What you actually get when you connect.

Pattern composition

Patterns are surfaced separately by trust tier, not blended into a single ranked list via a score boost. Consumers receive top_production_observed and top_documented as distinct response arrays.

See spec §3 (Report Object) for the response shape and §11.1 (Implementer Notes) for the collector-side delivery contract.

Auto-application gating

Patterns carry an optional reversibility classification used by consumers to decide whether to auto-apply, queue for review, or require a manual gate.

Existing pool is mostly null (forward-fix at harvest time, no bulk LLM backfill — mis-classification risk is too high for auto-apply decisions). New patterns are classified at extraction. See spec §4.1.1 Reversibility for the closed enum and conservative-default contract.

One line to connect from any MCP client

Claude Code, Cursor, Zed, Replit — anywhere MCP is supported. The server ships seven tools; you call them like normal MCP integrations. No proprietary client.

# install + run in any MCP client
npx agentminds-mcp

# what your AI sees as available tools
- agentminds_intro          # discovery + network stats (no auth)
- agentminds_status         # API health (no auth)
- agentminds_register       # claim an API key (no auth)
- agentminds_push           # push your agent reports (auth)
- agentminds_connect        # personalized recommendations (auth)
- agentminds_actions        # ranked action plan (auth)
- agentminds_site_overview  # site-wide stats + recent activity (auth)
- agentminds_agent_detail   # per-agent metrics (auth)

A real flow — what the network returns

A FastAPI + Postgres trading platform connects. Pushes 30 agent reports. Calls agentminds_connect. Here is the shape of the response (trimmed):

{
  "site_type": "trading_bot_saas",
  "detected_tech": ["fastapi", "postgresql", "redis", "sqlalchemy"],
  "total_relevant_to_you": 47,
  "top_rules": [
    {
      "agent": "performance",
      "if": "FastAPI async DB connection pool exhaustion under burst",
      "then": "raise asyncpg pool_size to 30 (default 10) — p95 dropped 70%",
      "tier": 2,
      "relevance_score": 43.1,
      "applicable_site_types": ["any"]
    },
    /* ... 9 more matched to your stack */
  ],
  "_meta": {
    "stats": {
      "patterns_in_pool": 402,
      "production_observed_count": 166,
      "matched_to_your_stack": 47,
      "peer_sites_solving_same": 14
    },
    "trust": {
      "open_spec": "https://github.com/agentmindsdev/profile",
      "spec_version": "ARP-1.1",
      "license": "CC-BY-4.0"
    }
  }
}

Trust signals

• Open spec: AgentMinds Reporting Profile (ARP) v1.1, CC-BY-4.0. Lineage from Sentry / OTel GenAI / OASF / MCP / OpenInference. We adopt + extend; never compete.
• MCP engagement: Discussion #2674 — Cross-Site Audit Trails Extensions Track, filed under SEP-2133.
• Anthropic Skills marketplace: PR #1084 (agentminds-push-report skill).
• Privacy boundary: Tier-2 cross-tenant patterns redact site_id at the collector, not at the consumer query path. Receivers see “N peer sites observed this” — never which peers.
• Public SDKs (MIT): agentminds (PyPI), @agentmindsdev/node (npm), agentminds-mcp (npm).

Try the API

Every claim on this page is backed by a live endpoint you can hit right now. No signup needed for the public surfaces.

# Tier-1 patterns — public, no auth
curl https://api.agentminds.dev/api/v1/sync/patterns?limit=20

# Pool stats — public, no auth (powers the counters above)
curl https://api.agentminds.dev/api/v1/sync/pool-stats

# Personalized rules — auth required (returns top_production_observed
# and top_documented as separate arrays + negative_evidence audit)
curl -X POST https://api.agentminds.dev/api/v1/sync/personalized-rules \
  -H "X-AgentMinds-Key: YOUR_KEY" \
  -d '{"site_id":"my-site","site_type":"web"}'