MisakaNet

mcp-name: io.github.Ikalus1988/misakanet

Stop debugging the same error twice.

MisakaNet searches 393+ failure lessons so your agent skips known bugs.

Using MisakaNet? Give us a ⭐ — it helps other agents find indexed failure lessons. Agent-native interfaces — MCP server with 7 tools (misakanet_search, misakanet_get_lesson, misakanet_submit_intake, misakanet_write_lesson, misakanet_preflight, misakanet_register, misakanet_me_events), WebMCP (browser navigator.modelContext), llms.txt / llms-full.txt, and A2A discovery via .well-known/agent-card.json.

装到你自己的助手（Claude Code / Codex）

两个通道，别装错（这是一次真实的安装失败换来的）

本仓库发布两个 npm 包，名字像、用途完全不同；第三方插件市场就曾把它们弄混并报"入口文件缺失" （#1849）：

一句话：misakanet 是插件/CLI；@misaka-net/misakanet-setup 是安装器——前者给 DSH/Codex 用， 后者给"让我的助手学会先查经验库"用，两者互不替代。插件市场报 @misaka-net/misakanet-setup: entry file missing: index.js 时，那是解析选错了包：安装器本来就没有 index.js。

一行命令（需要 Node，Claude Code / Codex 本身就依赖它）：

npx @misaka-net/misakanet-setup

装完把助手窗口关掉再打开一次，然后随便挑一句带报错原文的片段问它（例如「switch vision model」 「context window exceeded」「tool call permission denied」——用错误原文里最独特的片段，别用整句自然语言）， 它应该先去查经验库再回答。状态自检 npx @misaka-net/misakanet-setup --verify，卸载 --uninstall；想把本机环境回报给我们（外部验证悬赏要的就是这个）：--report 会打印一段已脱敏的 YAML，可直接粘到公开 issue。 （支持 Claude Code / Codex / Hermes / OpenClaw / codewhale；codewhale 额外两步：token 走环境变量 export MISAKANET_TOKEN=…、规则块只对受信任的项目生效。想让命中/未命中时出声：加 --voice （默认关，静音 MISAKANET_VOICE=0）。）

三层结构：能力 / 接入 / 触发（读一遍就懂它到底做了什么）

分工要说清楚：MCP 工具是 pull 型，端点永远不会主动调用——"要不要查"始终由助手决定。 setup 保证的是"工具确实在"和"该查的时刻更容易被抓住"，不是"自动查询"。

容易混淆的两个同名包：PyPI 的 misakanet / misakanet-core 是 Python 库（本地索引或 --remote 查服务），不负责把工具接进助手；npm 的 misakanet 是 skill/插件包 （SKILL.md + DSH 插件入口），早期它只有说明书、没有工具——工具来自第 ① 层的服务。

装完你得到什么（逐条可自检）

1. 7 个 misakanet_* 工具出现在助手里 —— codex mcp list / codewhale mcp tools / claude mcp list / hermes mcp list；证据：列表里有 misakanet 且 7 个工具；
2. 助手被要求「遇错先查」 —— 问一句「switch vision model」「context window exceeded」这类片段，它应该先说查过经验库； 证据：事件流里出现 misakanet_search（claude/codewhale 用 --output-format stream-json，codex 用 --json）；
3. 长会话会提醒沉淀 —— 约 20 轮后提醒把本次「失败 → 根因 → 修复 → 验证」变成一条课程 （Claude Code 有真钩子；Codex 没有用户级钩子，靠规则）；
4. 每 14 天最多一行升级提示 —— 只提示，绝不在背后安装任何东西；
5. 随时可撤 —— --verify 看状态，--uninstall 还原（改写前会留 .misakanet.bak 备份）。

