mcp-server

The Problem

Your AI assistant knows nothing about your internal services. Every time you ask "which teams own the payment service?" or "what breaks if I take down the DB?", it either hallucinates or burns tokens scanning dozens of repos.

DocScout-MCP solves this by pre-computing the answer graph and serving it deterministically over MCP.

How It Works

graph LR
    GH["GitHub Org\n(repos, manifests, docs)"]
    S["Scanner\n(concurrent, retry-safe)"]
    P["Parsers\ngo.mod · pom.xml · package.json\nCODEOWNERS · catalog-info.yaml\nDockerfile · Helm · Terraform · OpenAPI"]
    G["Knowledge Graph\nSQLite · PostgreSQL"]
    AI["AI Clients\nClaude · Cursor · Copilot · Gemini"]

    GH -->|"GitHub API + Webhooks"| S
    S --> P
    P -->|"entities + relations"| G
    G -->|"26 MCP tools"| AI

1. Scan — Crawls every repo in your org: docs, manifests, infra files, and root tooling files. Repeats on a configurable interval and reacts to GitHub webhooks for instant updates.
2. Parse — Extracts services, owners, dependencies, and relations from go.mod, pom.xml, package.json, CODEOWNERS, catalog-info.yaml, MCP config files, and more.
3. Graph — Persists everything as entities and relations in SQLite or PostgreSQL, surviving restarts.
4. Answer — AI clients query the graph via 26 MCP tools. No file-reading loops, no token waste, no guessing.

Why DocScout?

DocScout pre-computes the answer graph from your repos so your AI never reads files to answer architecture questions. See benchmark/RESULTS.md for methodology.

See It In Action

"What happens if I shut down component:db? Which systems go offline, and who do I notify?"

→ search_nodes("component:db")
  Found: component:db — incoming edge: payment-service depends_on

→ open_nodes(["payment-service"])
  Entity: payment-service (service)
  Observations: _source:go.mod, go_version:1.26, _scan_repo:myorg/payment-service

→ search_nodes("payments-team")
  Entity: payments-team (team)
  Observations: github_handle:@myorg/payments-team
  Relations: payments-team → owns → payment-service

Claude: "Shutting down component:db will impact payment-service.
         Notify @myorg/payments-team. No other services have a direct dependency."

The AI answers from verified graph facts — not file naming conventions or guesses.

Quick Start

1. Get a Fine-Grained GitHub PAT

Go to GitHub → Settings → Developer Settings → Fine-grained tokens. Grant Read-only access to Contents and Metadata for your org's repositories.

2. Add to Your AI Client

Claude CLI (recommended):

claude mcp add --transport stdio \
  --env GITHUB_TOKEN=github_pat_... \
  --env GITHUB_ORG=my-org \
  docscout-mcp -- go run github.com/doc-scout/mcp-server/cmd/docscout@latest

Or build and run locally:

git clone https://github.com/doc-scout/mcp-server
cd mcp-server

GITHUB_TOKEN="github_pat_..." GITHUB_ORG="my-org" go run ./cmd/docscout/

Docker:

docker run -i \
  -e GITHUB_TOKEN="github_pat_..." \
  -e GITHUB_ORG="my-org" \
  ghcr.io/doc-scout/mcp-server:latest

3. Ask Away

"Which services depend on the billing library?" "Who owns the checkout service?" "List all repos with a Helm chart." "What Go services have direct dependencies on pgx?"

MCP Tools (23)

What Gets Scanned

Root-level manifests (extracted into the knowledge graph):

Recursive directories: docs/ and .agents/ (.md files) · deploy/, infra/, .github/workflows/ (Helm, Terraform, K8s, workflows)

Key Configuration

See full environment variable reference for all options including SCAN_FILES, SCAN_DIRS, REPO_TOPICS, REPO_REGEX, EXTRA_REPOS, and more.

AI Client Setup

Architecture & Security

• Path-traversal protection: Only files verified by the scanner are accessible. The AI cannot read arbitrary files.
• STDIO safety: No text is ever written to stdout. All logs go to stderr. Corruption of the JSON-RPC stream is impossible by design.
• Rate limit resilience: Every GitHub API call uses exponential backoff with smart Retry-After handling.
• Graph integrity: Observations are sanitized before storage. Mass deletions (> 10 entities) require explicit confirmation.
• Audit log: Every graph mutation emits a structured slog line to stderr.

For a deep dive, see How It Works.

Roadmap

See ROADMAP.md for completed features and upcoming work, including:

• Semantic Search & RAG — vector embeddings via pgvector
• Custom Parser Extensions — plug in new manifest formats without forking
• Integration Topology Discovery — Kafka, gRPC, HTTP call graph from config files
• Multi-Cloud Adapters — GitLab, Bitbucket, Confluence
• Documentation Wiki (gh-pages) — move the detailed guides to a dedicated GitHub Pages site

Contributing

# Install dependencies
go mod tidy

# Build
go build -o docscout-mcp .

# Test (unit + E2E integration)
go test ./...

Review the Development Guidelines and AGENTS.md before submitting a PR.

License

GNU AGPL v3

Disclaimer

This software is provided "as is", without warranty of any kind. AI-generated output depends on indexed repository data — always verify before acting on it. See DISCLAIMER.md for full details.