Git.Top

Git.Top

The Knowledge Graph of Open Source.

GitHub provides code. Git.Top provides knowledge.

Git.Top turns unstructured GitHub repository data into structured project knowledge that humans and AI agents can search, compare, recommend, score, and call through REST APIs or MCP tools.

The Cloudflare Worker is the only product surface. It serves the human HTML pages, REST APIs, MCP endpoint, GitHub sync, and D1-backed knowledge data.

Production: git.top

Repository: github.com/haocn-ops/git-top

Connect An Agent

Use the focused MCP profile for the shortest first connection:

codex mcp add git-top --url https://git.top/mcp/coreclaude mcp add --transport http --scope user git-top https://git.top/mcp/core

Then ask your agent:

Use Git.Top to recommend three open-source browser-agent projects for Docker, then cite the data source and one caveat for each.

Full setup, client status, and REST alternatives: git.top/connect.

Dated compatibility evidence and the distinction between configuration verification and real-client support: git.top/compatibility.

Registry-ready copy, profiles, privacy boundaries, campaign links, and submission status: git.top/distribution.json. An installable REST-first Agent Skill is available at skills/git-top-project-selection.

What Git.Top Does

• Discover open-source AI projects by category, deployment target, and use case.
• Compare projects with decision summaries, ranking signals, and tradeoffs.
• Find alternatives with similarity signals and comparison-ready shortlists.
• Explore project relationships across alternatives, related projects, dependencies, deployments, and use cases.
• Explain Git.Top Score, Agent Score, quality signals, risk flags, and evidence.
• Serve agent-native REST, OpenAPI, llms.txt, llms-full.txt, and MCP surfaces.

Product Surfaces

• /discover - browse AI agents, MCP servers, RAG, browser automation, and AI IDE projects.
• /recommend - turn use cases and constraints into explainable project shortlists.
• /compare - compare project shortlists.
• /alternatives and /alternatives/:project - find replacement candidates.
• /graph/:project - inspect project knowledge graphs.
• /score/:project - explain project score dimensions.
• /atlas - explore ecosystem maps.
• /api/* - consume structured knowledge.
• /mcp - connect agents through MCP.
• /mcp/core - connect a focused five-tool MCP profile for first-time use.

Architecture

GitHub API  -> Git.Top ETL  -> Cloudflare D1  -> Worker pages, REST APIs, MCP tools, sync jobs, and quality gates

The project intentionally does not ship a separate Next.js app. Keeping Worker-rendered pages, APIs, sync, D1 access, OpenAPI, and MCP in one runtime avoids split product behavior.

Documentation

• Coding agent guide
• AI agent adoption and distribution plan 2026-07-31
• Agent client compatibility
• Agent distribution package
• Agent-native assessment and optimization plan
• Competitive analysis and 90-day optimization plan 2026-09-02
• Agent function test and improvement plan 2026-07-30
• Production freshness optimization plan 2026-07-14
• Site assessment and improvement plan 2026-07-12
• Automated operations and agent friendliness plan 2026-07-12
• Agent task evaluation
• Agent-native optimization implementation report
• Agent-native optimization PR summary
• Git.Top 2.0 product upgrade plan
• Next stage plan
• Agent friendliness optimization plan
• Deployment decision
• REST API guide
• MCP guide
• MCP tool behavior examples
• MCP conformance matrix
• SDK-oriented examples
• Adoption analytics runbook
• GRP examples
• Agent quickstart
• MailAgents agent workflow article
• Site assessment 2026-06-21
• Production runbook
• Operations and data governance plan
• Public trust benchmark: /benchmark and /api/benchmark
• Data coverage report
• Seed live check report
• Eval quality report
• Local eval report
• Ranking experiments
• Quality hardening plan
• Post-V1 improvement plan
• Contributing
• Security
• License

Development

For maintenance agents, start with AGENTS.md. It summarizes the runtime, high-risk areas, generated data, and validation tiers.

Fast checks:

pnpm docs:validatepnpm check

Agent surface checks:

pnpm api:validatepnpm mcp:validatepnpm eval:explanationspnpm check

Data and local D1 checks:

pnpm db:executepnpm db:seedpnpm db:integration

Release checks:

pnpm validatepnpm release:check

Common development commands:

pnpm installpnpm db:executepnpm db:seedpnpm seed:validatepnpm seed:coveragepnpm seed:candidatespnpm seed:live-check -- --limit 20pnpm seed:live-check -- --offset 20 --limit 20node scripts/run-governance-task.mjs daily-production-healthpnpm knowledge:validatepnpm db:seed-sqlpnpm db:validatepnpm db:integrationpnpm core:validatepnpm api:validatepnpm mcp:validatepnpm eval:qualitypnpm eval:explanationspnpm eval:localpnpm eval:rankingpnpm eval:agent-taskspnpm smoke:prodpnpm release:checkpnpm checkpnpm validatepnpm quality:checkpnpm dev