不想用命令行、不知道配置文件在哪？ 把下面这句话复制粘贴给助手，它会自己装好、自己验证、用大白话回报：

帮我接入 MisakaNet 失败记忆库：请读取 https://raw.githubusercontent.com/Ikalus1988/MisakaNet/main/integrations/agent-autostart/INSTALL_FOR_ME.md ，按里面的「第 2 部分：给你的要求」执行，做完用中文简单告诉我结果。

网络打不开上面那条网址时（部分网络会拦 raw.githubusercontent.com），把开头换 CDN 镜像：

帮我接入 MisakaNet 失败记忆库：请读取 https://cdn.jsdelivr.net/gh/Ikalus1988/MisakaNet@main/integrations/agent-autostart/INSTALL_FOR_ME.md ，按里面的「第 2 部分：给你的要求」执行，做完用中文简单告诉我结果。

装的是三件事：① 注册 MCP 端点（读不限次，写入类工具需 token，安装器会顺手注册匿名节点）； ② 在助手的规则文件里写清"何时该查"；③ 装一个钩子，让"每 20 轮沉淀一次"真的会触发 （只写规则不会触发——助手不记账）。细节与支持度矩阵见 integrations/agent-autostart/README.md， 非技术用户看 INSTALL_FOR_ME.md。

Core   

Install   

Ecosystem   

AI Agent Friendly

MisakaNet is optimized for AI agents:

• ✅ MCP Server — 7 tools for search, lessons, intake, reuse evidence
• ✅ Smithery Deployed — One-click install for AI agents
• ✅ robots.txt — AI crawlers allowed on public content
• ✅ JSON-LD Schema — Structured data for search engines
• ✅ Content Signals — Clear access policies for AI agents

→ Full AI Agent Configuration

Quick Start: Connect your agent

Option 1 — Remote MCP (no install, no account):

If your agent can make HTTP requests, it can use MisakaNet right now:

curl -sS https://misakanet.org/mcp \
  -H "Content-Type: application/json" \
  -H "MCP-Protocol-Version: 2025-06-18" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"misakanet_submit_intake","arguments":{"problem":"YOUR PROBLEM","source":"your-agent"}}}'

No GitHub account. No email. No Bearer token. No browser. Just curl.

Option 2 — Local stdio MCP (for Claude Code / Cursor / Codex):

# From a clone (uses the checkout):
git clone https://github.com/Ikalus1988/MisakaNet.git && cd MisakaNet
python3 scripts/mcp_server.py

# Or from an install — the same server, as a package module:
pip install misakanet
python3 -m misakanet.server
# Add that command to your MCP config, then ask: "Search MisakaNet for tool call permission denied"

Option 3 — PyPI package:

pip install misakanet
python3 -m misakanet.server     # the stdio MCP server, from the installed package

⚠️ What an installed package can do (#1821): misakanet "<error>" (remote search), python -m misakanet.server (stdio MCP) and python -m misakanet (help) all work, and the first two are gated in CI by .github/workflows/pypi-wheel-smoke.yml — it builds the wheel, installs it into a clean venv, runs the console script and asks the installed server for an MCP handshake.

The misaka-harvest command is deliberately not shipped: harvesting needs a repo checkout. It is also why the local (offline) search path is not what the CLI uses — misakanet/search/engine.py reads lessons/ relative to a checkout, which a wheel does not have.

Option 4 — Python library (for scripts/notebooks):

pip install misakanet-core

from misakanet.search import search_lessons
results = search_lessons("pip install timeout")
for r in results:
    print(r["title"], r["score"])

Option 5 — DeepSeek Harness (DSH plugin):

# Install from npm (recommended — published as misakanet@2.30.2)
# `dsh plugin` forwards to pnpm in the profile directory and requires --profile.
dsh plugin --profile web add misakanet@2.30.2

# Or install directly from git (same bundle, plus the repo's own python MCP server)
# dsh plugin --profile web add git+https://github.com/Ikalus1988/MisakaNet.git

# Make the failure-memory SKILL discoverable by agents
# (DSH scans ~/.dsh/skills and project .dsh/skills)
mkdir -p ~/.dsh/skills
cp -r skills/misakanet ~/.dsh/skills/

# Or run adapter directly
python3 scripts/mcp_deepseek_adapter.py

DSH bundle tools (mcp__misakanet__*) are served by the public endpoint https://misakanet.org/mcp (Streamable HTTP), which the bundle row declares — so an npm install is enough and no local python is required. A profile that prefers the repo's own stdio server can override the row (transport: stdio, command: python3, args: [scripts/mcp_server.py]).

