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)

npm install -g @shinobiapps/shinobi

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

From GitHub (latest, unreleased)

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)

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:

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 — <workspace>/.mcp.json
• Cursor — <workspace>/.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:

http://127.0.0.1:8765

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

shinobi.cmd dashboard

If you upgraded Node or copied an old node_modules, rebuild native dependencies:

npm rebuild better-sqlite3

Important: the code lives in the Shinobi folder, but the local memory database lives in:

~/.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 [<absolute path to dist/cli.js>, "mcp"]. Print the exact values for your machine:

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 <workspace>/.mcp.json. Restart Claude Code to pick up the server.

Cursor

Drops in automatically — shinobi init writes <workspace>/.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):

{
  "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:

{
  "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: [<absolute path to dist/cli.js>, "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:

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)

Architecture

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:

http://192.168.1.10:8765/?token=YOUR_TOKEN

Curl / scripts — any of these works:

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:

# Once: clone a private GitHub repo to act as the sync repo
git clone [email protected]: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 <path>.bak-<timestamp> before restore.

Configuration

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

Full reference: docs/configuration.md.

Plugins

Write a single .js file in ~/.shinobi/plugins/:

// ~/.shinobi/plugins/my-plugin.js
export default function register(registry, api) {
  registry.registerTool({
    name: 'plugin_count_open',
    description: 'Count open decisions across all projects',
    inputSchema: { type: 'object', properties: {}, additionalProperties: false },
    handler: (_args, api) => {
      const projects = api.listProjects();
      let total = 0;
      for (const p of projects) {
        total += api.listDecisions({ projectId: p.id, status: 'open' }).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.