Shinobi

MCPAI代理

提供方 numbererikson

AI代理 未标注 API Key 社区 / 第三方 ClaudeCursorVS CodeCline

Shinobi gives every AI coding agent on your machine — and every cloud Claude session — one shared task spine, decision log, and searched dead-ends ledger. Every failed approach is logged and semantically checked before an agent proposes a similar one, so agents stop repeating mis

一句话选型结论

按资料字段整理

Shinobi 属于AI代理类 MCP 服务器,传输方式未标注,API Key,可直接接入 Claude、Cursor、VS Code 等,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类AI代理 传输方式未标注 鉴权API Key 部署方式未标注
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

AI代理
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

mcpremote mcptask management

安装与配置

未标注 API Key
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
给代理接外部能力

让代理在任务中途调用搜索、数据库等真实工具。

多步任务编排

把一串操作拆成步骤,由代理按顺序执行。

自动化重复流程

把固定流程交给代理跑,减少手工操作。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 已适配 Claude、Cursor、VS Code,换客户端不用重新研究配置。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 使用前要准备好API Key对应的凭据,并按最小权限配置。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在AI代理场景里让模型直接调用外部能力的人。
  • 已经在用 Claude、Cursor 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 不方便管理密钥或第三方授权的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
Shinobi当前 未标注 未标注 未标注 无 未标注
Filesystem 560c4363 13 未标注 本地部署 无 2026-05-17
Server Everything (everything) 11 stdio · 本地进程 本地部署 有 2026-05-17
Time 7365cb6d 2 SSE · 服务端推送 远程部署 有 2026-05-18
Everything Claude Code 未标注 stdio · 本地进程 本地部署 有 2026-05-17
Openmemory 557e1585 未标注 stdio · 本地进程 本地部署 有 2026-05-17

服务器档案

AI代理
字段来自项目仓库与公开索引,中文说明由本站标注
提供方numbererikson上游标注的维护方 numbererikson
分类AI代理本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式API Key远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · Cross-platform本地安装时用到的包标识与运行时
部署类型未标注上游给出的部署形态
最近核验未标注上游最后核验时间

详细介绍

15.4 KB
来自项目仓库与上游目录的原文,未做删改

Shinobi

The task spine for AI coding agents. Shinobi holds the decisions that survive across sessions, actively searches your past dead ends — semantically — before the agent writes code, and routes approvals to your phone. One brain, every device: laptop, cloud session, and mobile all wired to the same store.

Works with Claude Code, Cursor, Cline, Continue.dev, Zed — any MCP-compatible client.

> Status: v0.3 — autonomous agents. Run it as a hosted HTTP /mcp brain (the default deploy) or self-host a local instance. Mobile push, a headless dispatch loop, parallel swarm over git worktrees, and audit→subtask ingestion on top of the remote MCP foundation. 39 MCP tools, web dashboard, mobile approvals, plugin system, optional semantic recall.

What it does

Most tools try to be a memory bolt-on. Shinobi is the task spine your agent
works along — the durable backbone of work, decisions, and known-bad paths
that outlives any single session and follows you across every device.

  • Task spine — projects + subtasks the agent claims, completes, or pivots; the persistent skeleton of multi-session work
  • Decisions that survive — record architectural choices with rationale so the next session (on any device) doesn't re-litigate them
  • Dead ends, searched before you build — every failed approach is logged and semantically matched the moment an agent plans a similar one, so it never burns a second session on the same wall. No other tool does this.
  • Approvals on your phone — request_approval pushes the decision to your pocket; the agent blocks until you tap yes/no, wherever you are
  • One brain, every device — laptop editor, Claude Code cloud session, and mobile chat all hit the same store over remote MCP; no sync step, no per-device drift
  • Plans — versioned plan snapshots, retrievable mid-task
  • Context — per-project conventions, "don't touch" rules, test patterns, deploy notes, file annotations
  • Recall — fulltext (FTS5) by default, semantic (embedding-backed) when an embedding provider is configured
  • Notes — free-form annotations and voice notes (audio_path field)
  • Activity timeline — every write path lands in the timeline so you can replay what happened
  • Git linking — link_commit ties commits to subtasks via [SHI-N] tags or via target_path attribution
  • Web dashboard — Hono-served Kanban + decisions + dead ends + notes + plans + context + timeline + analytics
  • Plugin system — drop a .js file in ~/.shinobi/plugins/ or install a @shinobi/plugin-* npm package and register custom plugin_* tools

Hosted or self-hosted, your call. The default deploy is one remote brain
behind an HTTP /mcp endpoint (we run ours at shinobi.shinobi-apps.com); the
same binary still runs as a fully local single-machine instance when you'd
rather keep everything on your own box. BYO embedding provider only if you want
semantic recall.