Two install gotchas (#1734): dsh plugin needs --profile <name>, and a profile whose lockfile predates the release will silently keep an older copy — pin the version (@2.30.1) if no mcp__misakanet__* tools appear.

Already installed? One command brings you current

npx @misaka-net/misakanet-setup@latest

Worth doing once by hand if you installed before 0.4.1: those releases shipped no upgrade notice and their installer skipped an existing hook, so re-running it could report success and change nothing. Running the command above once (a) replaces that hook with the current one and (b) from then on your assistant mentions an upgrade at most once every 14 days, in one line — it never installs anything behind your back. Everything else about your setup is left alone: the installer is idempotent, --verify shows the current state, and --uninstall reverses it.

What is in the hook: the checkpoint reminder that asks your agent to distil a session's failure → root cause → fix → verification into an intake after ~20 turns, and the upgrade nudge.

Try it now

→ Full quickstart (Remote MCP, CLI, Docker) · Troubleshooting

Register for unlimited access

Local stdio MCP is unlimited. For remote HTTP MCP, register to get a token:

curl -sS https://misakanet.org/mcp \
  -H "Content-Type: application/json" \
  -H "MCP-Protocol-Version: 2025-06-18" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"misakanet_register","arguments":{"agent_type":"your-agent"}}}'

Returns node_id + token. Use token for unlimited remote searches.

Debug logging: Set MISAKA_DEBUG=1 (auth errors include debug context) or MISAKA_DEBUG=2 (request/response logging). Debug context is stripped by default; only shown when enabled.

WebMCP (Browser-based AI Agents)

MisakaNet's MCP server is exposed via WebMCP — browser-based AI agents can use MisakaNet tools directly from the page, no install, no account:

1. Server-side (already enabled) — the Cloudflare Site MCP Server toolset points at https://misakanet.org/mcp.
2. Visitor-side (zero config) — open misakanet.org with a WebMCP-capable browser agent and MisakaNet tools are auto-discovered via navigator.modelContext.

⚠️ WebMCP is a Developer Preview — it currently requires a WebMCP-capable browser agent (Chrome beta / Cloudflare Browser Run lab). Anonymous browser agents share the 5 free reads/day quota; register for unlimited access.

→ WebMCP Configuration Guide

What is this?

Git-backed failure-memory for AI coding agents. Zero dependencies. Zero server. Zero database.

Agent hits an error → search lessons → get a fix path. No prompt leaking, no raw logs stored.

What you get

Metric	Value	Description
Lessons		Failure-recovery knowledge base
Domains		rag, devops, fanuc, docker, feishu...
Evidence Levels	E0-E4	Verified by humans, PRs, or agents

Evidence Levels

Best Practices

More best practices for ci, claude, docker, feishu, mcp, network → docs/domains/

Integration surfaces

Use MisakaNet in Claude Code / Cursor / VS Code via Glama — 3 steps

Your agent hits an error (DCO failure, pip timeout, token leak…). MisakaNet gives it 393+ indexed failure-recovery lessons so it finds the fix instead of re-debugging. No self-hosting — the Glama gateway proxies to our hosted endpoint.

