails if they drift.

Support

Best-effort, single-maintainer support runs through GitHub — see SUPPORT.md for how to file bugs, feature requests and security reports.

This project is built and maintained in my own time. If it helps, a tip keeps it going:

• GitHub Sponsors — one-off or recurring, no platform fee.
• Ko-fi — quick one-off support (also accepts PayPal), the fallback for anyone without a GitHub account.
• Donatree — every donation method on one page, including local payment options.

Donating buys no priority support and no SLA — see SUPPORT.md. If money is not an option, starring the repository or filing a good bug report helps just as much.

Trademark

This is an independent, community-built project and is not affiliated with, endorsed by, or sponsored by Meta Platforms, Inc. Facebook, Meta and related marks are trademarks of Meta Platforms, Inc., used here only nominatively to indicate compatibility.

License

MIT © 2026 Ivan Baev