> 🚀 New here? Follow Getting started — zero to a
> working brain in ten minutes. Going multi-device? Remote mode
> + $0/month cloud deploy.

Install

Requirements:

  • Node.js 18+ on PATH
  • C++ build toolchain for better-sqlite3 native build (most systems have prebuilt binaries; Windows may need Visual Studio Build Tools as fallback)

From npm (recommended)```bash

npm install -g @shinobiapps/shinobi


The binary is `shinobi` (e.g. `shinobi serve`, `shinobi dashboard`).

### From GitHub (latest, unreleased)```bash
npm install -g github:numbererikson/shinobi

Pulls from main. Useful for trying unreleased fixes. On Windows you may need
to add your Node directory to system PATH before this works, because the
prepare build script runs in a subshell that does not always inherit
per-session PATH (Laragon, portable installs). If install fails with 'node' is not recognized, prefer the npm install above.

From a cloned source folder (for development / contributing)```bash

git clone https://github.com/numbererikson/shinobi.git
cd shinobi
npm install # triggers prepare → builds dist/
npm install -g .


### Then bootstrap

In any project root where you want Shinobi available to your MCP client:```bash
shinobi init
shinobi dashboard

init will:
1. Create ~/.shinobi/ with config.json, .env template, and shinobi.db (migrations applied)
2. Drop a .mcp.json snippet for the current project
3. Print next steps

shinobi init writes config for the two clients with a workspace-local
MCP convention out of the box:

  • Claude Code — /.mcp.json
  • Cursor — /.cursor/mcp.json

Restart the client and the mcp__shinobi__* tools become available.

For other MCP clients (Cline, Continue.dev, Zed), see the
MCP client setup section below.

Then open:```text
http://127.0.0.1:8765


On Windows PowerShell, if script execution blocks `shinobi`, use the `.cmd` shim:```powershell
shinobi.cmd dashboard

If you upgraded Node or copied an old node_modules, rebuild native dependencies:```bash
npm rebuild better-sqlite3


Important: the code lives in the Shinobi folder, but the local memory database lives in:```text
~/.shinobi/shinobi.db

To move the tool only, copy/clone the Shinobi folder and run the install commands above. To move the existing projects, tasks, decisions, notes, and context too, either copy ~/.shinobi/ or use shinobi sync.

MCP client setup

Every snippet below uses the same JSON shape — command is the path
to the Node binary that's running Shinobi, args is [, "mcp"]. Print the exact values for your machine:```bash
shinobi init --print-config


(Or read `.mcp.json` from any project where you already ran
`shinobi init` — the values are identical.)

### Claude Code

Drops in automatically — `shinobi init` writes `/.mcp.json`.
Restart Claude Code to pick up the server.

### Cursor

Drops in automatically — `shinobi init` writes `/.cursor/mcp.json`.
Works on Cursor 0.43+. Restart Cursor or reload the workspace.

For a **global** Cursor config (every project sees Shinobi), paste the
same snippet into `~/.cursor/mcp.json` (or use Cursor Settings → MCP).

### Cline (VS Code extension)

Open Cline's settings file:

- Windows: `%APPDATA%\Code\User\globalStorage\saoudrizwan.claude-dev\settings\cline_mcp_settings.json`
- macOS: `~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json`
- Linux: `~/.config/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json`

Merge the contents of your project's `.mcp.json` into the file's
`mcpServers` object. Restart VS Code.

### Continue.dev

Edit `~/.continue/config.json`. Add Shinobi to the `mcpServers` array
(note: Continue uses an **array**, not an object like the others):```json
{
  "mcpServers": [
    {
      "name": "shinobi",
      "command": "/absolute/path/to/node",
      "args": ["/absolute/path/to/dist/cli.js", "mcp"]
    }
  ]
}

Use the values from your project's .mcp.json for command and args.

Zed

Edit ~/.config/zed/settings.json. Zed nests MCP servers under
context_servers:```json
{
"context_servers": {
"shinobi": {
"command": {
"path": "/absolute/path/to/node",
"args": ["/absolute/path/to/dist/cli.js", "mcp"]
}
}
}
}


Restart Zed.