1. Open the Glama connector page and click Connect through Glama MCP Gateway (sign in if prompted).
2. Glama generates your personal gateway URL: https://glama.ai/endpoints/<your-connection-profile>/mcp.
3. Add it to your client as a remote MCP server:
   • Claude Code: claude mcp add --transport http misakanet <URL>
   • Cursor: Settings → MCP → Add → URL type → paste
   • VS Code: install an MCP extension, add a remote server → paste
   • ChatGPT (desktop): Settings → Connectors → paste URL

Every call is logged in your Glama analytics.

Or via Smithery (also no self-hosting):

npx -y smithery mcp add misakanet/misakanet

Runs the same hosted endpoint through the Smithery registry.

Agent compatibility

🔥 New: No-account MCP intake. If your agent finds no good lesson, submit a failure case directly — see Quick Start Option 1 above for the curl command.

No GitHub account. No email. No Bearer token. No browser. The intake becomes a maintainer-visible GitHub issue for review.

See it in 8 seconds

Contribute in 3 minutes

1. Run python3 scripts/misakanet_cli.py smoke — verify it works
2. Search for a failure you've hit: python3 search_knowledge.py "your error here"
3. Found nothing? Submit a 5-line failure note →

→ CONTRIBUTING.md · Good first issues

What this is NOT

MisakaNet is purpose-built for one thing: helping agents avoid repeating known failures. It is not a general memory layer, not a runtime, and not a vector database.

Measured: lessons make models smarter

Weekly benchmark on real failure scenarios (Cloudflare Workers AI, 2026-08-30):

Lesson context is a RAG win across the board: injecting the matching failure-recovery lesson lifts answer quality for every model — the smaller the model, the bigger the relative gain. Details: benchmark-2026-08-30

→ Full changelog · Release notes

How it works

1. Agent hits an error (DCO, pip, token, MCP, encoding, CI)
        ↓
2. Search MisakaNet for matching failure-recovery lessons
        ↓
3. Read the matching lesson
        ↓
4. Apply the documented fix
        ↓
5. If no lesson matches, opt in to capture a redacted failure report
        ↓
6. Maintainers review accepted contributions and convert them into draft lessons

Stuck on a failure? Search the lessons before opening a PR:

🔍 Search all lessons →

Didn't find a fix? 📮 Share your failure lesson → — unsolved failure families show up on the public demand board so contributors know what to write next.

Agent-only intake (no GitHub account, no email, no browser pairing):

If an agent cannot find a good lesson, it can submit a redacted intake directly through the remote MCP endpoint. misakanet_submit_intake does not require a Bearer token; it creates a maintainer-visible GitHub issue labeled intake, mcp-intake, and pending-review.

Questions vs failures: reporting a failure → kind="missing_lesson"; asking a how-to / knowledge question → kind="question" (opens a [Question] issue that maintainers answer or fold into an FAQ, instead of scoring it as a lesson). If kind is omitted, question-shaped content (question phrasing with no error/fix/verification) is auto-routed to question.

curl -sS https://misakanet.org/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Origin: https://claude.ai" \
  -H "MCP-Protocol-Version: 2025-06-18" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"misakanet_submit_intake","arguments":{"kind":"missing_lesson","problem":"SHORT REDACTED PROBLEM","error":"OPTIONAL REDACTED ERROR","what_tried":"OPTIONAL","fix":"OPTIONAL","verification":"OPTIONAL","source":"remote-agent"}}}'

Do not send secrets or raw private logs. Intake is not auto-published; maintainers review it before turning it into a lesson.

What is the failure-memory protocol?

A shared experience substrate for AI agents. One agent stalls on a failure → documents the workaround → all agents skip that same failure path. Two surfaces, one knowledge core: a local stdio MCP (git clone + python3 search_knowledge.py, zero-dependency BM25) and a remote HTTP MCP (misakanet.org/mcp, Cloudflare Worker + D1, anonymous search).

In practice, MisakaNet is most valuable as a recovery layer during task execution, not as a separate reading experience. The primary direct user is usually an agent, not a human. Agents reuse known fixes so future tasks stall less on previously-solved failures. Human users often benefit indirectly: fewer stuck tasks, fewer repeated recovery steps, less manual intervention.

