What one site learns,
every connected site uses.
AgentMinds is an early-access private pool of patterns, knowledge, and functions — connected sites contribute, the pool grows. Connect once — your code can search the pool, and rule rankings get matched to your stack, your routes, your agents. Anonymized in, ranked out.
claude mcp add agentminds -- npx agentminds-mcppip install agentmindspython -m agentminds connectnpm install @agentmindsdev/nodenpx @agentmindsdev/node connectPulling is free, unlimited, and never requires you to give anything back. If your agents observe useful patterns — a fix that worked, a configuration that broke production — push them back voluntarily with agentminds_push. Every voluntary contribution helps the pool grow for everyone, including you. Pushing is always optional and never gates what you can pull.
npx agentminds-mcpGet top production-observed patterns from the network in your terminal. 30 seconds, no API key, no card. (no daily cap.)
How it's different
Three pictures. One for the flow, one for what makes us not-Sentry, one for what stays private vs public.
Many sites push reports. The pool fingerprints, dedups, learns. Each site gets back only what matches their stack.
Confidence ↑ · status tracked
Sentry tells you what broke on YOUR site. AgentMinds tells you what broke on yours, who else hit it, and how they fixed it.
Useful: you see your error count + stack trace. Limit: you only know about your site. The same error has hit dozens of others — you have no signal.
+ N other sites hit this
+ M solved with `pool.dispose()`
+ confidence:
k/(N+2)Same error, but you also see how peer sites solved it, ranked by Beta-Bernoulli confidence. Numbers above are an illustrated example — the network is early-access and the real values you see depend on what other connected sites have observed and resolved.
We publish how we work. We do not publish what we learned about your competitors.
- · ARP profile spec + JSON Schema (CC-BY-4.0)
- · Tier-1 generic web hygiene rules (12 sample)
- · Aggregate network metrics (sites count, scan count, top issues)
- · A2A AgentCard + OASF descriptor + MCP server card
- · Standards-watcher reorientation log
- · Tier-2 site-specific learned patterns
- · Cross-site recommendations matched to YOUR stack
- · Pattern confidence + drift signals + status lifecycle
- · Network position vs sites with similar stacks
- · Beta-Bernoulli posterior on each pattern
This is the moat. We never expose it via public/no-auth API. Read the spec to see exactly which API surfaces are gated.
What you do
Three steps. The first one is one command; the other two happen automatically.
Install + connect
One command. Registers your site, builds your DSN, edits your entry file in place.
pip install agentminds && python -m agentminds connect
Your agents push
Errors, learned patterns, runtime telemetry — anonymized — flow into the pool automatically. No extra code.
# nothing to do — agentminds.init() handles it
Your code pulls
Query the pool any time. Tier-1 rules are public; tier-2 (derived from solved patterns at peer sites) is gated behind connect.
GET /api/v1/sync/playbook?agent=health # or category=performance, stack=fastapi
Free public-surface scan — no signup
Security headers, SEO meta, perf metrics, tech stack inferred from public assets. 30 seconds.
Scan
Free instant analysis. Real headless browser, real headers, no fake heuristics.
Connect
Push agent reports via API or Claude Code. Everything anonymized.
Learn
Pull patterns and fixes from the cross-site pool — including ones from sites running similar stacks.
The playbook is just the start.
Connecting unlocks features that can't exist without your metrics flowing into the network.
Personalized rules
Rules re-ranked for your exact tech stack + site type. A Next.js SaaS gets different top-20 than a Unity game.
GET /sync/personalized-rulesYour benchmarks vs the network
Your LCP is 2,100ms. Network p50 is 1,400ms — bottom third. Every metric compared automatically.
GET /sync/benchmarks/{site_id}Real-time auto-fixes
When a new critical bug appears in the network, you find out immediately — with the tested fix, not a description.
POST /connect (returns auto_fixes)Outcome tracking
Apply a fix, report the outcome. Your confidence score feeds back into the network — everyone gets smarter.
POST /feedbackConnect from anywhere
REST API, Claude Code MCP, CLI, VS Code extension. Pick your tool.
# 1. Free scan — no signup
curl -X POST https://agentforge-20ng.onrender.com/api/v1/free-scan \
-d '{"url":"https://yoursite.com"}'
# 2. Register and get an API key
curl -X POST https://agentforge-20ng.onrender.com/api/v1/sync/onboard \
-d '{"url":"https://yoursite.com","name":"My Site"}'
# 3. Connect — push data, receive tips
curl -X POST https://agentforge-20ng.onrender.com/api/v1/connect \
-H "X-AgentMinds-Key: sk_..." -d '{}'Latest from the blog
All posts →Production Deployment Patterns That Prevent Cascading Failures on a Shared VPS
We lost eight days to a silent cron failure. Here’s how disk exhaustion, concurrency races, and database pool drains taught us to harden our deployment pipeline.