/api/health reports D1 availability and the current project count. Run pnpm db:execute and pnpm db:seed before local API checks when the D1 state is empty. pnpm db:execute prepares the local D1 schema and backfills optional columns that may be missing from older local databases. seed.sql is generated from hand-authored seed knowledge and generated eval fixtures; run pnpm db:seed-sql after changing those fixtures. pnpm db:integration seeds local D1, starts a temporary local Worker, and validates the D1-backed HTTP API path.

pnpm eval:quality is the CI-safe recommendation and classification regression gate. pnpm eval:explanations checks that agent-facing responses include source metadata, classification evidence, quality signal confidence, recommendation reasons/tradeoffs, health count semantics, and GRP reasoning. pnpm eval:agent-tasks gates complete trust, fallback, pagination, evidence, comparison, change-feed, feedback, multilingual, typo, and alias workflows and writes docs/AGENT_TASK_EVAL.md. pnpm eval:local runs broader generated category and deployment probes across the fixture-backed project set and writes docs/EVAL_LOCAL.md; use it before tuning ranking heuristics, but keep it out of the default validation path. pnpm eval:ranking compares offline ranking strategies in docs/RANKING_EXPERIMENTS.md. Runtime search keeps exact-intent ranking by default; broad scoped discovery can opt into ranking=browse.

pnpm seed:candidates discovers seed expansion candidates from live GitHub organization metadata and writes docs/SEED_CANDIDATES.md. Treat it as a review queue, not an automatic append step; candidates still need live-check and category review before entering data/seed-repositories.json.

pnpm smoke:prod validates the deployed Worker at https://git.top and requires D1-backed responses. Use pnpm smoke:prod -- --base-url http://localhost:8787 for a local or preview Worker, or add --allow-seed only when intentionally checking seed fallback behavior.

pnpm quality:check validates the production /api/quality endpoint at https://git.top by default, requires D1-backed metadata, and uses a default minimum score of 90. Use --base-url, --target, --min-score, or --allow-seed for preview and fallback checks.

Operations and data governance are exposed through /operations, /api/governance/summary, and /api/governance/runs. The Cloudflare Worker cron runs hourly maintenance plus daily, weekly, biweekly, and monthly governance checks, then records results in governance_runs.

pnpm release:check runs the public V1 release gate: local validation, local D1 integration, production quality, and production smoke. Use pnpm release:check -- --skip-prod-smoke only when validating a build before the production deployment exists; it skips production-only checks. Use pnpm release:check -- --base-url <origin> to run the same gate against a Worker preview or local origin.

When updating an existing D1 database, apply SQL files in migrations before deploying code that reads the new columns.

Sync

Git.Top can sync real repository data from the GitHub API into D1.

Required production secrets:

wrangler secret put GITHUB_TOKENwrangler secret put SYNC_SECRET

For local development, copy .dev.vars.example to .dev.vars and fill in real values.

Trigger a protected local or production sync:

curl -X POST http://localhost:8787/api/admin/sync \  -H "authorization: Bearer $SYNC_SECRET" \  -H "content-type: application/json" \  -d '{"limit": 40, "offset": 0, "signal_depth": "lite"}'

Use {"limit":40,"signal_depth":"lite"} for manual catch-up syncs after checking recent /api/sync/status runs. Cron uses a smaller lightweight batch size to stay under Worker subrequest limits, while manual catch-up remains capped at 50.

Run repeated production catch-up rounds:

SYNC_SECRET=... pnpm sync:prod:catchup --rounds 13 --limit 40

Omit offset to use the stored seed cursor. Cron syncs use this cursor and advance through the seed list in bounded batches. Add --refresh-cycle when all seed repositories have been indexed but freshness recovery requires revisiting the full seed cursor despite remaining_count=0.

API

• GET /api/project/:owner/:name
• GET /api/health
• GET /api/quality
• GET /api/benchmark
• GET /api/governance/summary
• GET /api/governance/runs
• GET /api/sync/status
• GET /api/schema/agent-card.v1
• GET /api/schema/project-knowledge.v1
• GET /api/search
• GET /api/trending
• GET /api/category/:name
• GET /api/recommend
• GET /api/compare
• GET /api/alternatives/:owner/:name
• POST /api/grp/query
• POST /api/admin/sync
• POST /api/admin/alternatives
• POST /api/admin/governance/runs

See git_top_v1.md for the V1 development spec and git_top_grp_v1.md for the GRP v1 reasoning protocol implementation spec.

License

MIT