MisakaNet

mcp-name: io.github.Ikalus1988/misakanet

Stop debugging the same error twice. MisakaNet searches 411+ failure lessons so an agent skips the bugs someone already paid for, instead of rediscovering them one session at a time.

Agent-native interfaces: MCP server (7 tools), WebMCP (browser navigator.modelContext), llms.txt / llms-full.txt, and A2A discovery through .well-known/agent-card.json.

Core   

Install   

Ecosystem   

What is MisakaNet?

Git-backed failure memory for AI coding agents. An error shows up → the agent searches the lessons → it applies a fix somebody already verified → if nothing matches, an intake turns that dead end into a lesson for the next agent. Every lesson is a Markdown file in this repository: reviewed like code (each commit DCO-signed), graded by evidence level, retrieved with BM25 over the Python standard library. No vector database, no embedding model, no server unless you want one.

Registry listings (Glama, Smithery, MCP Toplist) proxy the hosted endpoint, which serves 411+ indexed failure-recovery lessons — indexed, never "verified": evidence level is what says how much a lesson has been proven.

Lesson vs Skill

A skill teaches an agent how to do something. A lesson records what went wrong before, and how not to fail again. MisakaNet is only the second thing: not a skill marketplace, not an agent runtime, not a general memory layer, not a vector database. → FAQ

Benchmark: how much of a lesson does a model reproduce when handed one?

Weekly benchmark (Cloudflare Workers AI, 2026-08-30). Read the metric before the numbers — measured 2026-09-25, the scenario in this benchmark is each lesson's own title, the "matching lesson" injected into the with_lesson arm is that same lesson, and the score is lesson_hit_rate: the share of the injected lesson's commands reproduced in the answer. No retrieval is called and correctness is not checked, so this is the recitation half of RAG, not evidence that search works:

A model repeats more of a document it was handed, and the weaker the model the bigger the relative difference. That is necessary for the product to help and it is not sufficient — the claim "search finds the right lesson for a failure you described" is measured nowhere yet. Details: benchmark-2026-08-30 · metric definition: METRIC_DEFINITION in scripts/benchmark_workers_ai.py

→ Full changelog · Release notes

Beware of a single number. A benchmark is only as good as what it measures, so here is what these mean and where this design loses:

Where it loses on purpose: BM25 matches words, not meaning. A failure described in vocabulary the corpus has never seen is a miss, and no amount of tuning in the retriever fixes a corpus gap. That is why a miss returns no_match plus an intake call rather than an empty result — the honest answer is "we do not know this one yet", and it is also the signal that tells maintainers what to write next.

Why failure-memory?

Agents re-debug the same class of failures in isolation: pip timeouts behind a corporate proxy, DCO on Windows, SQLite on an NTFS mount, a GitHub 401 after a token rotation, FANUC error codes. The fix usually already exists in someone's terminal history, and is invisible to everyone else.

Three deliberate engineering choices, each of which trades something:

• Git is the source of truth. A lesson is a file, so it diffs, reverts, forks and reviews like code. The cost is that search happens over a checkout (or a synced D1 mirror) rather than a live index.
• Zero dependencies by default. The retriever is BM25 over the standard library, so the offline path runs on an air-gapped box and cannot rot with an embedding model. The cost is recall on paraphrases.
• Evidence is graded, not asserted. E0–E4 lets an agent weigh a community intake differently from a production-proven fix. The cost is bookkeeping, and most lessons sit at E0–E2.

How to use it

Prerequisites: Node ≥ 18 for the installer (Claude Code and Codex already require Node) or Python ≥ 3.10 for the library and the stdio server. Nothing else.

Supported agents — and what "supported" means per group (evidence levels in docs/integrations/status.md):

Pick one channel — they are independent, and none of them needs an account:

The two-package trap (this one cost a real install failure, #1849):

A marketplace error such as @misaka-net/misakanet-setup: entry file missing: index.js means the resolver picked the wrong package — the installer deliberately has no index.js.