When Our AI Pipeline Went Silent for 8 Days: Hardening a $6 VPS with Five Production Patterns
A cascade of silent failures—cron death, log overflow, parallel job corruption, UFW Docker conflicts, and tenant isolation breaches—crippled our AI monitoring pipeline. We dissect each failure and the concrete patterns that now keep our $6 VPS self-healing.
Cross-Site Agent Intelligence: Why We Built the ARP Profile
Production AI agent teams learn the same lessons in isolation. The AgentMinds Reporting Profile (ARP) is the open wire format that lets cross-site pattern fingerprints + lifecycles travel safely between organisations — without leaking customer data. This is the long-form mission post.
Early-stage but actively used
The community shaping the spec is part of what AgentMinds is. Below: real public discussions, threads, and issues from external engineers — not testimonials we curated.
PBN-compatible data_class for Pattern primitive
External developer extending the spec with PBN (data-flow enforcement) compatibility. Active design feedback shaping ARP v1.4.
Cross-site pattern pool for production AI agent failures
Reddit thread with peer technical feedback that informed v1.3 design (averageuser612 trust-layer notes, eval-set request).
Eval Set v0 strawman — delta welcome
Public eval set proposal with the request for community deltas. Independent benchmarks part of the project's transparency commitment.
Common questions
Real questions we've heard. Honest answers.
+How is this different from ChatGPT or Claude?
Large language models like Claude, Gemini, and ChatGPT are excellent for general questions about AI agent development. But three things they can't do today:
1. Real-time production data. LLM training cutoffs are months behind. New library versions, fresh failure patterns — the model doesn't know yet. Patterns we surface include observations from sites that hit a bug the day it shipped.
2. Cross-site private knowledge. Patterns learned inside private codebases never reach public training data. AgentMinds' opt-in network shares them 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, average resolution time 12 minutes, reversibility safe_config.' That's production data, not training data.
We're complementary, not competitive. Use Claude for general questions; use AgentMinds when you need to know what *actually worked* in production for someone with your exact stack.
+What's the catch? Why is it free?
There is no catch. AgentMinds is a collective-intelligence pool — the value grows with the number of contributors. Paywalls would gate that growth, so we removed them entirely.
Pull what you need, push what you can. Pushing is optional, never required to keep pulling. Cross-site `learned_patterns` are still private to authenticated callers (per-site personalised delivery only), so opening the pool doesn't expose what other sites contributed.
If you want to support the project, the most valuable thing you can do is connect your site and push useful agent reports back to the pool.
+Can I try without committing to anything?
Yes. Run `npx agentminds-mcp` in your terminal. 30 seconds, no signup, no card, no API key. You get top production-observed patterns from the network, pull as much as you want.
If it is useful, registering takes 30 seconds (URL + name) and you start getting stack-matched personalised recommendations on the first call. No upgrade between modes — they are all free.
We believe in pull-first: you should know if AgentMinds gives you value before you give us anything.
+What about privacy? You're collecting agent reports, right?
Anonymous trial: nothing leaves your machine other than the IP used for rate limiting. No payload, no tracking pixel.
Registered: only the URL + site name you provide.
Pushed: agent reports you submit — and you control the content. The push tool sends what you put in the payload, nothing else. No code analysis, no implicit telemetry, no third-party tracking.
URLs in patterns are stripped before they reach the cross-site pool other sites read. Cross-site `learned_patterns` are private to the network — never exposed via no-auth API. Per-site personalised delivery only.
+Why should I trust this with only 6 sites in the network?
You shouldn't trust based on size. We're early-stage and we say so — 6 active contributing sites, 3,983 patterns as of this writing.
What earns trust here isn't size; it's the spec. ARP v1.3.0 is published openly at [github.com/agentmindsdev/profile](https://github.com/agentmindsdev/profile), formally versioned with extension points, and includes a [reorientation clause](https://github.com/agentmindsdev/profile#reorientation-clause): if OpenTelemetry GenAI semantic conventions cover your case, use that instead.
We're not trying to lock you in. We're trying to build a profile that outlasts us.
+What languages and frameworks do you support?
Official SDKs:
- Python 3.8+ — auto-instruments FastAPI, Flask, Django (manual capture API also available)
- Node.js 14.17+ — auto-instruments Express, Fastify, Next.js (logger integrations for Winston, Pino)
Any other language can use the HTTP API directly — `POST /api/v1/sync/bulk` with your agent reports. The full API contract is documented at [/docs](/docs).
MCP server (`agentminds-mcp` on npm) works with any MCP-compatible client: Claude Code, Cursor, Continue, Cline, Zed, etc.
Patterns of the week → your inbox
Every Monday: the 3 most-pulled fix patterns from the network, one new case study, no fluff. Unsubscribe anytime.
Ready to make your site smarter?
Join the collective intelligence network. What you share makes everyone's site stronger.
No credit card · No signup · Results in 30 seconds