gemini-faf-mcp — The Full-Facts Edition

Persistent Project Context for Google Gemini. Define once. Sync everywhere.

FAF defines. MD instructs. AI codes.

⭐ A star helps other devs discover gemini-faf-mcp — despite the downloads, ~3 of 4 devs check stars first.

Stop re-explaining your project to every new Gemini session. Every Gemini conversation starts cold — you re-state your stack, your goals, your conventions every single time. .faf is one structured file that captures all of it. This package is the MCP server that lets Gemini read it.

Tests

Before and after

Without FAF                           With FAF (.faf at 85%+ Bronze)
─────────────────────────             ─────────────────────────
You: "I'm using FastAPI with...       You: "Add a /users/me endpoint"
      PostgreSQL, pytest, and..."     Gemini: [writes correct code,
Gemini: "Got it. What's the              uses your auth pattern,
        codebase like?"                  matches your test style]
You: "It's a REST API for..."
[5 minutes of re-explaining]
Gemini: [now ready to help]

.faf is read once at session start. Every tool call lands on a Gemini that already knows your project.

What's New in v3.0.0 — The Always33 Edition

One engine, one number: gemini-faf-mcp scores with the always-33 engine, exactly like faf-kernel — the same score faf-cli 8, claude-faf-mcp 7, faf-mcp 4 and grok-faf-mcp 2 give.

Built on faf-python-sdk>=2.0.0. The 12 enterprise slots count unless marked slotignored: 21 slots filled with no markers scores 64% (21/33); the same file plus the 12 markers scores 100% (21/21). faf auto (faf-cli) writes the markers. The 15 faf_model templates carry them and stay 100% Trophy.

v2.8.2 — The Full-Facts Edition

Privacy patch — FAFClient's start-up ping is now opt-in (FAF_TELEMETRY=1).

FAFClient (the Python SDK client — the MCP server never used it) sent a start-up ping, package name and version, by default, and the opt-out wasn't documented. It now sends only when you set FAF_TELEMETRY=1; FAF_TELEMETRY_OFF still turns it off. See faf.one/privacy.

v2.8.1 was a dependency patch — faf-python-sdk floor >=1.4.0, where the interop functions are now named author_agents_md / author_gemini_md. Authored AGENTS.md / GEMINI.md output is byte-identical.

v2.8.0 — faf_auto now grounds its detection in the repo's own files, and every faf_model reference template scores 100% Trophy.

faf_auto used to read only the root manifest (pyproject.toml, package.json, …). It now also reads the files that carry the real stack: docker-compose service images map onto database / cache / search / storage (a running Postgres service is the database — it beats a dependency guess), Makefile / justfile targets map onto the commands block (test / build / check-all, root file or a nested backend/Makefile), and .github/workflows/ sets cicd. A polyglot repo that reported library / JavaScript now reports its real Postgres + Redis + FastAPI stack. In parity with faf-cli 7.10.

Separately: the 15 faf_model reference templates were scoring 48–57% — they filled 4 stack slots and used null. All rewritten to the full 21-slot schema; every one is 100% Trophy now, and a test keeps it that way. 13 tools · 265 tests.

v2.7.1 / v2.7.0 — The Interop Edition — faf_agents / faf_gemini rewritten as faf-python-sdk authoring-tool wrappers (were 4-field stubs); new faf_migrate brings a .faf up to the current format. v2.6.0 — The Agent Card Edition added a real agent.fafa passport, an MCP Server Card (SEP-2127), and an AI Catalog entry. v2.5.0 — The Dart Edition detects Dart/Flutter from pubspec.yaml. v2.4.2 — The Confinement Edition confined every caller path argument. v2.4.0 — The Chameleon Edition auto-selects its transport: stdio locally, Streamable HTTP on Cloud Run.