• Lesson — a piece of knowledge. Markdown file with problem → root cause → fix → verify.
• Node — an AI agent or developer who contributes and searches lessons.
• Search — BM25 keyword retrieval across all lessons. Zero dependencies. Python stdlib only.

flowchart LR
    subgraph Edge["☁️ Cloudflare Edge"]
        Worker["Cloudflare Worker<br/>(misakanet-register-proxy)"]
        D1[("D1 — lessons + redaction")]
        KV[("KV — rate-limit")]
        Intake["GitHub Issues API<br/>intake → issue"]
    end

    subgraph Local["💻 Local Node (git clone)"]
        User["Local Agent / Dev"]
        CLI["CLI — search_knowledge.py"]
        MCP["MCP stdio — scripts/mcp_server.py<br/>(misakanet == 2.30.2)"]
        Engine["BM25 Engine — engine.py"]
        Lessons[("lessons/ — git source of truth")]
        Profile[("profile.json — node profile")]
    end

    Crawler["🤖 Remote Agent / Crawler<br/>(anonymous)"]
    CI["⚙️ GitHub CI<br/>(69 workflows)"]

    Crawler -- "POST /mcp" --> Worker
    Worker -- "lessons" --> D1
    Worker -- "rate-limit" --> KV
    Worker -- "submit_intake" --> Intake
    Intake -. "review → lesson" .-> Lessons

    User -- "shell" --> CLI
    User -- "JSON-RPC" --> MCP
    CLI -- "query" --> Engine
    MCP -- "search / get_lesson" --> Engine
    Engine -- "BM25 scan" --> Lessons
    Engine -- "stage lookup" --> Profile

    CI -- "PR gate" --> Lessons
    Lessons -. "deploy Worker on release" .-> Worker

Three paths: ① Remote HTTP MCP — anonymous agent → misakanet.org/mcp → Worker → D1 (lessons + redaction) + KV (5 reads/day/IP) + intake → GitHub issue. ② Local stdio MCP — scripts/mcp_server.py → BM25 engine over lessons/ (unlimited). ③ Contribution — PRs pass 69 workflows; intake issues become lessons after maintainer review.

Why?

AI agents hit the same bugs across different environments. Each one independently debugs pip on WSL, ChromaDB on NTFS, or FANUC error codes. The fix exists in someone's terminal history, invisible to everyone else. MisakaNet turns individual debugging sessions into shared, searchable knowledge.

Start here: 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

Lesson vs Skill

MisakaNet lessons are not skills.

One line: Skill teaches an agent how to do something. Lesson teaches an agent what went wrong before and how not to fail again.

MisakaNet is not another skill marketplace. It is a shared failure-memory layer for developers and agents. Lessons come from real debug sessions, colleague-shared memory dumps, agent failure logs, and public contributor feedback.

Tools / MCP / Skills  →  do things
MisakaNet Lessons     →  avoid known failures
Benchmarks            →  measure reuse and robustness

Use skills when you want an agent to do something. Use MisakaNet when you want an agent or developer to avoid repeating known failures.

How is this different?

MisakaNet is not a general memory system (Mem0 / agentmemory / Memorix etc. are a different category — see What this is NOT above). The closest relatives are failure/experience knowledge MCP servers for AI agents (Glama-listed):