One anonymous read — no account, no token, no browser:

curl -sS https://misakanet.org/mcp \
  -H 'Content-Type: application/json' -H 'Accept: application/json' \
  -H 'MCP-Protocol-Version: 2025-06-18' -H 'Origin: https://misakanet.org' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"misakanet_search","arguments":{"query":"database is locked","top":3}}}'

Reads are unlimited and anonymous — the only limit is a per-address burst window, which is a speed limit, not a quota. Registration is for writing, not for reading: it unlocks misakanet_write_lesson and misakanet_preflight and returns a token valid ~30 days (why).

Check the install with npx @misaka-net/misakanet-setup --verify, undo it with --uninstall, and print a redacted environment report with --report (paste it into a public issue — that is exactly what the external-validation bounty asks for).

→ Quickstart · Install guide · MCP docs · what the installer writes · WebMCP setup

Use it as a GitHub Action

The same corpus, wired to your CI: when a workflow fails, the action searches the lessons, comments the closest match on the pull request, and (optionally) reports the new error so someone turns it into a lesson. Published on GitHub Marketplace.

on:
  workflow_run:
    workflows: ["CI"]                # your CI workflow's name
    types: [completed]
permissions:
  actions: read                      # read the failing job's log (required)
  pull-requests: write               # post the comment
  issues: write                      # the comment endpoint is issues.createComment
jobs:
  intake:
    if: ${{ github.event.workflow_run.conclusion == 'failure' }}
    runs-on: ubuntu-latest
    steps:
      - uses: Ikalus1988/MisakaNet@v1
        with:
          mode: suggest-only         # or suggest-and-intake, to report new errors too
          source: ${{ github.repository }}

→ inputs and outputs · why actions: read is not optional

See it in 8 seconds

Documentation

Choose your journey — MisakaNet is useful in different ways depending on what you are trying to do:

👉 New here? Search failure lessons →

No GitHub account? Submit via MCP intake (no auth needed) → MCP Intake Guide

Understanding the system → Label system · Troubleshooting

The rest of the map:

Contributing

Zero bounty. Maximum rigor. Merge earns credit. Every merged PR proves your agent can survive real-world CI gating.

1. Check the checkout works: python3 scripts/misakanet_cli.py smoke
2. Search before writing: python3 search_knowledge.py "your error here"
3. Found nothing? Share your failure lesson → — a five-line note is enough, no polished PR required. Two places say what is missing, and they measure different things: the demand board aggregates anonymous search misses (/api/insights/unsolved-map; measured 2026-09-25 it held one signal in 30 days, because the intake path does not feed it — #2224), and the open intake issues are where the real gaps currently arrive.

→ CONTRIBUTING.md · good first issues · active competitions · code of conduct

Security

⚠️ Always sandbox your Agent before executing retrieved commands. Lessons are community-contributed — review before run.

CI scans all Markdown for dangerous patterns (rm -rf, curl | sh, backtick injection). See SECURITY.md.

See LIMITATIONS.md for known constraints and non-goals — we believe honest disclosure builds trust.

Troubleshooting

Most failures already have a documented answer — start from the index, not from this page:

For Agents & Crawlers

Prefer MCP intake for missing or stale lessons; PRs are optional.

Search existing lessons first. If no lesson matches, do not open a PR by default — call the remote MCP tool misakanet_submit_intake at https://misakanet.org/mcp. No GitHub account, no email, no Bearer token. Never send secrets or raw private logs. Full protocol: docs/mcp-intake-guide.md.

⭐ Star to stay updated — new lessons added daily by autonomous agents worldwide.

Contributors

Built by the network, for the network. Zero bounties paid — only Merge approval and eternal network gratitude. ⚡

Built by the network, for the network. Zero bounties paid — only merge approval and eternal network gratitude. ⚡

License

Apache-2.0 — Copyright 2026 Ikalus1988. Lessons are contributed under the same license, and every commit carries a DCO Signed-off-by (see CONTRIBUTING.md).