agentminds-mcp

MCP server for AgentMinds — cross-site collective intelligence for production AI agents. Pull patterns from the network, push your agent reports, get personalised recommendations matched to your stack. No signup needed for the trial.

Try it now (30 seconds, no API key)

npx agentminds-mcp

Then call agentminds_connect from any MCP-aware client. You'll get top production-observed patterns from the network — no registration required, no daily cap.

How it works — three access shapes

AgentMinds is free for everyone. There are no tiers and no upgrade between modes — just three different ways to use the same pool.

| Mode | What you give | What you get | |---|---|---| | Anonymous | nothing | Top patterns from the public pool | | Registered | URL + name (run agentminds_register) | Stack-matched personalised recommendations | | Push back | agent reports (run agentminds_push) | Pool grows for everyone; cross-site references surface when matches exist |

The backend auto-routes between modes based on your auth state. agentminds_connect returns the richest content available given your current mode. Pushing is optional and never required to keep pulling.

Why not just ask ChatGPT or Claude?

Fair question. Large language models are excellent for general AI agent development questions, and you should keep using them. But they have three blind spots that AgentMinds fills:

1. Real-time production data

LLM training cutoffs are months behind. The vLLM threading bug that broke production agents in March? Claude can describe it now (after the cutoff caught up), but couldn't have warned you then. AgentMinds patterns include observations from sites that hit the bug the day it shipped.

2. Cross-site private knowledge

Patterns learned inside private codebases never reach public training data. A FastAPI + Pydantic + Claude SDK failure that one team solved in their internal monorepo — that knowledge stops there. AgentMinds' opt-in network shares it safely (URLs anonymised, push is explicit, GDPR-compliant).

3. Quantified pattern data

Claude can suggest a fix. AgentMinds can tell you:

• 14 sites tried this fix
• 9 solved it
• 5 it didn't (and why — negative_evidence)
• Average resolution time: 12 minutes
• Reversibility: safe_config (no rollback risk)

That's production data, not training data.

Numbers above are illustrative for the format. Live counts vary by pattern fingerprint and current network state — see /sync/pool-stats.

Use both

We're complementary, not competitive:

• General agent development questions → Claude / Gemini / ChatGPT
• "What worked for someone with my exact stack in production" → AgentMinds

The MCP server makes both available in your terminal. Most users ask Claude first, then call agentminds_connect to verify against production patterns before shipping.

Install

Claude Code

claude mcp add agentminds -- npx agentminds-mcp

Or add manually to ~/.claude/mcp.json:

{
  "mcpServers": {
    "agentminds": {
      "command": "npx",
      "args": ["agentminds-mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "agentminds": {
      "command": "npx",
      "args": ["agentminds-mcp"]
    }
  }
}

Other MCP clients

Any client following the MCP spec works. Spawn npx agentminds-mcp over stdio.

Tools (8)

| Tool | Auth | What it does | |---|---|---| | agentminds_intro | None | Onboarding overview + live network stats. Call this first if unsure. | | agentminds_status | None | Backend health (/health): up/down, last pipeline, open circuits. | | agentminds_connect | Optional | Tier-aware pull: anonymous trial / registered no-push / personalised. The main value tool. | | agentminds_register | None | Create a site, receive an API key. Saves to .agentminds.json in cwd. | | agentminds_push | Required | Submit agent reports (severity, summary, metrics, warnings, learned_patterns). Returns server-graded data quality. | | agentminds_actions | Required | Personalised action plan for your site. | | agentminds_agent_detail | Required | Inspect a specific agent (metrics, warnings, patterns). | | agentminds_site_overview | Required | Dashboard view of all your agents and their status. |

Configuration

AGENTMINDS_API_KEY=sk_...                      # required for push + authed tools
AGENTMINDS_API_URL=https://api.agentminds.dev  # default

The server also auto-reads .agentminds.json and .env from the calling project's cwd if AGENTMINDS_API_KEY is unset:

{
  "site_id": "yoursite",
  "api_key": "sk_yoursite_...",
  "site_url": "https://yoursite.com"
}

Privacy

• Anonymous trial: no payload sent — only your IP is used for the 3/day rate limit (in-memory at the backend, not logged per-request).
• Registered: the URL + name you pass to agentminds_register are stored. No telemetry beyond that.
• Push: agent reports you submit are stored in the pool. You control the content — anonymise before sending if needed. The backend strips site identity before reports are surfaced to other sites' personalised flows.
• No analytics, no tracking. The MCP server makes HTTP calls only when you explicitly invoke a tool.

Honest status (2026-05-11)

This is early-stage. AgentMinds is free for everyone — there are no tiers, no paywalls, no upgrade path. Pull what you need, push what you can. Live numbers:

| Metric | Value | |---|---| | Contributing sites (active) | 6 | | Production-observed patterns | 3,233 | | Documented patterns | 702 | | Total tier-1 patterns | 3,983 |

The cross-site "peer sites solving the same problem" feature activates as the network grows. Today most patterns come from the external harvester (public GitHub issues, MCP corpora, awesome lists) rather than peer sites — the personalised flow surfaces them with stack-matching, but the network-effect moat is still forming.

If you're evaluating this for your team: the ARP spec is the most mature surface (formally versioned at v1.3.0, with extension points and a reorientation clause explicitly telling readers to prefer OpenTelemetry GenAI / MCP when those cover your need). The MCP server and SDKs are v1.3.x — actively iterated, may have rough edges. Bug reports welcome.

Lineage

ARP is a profile built on top of OpenTelemetry GenAI semantic conventions, MCP, Sentry-style runtime ergonomics, Anthropic Claude Skills, and AGNTCY OASF. The single primitive AgentMinds owns is the cross-site learned-pattern lifecycle — see AGENT_REPORTING_PROFILE.md §4.1.

Resources

• Site: https://agentminds.dev
• Spec (ARP v1.3.0): https://github.com/agentmindsdev/profile
• API base: https://api.agentminds.dev
• Public pool stats: https://api.agentminds.dev/api/v1/sync/pool-stats
• Issues: https://github.com/agentmindsdev/mcp-server/issues
• Changelog: CHANGELOG.md

License

MIT.