mcp-fb-publisher

An MCP server that lets Claude (or any MCP-compatible LLM) safely publish posts to multiple Facebook Pages through the Meta Graph API, with built-in guardrails: brand-voice config, banned-topic blocklists, image-required enforcement, and anti-duplication checks across recent feed posts.

![Tests](https://github.com/anthonyjbolo/mcp-fb-publisher/actions/workflows/ci.yml) ![License: MIT](https://opensource.org/licenses/MIT) ![Python 3.12+](https://www.python.org/downloads/) ![#BuiltOnClaudeCode](https://claude.com/claude-code)

Why this exists

Letting an LLM agent post directly to your Facebook Pages is a footgun unless you put guardrails in front. Common failure modes I have hit running 5+ pages:

Token expired silently — post fails at midnight, nobody notices for 3 days.
Same angle posted twice in 14 days — audience tunes out, reach drops.
Text-only post when the brand voice mandates an image — engagement craters.
Banned topic leaked — competitor name, internal codename, or deprecated product mentioned.

mcp-fb-publisher ships an MCP server that wraps the Meta Graph API behind 4 deterministic tools, all driven by a single config.yaml. Every publish call goes through validation by default. Validation is pure-Python (no LLM), reproducible in CI, and runs offline.

What it does

4 MCP tools:

| Tool | What it does | |------|--------------| | fb_publish_post | Publishes (or schedules) a post on a configured page. Runs full validation by default; pass skip_validation=True to bypass. | | fb_validate_pre_publish | Dry-run all guardrails. Returns verdict: go|block plus per-check details. Network-optional. | | fb_anti_duplicate_check | Compares a candidate message against the page's recent posts using Jaccard similarity over word 4-grams. | | fb_generate_post_with_image | Generates an image via OpenAI (gpt-image-1) or fal.ai (flux-pro) and returns a URL ready for fb_publish_post. |

5-minute quickstart

# 1. Install
pip install mcp-fb-publisher

# 2. Copy and edit the example config
cp config.example.yaml config.yaml
# -> set page_id values, brand voices, banned_topics

# 3. Set required env
export META_USER_TOKEN="<your long-lived Meta page/user token>"
export MCP_FB_PUBLISHER_CONFIG="$PWD/config.yaml"

# 4. (Optional) for image generation
export OPENAI_API_KEY="sk-..."        # or
export FAL_KEY="..."

# 5. Run the MCP server (stdio transport)
mcp-fb-publisher

Wire it into Claude Desktop / Claude Code

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "fb-publisher": {
      "command": "mcp-fb-publisher",
      "env": {
        "META_USER_TOKEN": "your_token_here",
        "MCP_FB_PUBLISHER_CONFIG": "/absolute/path/to/config.yaml",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Then ask Claude things like:

"Post on the marketing page: 'New collection drops Friday'. Generate an image first, validate, then publish."

Claude will call fb_generate_post_with_image → fb_validate_pre_publish → fb_publish_post.

Architecture

flowchart LR
  A[Claude / MCP client] -->|tool call| B[FastMCP server]
  B --> C{tool}
  C -->|generate| D[OpenAI / fal.ai]
  C -->|validate| E[Validator<br/>pure Python]
  C -->|anti-dup| F[Meta Graph API<br/>fetch_recent_posts]
  C -->|publish| G[Meta Graph API<br/>POST /feed or /photos]
  E -.reads.-> H[(config.yaml<br/>per-page rules)]
  G -.reads.-> H
  F -.reads.-> H

The split between generation, validation and publish is intentional: it lets the LLM iterate on the visual without burning Meta API quota, and it makes the guardrails inspectable in CI without a Meta account.

Anti-duplication strategy

We compare the candidate message against every post within anti_duplicate_lookback_days (per-page, default 14) using Jaccard similarity over word 4-grams:

Normalize: lowercase, strip accents (NFKD), drop URLs, drop punctuation, collapse whitespace.
Build the set of word-level 4-grams for both texts.
similarity = |A ∩ B| / |A ∪ B|
If similarity >= similarity_threshold (default 0.5), block.

Why this and not embeddings: deterministic, free, no extra API key, fast enough for 50 candidates per call. If you want LLM-grade semantic comparison, add another layer on top — the fb_validate_pre_publish tool returns the score so your agent can decide.

Brand voice config

defaults:
  language: en
  brand_voice: |
    Direct, professional, no fluff.
  banned_topics: []
  image_required: true
  anti_duplicate_lookback_days: 14

pages:
  marketing_main:
    page_id: "0000000000000000"
    name: "My Brand — Main"
    brand_voice: |
      Confident, concise, customer-first.
    banned_topics:
      - competitor_brand_a
      - leaked_codename
    image_required: true

  community:
    page_id: "0000000000000002"
    name: "Community"
    image_required: false

Note: the brand_voice field is informational — it's surfaced to the calling LLM via the tool description but the server itself does not LLM-validate against it. This is by design (tests must run offline). Layer your own LLM check on top if you want enforcement.

Use cases

1. Multi-page agency

You manage 5 Facebook pages for clients. Each has its own brand voice, banned topics (competitor names), and image policy. Configure them all in one config.yaml, give Claude the tool, and let the agent draft + validate + publish across all of them with safety rails.

2. Solo founder

You run a single product page and want Claude to schedule the next 30 days of posts. Set image_required: true, give the agent your product brief, and use fb_anti_duplicate_check to make sure no two posts land on the same angle within a fortnight.

3. E-commerce

You rotate flash promos. Set anti_duplicate_lookback_days: 7 for the promo page and 14 for the evergreen content page. Combine with scheduled_at to queue a week's worth of posts in one shot.

Development

git clone https://github.com/anthonyjbolo/mcp-fb-publisher.git
cd mcp-fb-publisher

python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,openai,fal]"

# Run the test suite (offline, no Meta credentials needed)
pytest

# Lint
ruff check .

All tests use httpx.MockTransport and pytest-mock. No real Meta API calls in tests.

Security

The Meta token is read from META_USER_TOKEN only. Never hard-code it.
Token strings are redacted from error messages (REDACTED) before they leave the process.
The validator is sync and offline — safe to run in CI without exposing credentials.
config.yaml is in .gitignore. Only config.example.yaml is committed.

Roadmap

[ ] Instagram Graph API support (the validator already works, only the meta_client publish path needs adapting).
[ ] Optional ntfy webhook on publish failure.
[ ] LLM-grade brand-voice scoring as an opt-in tool.
[ ] Token rotation helper (fb_check_token_expiry).

En français — pourquoi ce projet

Construit en Nouvelle-Calédonie pour gérer 5 pages FB en parallèle (auto-école, marketplace de bingo, atelier d'apps, marque de tee-shirts, page média). Tous les écueils ci-dessus sont des bugs que j'ai vraiment vécus. Le serveur MCP est la couche que j'aurais aimé avoir le premier jour — maintenant elle est libre.

License

MIT — see LICENSE.

Contributing

Issues + PRs welcome. Please run pytest and ruff check . before opening a PR. For substantial features, open an issue first to discuss.

#BuiltOnClaudeCode

Facebook Publisher