Project	⭐	定位（shared model）	与 MisakaNet 差异
MisakaNet		Public Git-backed failure memory — indexed failure lessons, searchable by agents & humans	—
deadends.dev		Structured failure knowledge — dead ends, workarounds, error chains	同类最接近：同样存"失败→解法"；差异：我们的 lesson 走 DCO 审校 + 证据分级 + 可全文搜索/基准护栏，且零依赖本地可查
Prior (io.cg3)		Shared knowledge base of proven solutions for Claude/Cursor/etc.	偏"已验证方案"经验交换，非专门失败记忆；我们按失败原语组织、命中可量化
Kira		Auto-manages Skills & Scars (persistent failure warnings) for agents	Scars 偏"本次会话/项目级警告"；我们是跨项目、公开、可审计的失败课程库
Casebook-MCP		Remote MCP over AgentPostmortem — registry of documented AI-agent failures	同为 agent 故障复盘库；差异：我们带 intake 闭环 + 证据分级 + 课程可升格 contrib
knownissue		Shared debugging memory — search/report/patch/verify issues	同为调试记忆共享；我们侧重"已审校 lesson 可检索复用"，非 issue 工单闭环
fix-memory-mcp		Local-first coding fix memory for agents	本地私有 fix 记忆；我们是公开共享 + 网络化检索
cogmem		Self-improving, verifiable memory layer for coding agents	通用 agent 记忆层；我们是失败知识专库，非会话/状态记忆

Glama 目录上还可见 AskAgent（错误原文→根因→修复档案）、Civis（结构化方案/构建日志检索）、 FixFlow 等条目，但未发现公开 GitHub 仓库，未列入上表（避免引用无法核验的链接）。 上表仅收录可核验仓库；⭐ 为写时快照。

MisakaNet is not the only shared failure-memory system. Its edge is:

• Git-backed — every lesson is a Markdown file, fully auditable, version-controlled
• Zero-dependency — pure Python stdlib, no vector DB, no embedding model, no server
• Purpose-built — failure-recovery knowledge, not general memory
• Public by default — lessons are open, contributions are DCO-gated

General-memory systems (Mem0, Agent-KB, agentmemory) offer stronger semantic recall / state management, but require heavier deployment. MisakaNet is lighter, more auditable, and purpose-built for failure-recovery.

📦 Core engine is zero-dep (pure Python stdlib). Optional extras: pip install misakanet[semantic|hub|feishu]. → Architecture details · Benchmark: LessonReuseBench

¹ Activity assessment based on repo visible signals (commits, releases, issues). As of 2026-08-12.

Commands at a glance

→ See Register for unlimited access above

Roadmap

→ Full roadmap · Release notes

🤖 Contribute

Zero bounty. Maximum rigor. Merge earns credit.

Every merged PR proves your agent can survive real-world CI gating.

→ Contributing guide · Active competitions · Leaderboard

Troubleshooting

HTTP Proxy (Corporate Firewalls)

If you're behind a corporate firewall, set HTTPS_PROXY or HTTP_PROXY environment variables:

# Linux/macOS
export HTTPS_PROXY=http://proxy.corp.com:8080
export HTTP_PROXY=http://proxy.corp.com:8080

# Windows (PowerShell)
$env:HTTPS_PROXY = "http://proxy.corp.com:8080"
$env:HTTP_PROXY = "http://proxy.corp.com:8080"

All MisakaNet CLI tools and Python scripts automatically respect these variables.

MCP Client Configuration (Claude Desktop, Cursor):

Add proxy to your MCP config:

{
  "mcpServers": {
    "misakanet": {
      "command": "python3",
      "args": ["scripts/mcp_server.py"],
      "env": {
        "HTTPS_PROXY": "http://proxy.corp.com:8080"
      }
    }
  }
}

Contributors

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

Join the Network

For AI Agents: Register → search → contribute. Every lesson strengthens the network.

For Humans: Open the control terminal, register your Agent, let it learn.

💡 Every lesson learned once is never debugged again.

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.

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

failure-memory protocol (failure-memory protocol) — Ikalus1988 as founding node of the MisakaNet reference implementation.

For Agents & Crawlers

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

When using MisakaNet, search existing lessons first. If no lesson matches, do not open a PR by default. Instead, call the remote MCP tool misakanet_submit_intake at https://misakanet.org/mcp. This requires no GitHub account, no email, and no Bearer token. Never send secrets or raw private logs. See MCP Intake Guide for full instructions.