### Generic / other clients

Any MCP client that supports the standard `{ command, args }` server spec
should work. Use the same values your `.mcp.json` has:

- `command`: absolute path to the Node binary running Shinobi
- `args`: `[, "mcp"]`

Avoid the bare `shinobi` command in MCP config — many clients spawn
servers with `shell: false`, which skips the OS PATH resolution that
makes `shinobi` work in a terminal.

### Remote mode (the default deploy)

Host one Shinobi brain on a server and connect every device to it — your
desktop editor, Claude Code web/mobile sessions, any remote-MCP-capable
client. `shinobi serve` exposes the MCP endpoint at `/mcp` (streamable HTTP,
**stateless**, bearer-token auth) alongside the dashboard:```bash
claude mcp add --transport http shinobi https://your-host/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

This is the recommended way to run Shinobi — one brain, reachable from every
device. The same binary still runs as a local single-machine instance if you'd
rather self-host everything on your own box. Full deployment guide (Docker,
Cloudflare Tunnel, GCP Always Free, client config):
docs/remote-mcp.md.

CLI```

shinobi <command> [options]

Commands:
init Bootstrap ~/.shinobi/ and drop .mcp.json in the current directory
mcp Run the MCP server over stdio (invoked by the MCP client)
migrate Apply pending SQL migrations
dashboard Start the web dashboard on localhost (default port 8765)
serve [--host H] [--port P] Dashboard + MCP HTTP endpoint (/mcp) in one process — see docs/remote-mcp.md
sync init <path> [branch] Configure a local git repo as the cross-machine sync target
sync push Snapshot the DB and commit it to the sync repo
sync pull Restore the DB from the sync repo's snapshot
sync status Show last push/pull timestamps and git status
dispatch [--once|--drain] Autonomous loop: pull next_task → run worker → complete/unblock → repeat
[--project N] [--interval S] [--max-failures N] Worker via SHINOBI_WORKER_CMD (e.g. 'claude -p "$SHINOBI_TASK_PROMPT"'); unset → dry-run
swarm --agents N N dispatch loops in parallel, each in its own git worktree/branch, one shared
[--project N] [--drain] brain. Atomic claim → no two agents take the same task. --no-worktree / --keep-worktrees


## MCP tools (39)

| Group | Tools |
|---|---|
| Projects | list_projects, get_project, create_project, update_project, archive_project, unarchive_project, delete_project |
| Subtasks | list_tasks, get_task, create_task, bulk_create_tasks, update_subtask, delete_subtask, claim_task, complete_task, next_task |
| Decisions | log_decision, decisions_for_file, update_decision_status |
| Dead ends | log_dead_end, check_dead_ends |
| Notes | add_note, list_notes |
| Plans | save_plan, get_plan |
| Context | get_context, update_context |
| Recall | recall (FTS5 or semantic) |
| Timeline | history, link_commit |
| Workflow | agent_bootstrap, session_closeout, file_context |
| Extraction | extract_decisions, compress_session_summary |
| Approvals | request_approval |
| Notifications | notify |
| Findings | ingest_findings |
| Plugins | plugin_hello |

## Architecture

| Layer | Tech |
|---|---|
| Language | TypeScript (strict mode, ES2022, NodeNext) |
| Runtime | Node 18+ |
| MCP | `@modelcontextprotocol/sdk` 1.x |
| Storage | SQLite via `better-sqlite3` (WAL mode) |
| Dashboard | Hono + `@hono/node-server` (same process, localhost:8765) |
| Embeddings (optional) | OpenAI text-embedding-3-small / Voyage voyage-3-lite / Ollama nomic-embed-text |
| Migrations | Forward-only, sha256 checksum, `schema_migrations` table |

See docs/architecture.md for the request lifecycle and module layout.

## Dashboard auth

The dashboard is open on loopback binds (`127.0.0.1`, `localhost`, `::1`) and token-protected on any non-loopback bind. The token is read from `SHINOBI_DASHBOARD_TOKEN`, otherwise loaded from `~/.shinobi/dashboard-token`, otherwise auto-generated and persisted there. `/health` is always open for probes.

Browser flow — open the dashboard with the token once and the cookie sticks:```text
http://192.168.1.10:8765/?token=YOUR_TOKEN

Curl / scripts — any of these works:```bash
curl -H "Authorization: Bearer $SHINOBI_DASHBOARD_TOKEN" http://192.168.1.10:8765/api/projects/1/snapshot
curl -H "X-Shinobi-Token: $SHINOBI_DASHBOARD_TOKEN" http://192.168.1.10:8765/api/projects/1/snapshot
curl --cookie "shinobi_token=$SHINOBI_DASHBOARD_TOKEN" http://192.168.1.10:8765/api/projects/1/snapshot


See docs/configuration.md for the full env-var reference.

