# subtask MCP server

> A local delegation scheduler that lets coding agents start parallel subtasks with spend caps and verifiable evidence.

- Repository: https://github.com/tedin7/subtask
- Page: https://hostedmcp.org/servers/tedin7/subtask
- Transport: stdio
- Stars: 0
- License: MIT
- Latest version: 0.5.9

## Install

### Claude Code

```sh
claude mcp add subtask -- uvx subtask-mcp
```

### Cursor (`~/.cursor/mcp.json`)

```json
{
  "mcpServers": {
    "subtask": {
      "command": "uvx",
      "args": [
        "subtask-mcp"
      ]
    }
  }
}
```

### VS Code (`.vscode/mcp.json`)

```json
{
  "servers": {
    "subtask": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "subtask-mcp"
      ]
    }
  }
}
```

### Claude Desktop (`claude_desktop_config.json`)

```json
{
  "mcpServers": {
    "subtask": {
      "command": "uvx",
      "args": [
        "subtask-mcp"
      ]
    }
  }
}
```

### Codex (`~/.codex/config.toml`)

```toml
[mcp_servers.subtask]
command = "uvx"
args = ["subtask-mcp"]
```

### Gemini CLI (`~/.gemini/settings.json`)

```json
{
  "mcpServers": {
    "subtask": {
      "command": "uvx",
      "args": [
        "subtask-mcp"
      ]
    }
  }
}
```

## Tools (5)

- `context_refs`
- `invariants`
- `allowed_paths`
- `oracle_cmd`
- `tier`

## FAQ

### What does subtask do for coding agents?

It acts as a local delegation scheduler that allows agents like Claude Code or Codex to start work in parallel, wait for completion, and receive a summary with evidence.

### How does the server handle spending and budgets?

The server treats money as a primitive, reserving budget for paid engines before work begins and ensuring every reservation is closed exactly once.

### Can I verify the work performed by a subtask?

Yes, you can use the verify option to run checks on the commit left by the worker, and the result tool can provide evidence including observed commands and exit codes.