One-Minute Setup

1. Install

uvx gemini-faf-mcp          # zero-install run via uvx (fetched from PyPI)
# or: pip3 install gemini-faf-mcp

2. Add to Gemini CLI

gemini extensions install https://github.com/Wolfe-Jam/gemini-faf-mcp

3. Author your project context

In your Gemini CLI:

> /faf:setup

You should see: Created project.faf — Score: 85% (BRONZE). From this point, every Gemini session in this project reads it automatically.

Tip: A score of 85% (BRONZE) is the minimum where Gemini stops guessing. Run /faf:score to see what's missing and how to push to 100% (TROPHY).

The "One-File" Advantage

A .faf file is structured YAML that captures your project DNA. Every AI agent reads it once and knows exactly what you're building.

# project.faf — your project, machine-readable
faf_version: "3.0"
project:
  name: my-api
  goal: REST API for user management
  main_language: Python
stack:
  backend: FastAPI
  database: PostgreSQL
  testing: pytest
human_context:
  who: Backend developers
  what: User CRUD with auth
  why: Replace legacy PHP service

Result: Gemini reads this once and knows your project. No 20-minute onboarding. No wrong assumptions. Every session starts aligned.

FAF defines. MD instructs. AI codes.

What about my GEMINI.md?

You don't replace it. .faf authors it. Run faf_gemini and you get a fresh GEMINI.md in Gemini CLI's own hierarchical, @file-importable convention — setup, verify, key files, stack, confirm-first actions — authored from a single source of truth instead of hand-maintained. Your hand-written content outside the faf-managed block is preserved.

> /faf:export
# Authors GEMINI.md from project.faf

.faf is the source. GEMINI.md is one of its outputs. Same logic for AGENTS.md (OpenAI Codex), .cursorrules, CLAUDE.md, and others — write once, render everywhere.

Auto-Detect Your Stack

faf_auto scans your project's manifest files and its docker-compose services, Makefile targets, and CI config, then authors a .faf with accurate slot values. No manual entry needed.

> Auto-detect my project stack

{
  "detected": {
    "main_language": "Python",
    "package_manager": "uv",
    "framework": "FastAPI",
    "api_type": "REST",
    "database": "PostgreSQL",
    "cache": "Redis",
    "hosting": "Docker Compose",
    "cicd": "GitHub Actions",
    "commands": { "test": "make test", "build": "make build", "lint": "make check-all" }
  },
  "score": 79,
  "tier": "GREEN"
}

database and cache came from docker-compose.yml, commands from the Makefile, cicd from .github/workflows/ — none of which the manifest scan sees. Fill in the six W's and you are at Trophy.

What it scans:

Priority rule: pyproject.toml / Cargo.toml / go.mod take priority over package.json. File-facts (a real compose service, a Makefile target) win over dependency guesses. Only sets values that are actually detected — no hardcoded defaults.

All 13 Tools

Create & Detect

Validate & Score

Read & Transform

Export & Interop

Migrate

Reference

Score and Tier System

Your .faf file is scored on completeness — how many slots are filled with real values.

Aim for Bronze (85%+). That's where AI stops guessing and starts knowing.

Using with Gemini CLI

> Create a .faf file for my Python FastAPI project
> Auto-detect my project and fill in the stack
> Score my .faf and show what's missing
> Export GEMINI.md for this project
> Show me a 100% example for an MCP server
> What is FAF and how does it work?
> Read my project.faf and summarize the stack
> Validate my .faf and fix the warnings
> Migrate my project.faf to the current format

Architecture

gemini-faf-mcp v3.0.0
├── server.py              → FastMCP MCP server (13 tools, dual-transport, Mk4 scoring)
├── safe_path.py           → path confinement for caller-supplied `path` args
├── inject.py              → non-destructive faf-managed-block injection
├── interrogate.py         → Full-Facts grounding (docker-compose + Makefile signals)
├── main.py                → Cloud Run REST API (GET/POST/PUT)
├── models.py              → 15 project type examples
└── src/gemini_faf_mcp/    → Python SDK (FAFClient, parser)

The MCP server delegates to faf-python-sdk for parsing, validation, Mk4 scoring, and AGENTS.md / GEMINI.md authoring. Stack detection in faf_auto — including the Full-Facts grounding — is Python-native, no external CLI dependencies.