## Cross-machine sync

Shinobi syncs your local SQLite database via a private git repo. Setup once per machine:```bash
# Once: clone a private GitHub repo to act as the sync repo
git clone git@github.com:USER/shinobi-sync.git ~/shinobi-sync

# Once per machine: tell shinobi about it
shinobi sync init ~/shinobi-sync main

# Push from the writer machine:
shinobi sync push

# Pull on the reader machine:
shinobi sync pull

The DB file lands on the main branch as a binary; pre-existing local DB is backed up to .bak- before restore.

Configuration

Edit ~/.shinobi/.env or set env vars before running shinobi mcp:

Variable Default Purpose
SHINOBI_DB_PATH ~/.shinobi/shinobi.db SQLite location
SHINOBI_CONFIG_DIR ~/.shinobi Config + plugins directory
SHINOBI_PLUGINS_DIR ${configdir}/plugins Plugin discovery directory
SHINOBI_DASHBOARD_PORT 8765 Dashboard port
SHINOBI_DASHBOARD_HOST 127.0.0.1 Dashboard bind host. Loopback skips auth; non-loopback auto-enables token auth.
SHINOBI_DASHBOARD_TOKEN (auto) Override the dashboard token; auto-generated to ~/.shinobi/dashboard-token when needed.
SHINOBI_EMBED_PROVIDER none openai / voyage / ollama / none
SHINOBI_EMBED_API_KEY — Auth for OpenAI / Voyage (else falls back to OPENAI_API_KEY / VOYAGE_API_KEY)
SHINOBI_EMBED_MODEL provider default Override model id
SHINOBI_EMBED_DIMS provider default Override dimensions
SHINOBI_OLLAMA_URL http://localhost:11434 Ollama endpoint
SHINOBI_MIGRATIONS_DIR /migrations Override migrations source

Full reference: docs/configuration.md.

Plugins

Write a single .js file in ~/.shinobi/plugins/:js // ~/.shinobi/plugins/my-plugin.js export default function register(registry, api) { registry.registerTool({ name: &#x27;plugin_count_open&#x27;, description: &#x27;Count open decisions across all projects&#x27;, inputSchema: { type: &#x27;object&#x27;, properties: {}, additionalProperties: false }, handler: (_args, api) =&gt; { const projects = api.listProjects(); let total = 0; for (const p of projects) { total += api.listDecisions({ projectId: p.id, status: &#x27;open&#x27; }).length; } return { open_decisions: total }; }, }); }

Restart the MCP client and mcp__shinobi__plugin_count_open is available. See docs/plugin-development.md.

Roadmap

  • v0.2 — remote MCP foundation: stateless HTTP /mcp endpoint,
    Docker + GCP Always Free deploy, Cloudflare Tunnel, env-driven .mcp.json
    for cloud sessions. One brain across laptop, cloud, and mobile.
  • v0.3 (current) — autonomous agents: push notifications (task-completed /
    agent-blocked) → dispatch loop (headless agent pulls next_task and works
    while you sleep) → Shinobi Swarm (N parallel agents, worktree isolation,
    shared dead ends), plus ingest_findings (audit → subtask graph → swarm)
    pulled forward from v0.4.
  • v0.4 — unit-test coverage pass on tools/*; richer remediation templates
    and finding-source adapters on top of ingest_findings.
  • v1.0 — stability, security audit, performance pass; optional team mode /
    hosted SaaS only on demand signal.

Changelog

See CHANGELOG.md for release notes.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

By participating you agree to our Code of Conduct.

Security

For vulnerability reports, see SECURITY.md — please do
not open public issues for security matters.

License

MIT — see LICENSE.

常见问题

7 问
答案由本站依据条目字段整理,不引用上游原文
Shinobi 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

Shinobi 怎么在 Claude、Cursor、VS Code 里配置?

它的传输方式未标注,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

Shinobi 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

Shinobi 需要 API Key 吗?

资料标注需要API Key认证,使用前准备好对应凭据,并按最小权限配置。

Shinobi 支持哪些客户端?

资料标注支持 Claude、Cursor、VS Code、Cline、Codex、Continue;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

Shinobi 能提供哪些能力?

资料显示能力标签包括 mcp、remote mcp、task management、memory。具体工具清单以官方文档为准。

Shinobi 需要什么运行环境?

运行时为 Cross-platform。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 未标注。