Testing

pip3 install -e ".[dev]"
python -m pytest tests/ -v

266 tests passing across the FastMCP server, Cloud Function, Mk4 WJTTC championship, Full-Facts grounding, path-confinement, write-guard, model-parity, Dart detection, and client-telemetry suites. Championship-grade test coverage — WJTTC certified.

FAF Ecosystem

One format, every AI platform.

Python SDK

Use FAF directly in Python without MCP:

from gemini_faf_mcp import FAFClient, parse_faf, validate_faf, find_faf_file

# Parse and validate locally
data = parse_faf("project.faf")
result = validate_faf(data)
print(f"Score: {result['score']}%, Tier: {result['tier']}")

# Find .faf files automatically
faf_path = find_faf_file(".")

# Or use the Cloud Run endpoint
client = FAFClient()
dna = client.get_project_dna()

FAFClient sends no telemetry unless you set FAF_TELEMETRY=1, which adds one start-up ping (package name and version). FAF_TELEMETRY_OFF always turns it off. Remote mode sends your requests to the Cloud Run endpoint. Privacy: faf.one/privacy.

Cloud Run REST API

Live endpoint for badges, multi-agent context brokering, and voice-to-FAF mutations.

https://faf-source-of-truth-631316210911.us-east1.run.app

Supports agent-optimized responses (Gemini, Claude, Grok, Jules, Codex/Copilot/Cursor) via X-FAF-Agent header. Voice mutations via Gemini Live through PUT endpoint. Auto-deploys via Cloud Build on push to main.

If gemini-faf-mcp has been useful, consider starring the repo — it helps others find it.

Links

• FAF Specification
• IANA Registration
• faf-python-sdk
• faf-cli
• Changelog

Citation

If you use gemini-faf-mcp or the .faf / .fafm / .fafa formats in research or production, please cite the format papers:

Wolfe, J. (2025). Format-Driven AI Context Architecture: The .faf Standard for Persistent Project Understanding. Zenodo. https://doi.org/10.5281/zenodo.18251362

Wolfe, J. (2026). Permanent Memory and Instant Recall: The .fafm Standard for Multi-Profile AI Agent Memory. Zenodo. https://doi.org/10.5281/zenodo.20348942

Wolfe, J. (2026). Why Agents Need a Passport: .fafa — Portable Identity for the Agentic Era. Zenodo. https://doi.org/10.5281/zenodo.21951641

BibTeX

@article{wolfe2025faf,
  title     = {Format-Driven AI Context Architecture: The .faf Standard for Persistent Project Understanding},
  author    = {Wolfe, James},
  year      = {2025},
  month     = {nov},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.18251362},
  url       = {https://doi.org/10.5281/zenodo.18251362}
}

@article{wolfe2026fafm,
  title     = {Permanent Memory and Instant Recall: The .fafm Standard for Multi-Profile AI Agent Memory},
  author    = {Wolfe, James},
  year      = {2026},
  month     = {may},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20348942},
  url       = {https://doi.org/10.5281/zenodo.20348942}
}

@article{wolfe2026fafa,
  title     = {Why Agents Need a Passport: .fafa — Portable Identity for the Agentic Era},
  author    = {Wolfe, James},
  year      = {2026},
  month     = {aug},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21951641},
  url       = {https://doi.org/10.5281/zenodo.21951641}
}

License

MIT

Built by @wolfe_jam | wolfejam.dev

Get the CLI

faf-cli — The original AI-Context CLI. A must-have for every builder.

npx faf-cli auto

Anthropic MCP #2759 · IANA Registered: application/vnd.faf+yaml · faf.one · npm