local-worker-mcp

Status (2026-09-10): kept as our own custom delegation tool, but not the default anymore -- DeepSeek Harness (dsh --profile worker "<task>") is, after a head-to-head benchmark on the same task/model showed its real bash/fs tools verify their own work more reliably out of the box. See ~/local-llm-bench for the comparison and tools/claude-hooks/README.md for the delegation guardrail that now points at dsh first. This server is still useful for cases dsh doesn't fit, and got a run_script tool added as a direct result of that benchmark (closing the verification gap it found).

An MCP server that exposes a local Ollama model (e.g. Qwen3.6) as an agentic worker -- with its own tool-calling loop (read/list/write files inside an allow-listed folder, search the web, get the real current date/time) -- for use from any MCP client: Claude Code, Claude Desktop, Continue, Cursor, whatever.

"Foreman/worker" pattern: the MCP client (you, or a larger LLM like Claude) keeps the full context and decides what's worth delegating; this server only ever executes the one scoped, well-specified task it's handed, with no memory of anything else.

Why this exists

Grew out of using Continue.dev with local models and hitting real instability in that extension (it hangs). Instead of continuing to patch a VS Code extension, this project separates the concern: a plain MCP server, no GUI in the loop, talking straight to Ollama.

Related MCP server: ollama-mcp

Architecture

MCP client (Claude Code, Claude Desktop, ...)
        │  tools/call "delegate_task"
        ▼
local-worker-mcp (this server, streamable HTTP)
        │  /api/chat with tools attached
        ▼
Ollama (localhost:11434, or remote over Tailscale)
        │  tool_calls: read_file / write_file / search_web / ...
        ▼
local-worker-mcp executes each tool call, scoped to workspace_root

Security (read this before using it)

• allowedRoots in the config is a real allow-list: the server refuses any workspace_root or path that isn't literally inside one of those folders -- empty by default, by design. You have to explicitly add folders via the dashboard or the config file before the worker can touch anything.

• trustedParents is the ergonomic escape hatch for the above: any subdirectory under one of these is auto-allowed without registering it individually -- defaults to ~/code, so any project you keep there (and whatever workspace_root your MCP client passes for it, e.g. the current Claude Code project directory) just works with zero manual setup. Use allowedRoots instead when you want something scoped tighter than "everything under this folder".

• Every write (write_file) is logged to ~/.local-worker-mcp/activity.log with the full path and size -- so you can audit what the model touched afterward.

• Explicitly tested that a directory-escape attempt (../../etc/...) is rejected by the server itself, not just "trusted" to the model behaving.

• The local model can get things wrong (see ~/local-llm-bench/RESULTS.md for real data on when and how much) -- always review steps in the response before trusting a file change.

Install / run

git clone https://github.com/deatherick/local-worker-mcp.git
cd local-worker-mcp
npm install
npm run build
node dist/server.js

Open http://localhost:8787 for the config dashboard (model, context length, allowed folders, etc. -- saved to ~/.local-worker-mcp/config.json).

To run it centralized (one Mac with Ollama, reachable from other machines over Tailscale/LAN) and connect from an MCP client on another machine, point that client's MCP config at http://<host>:8787/mcp (streamable-http transport), e.g. in Continue:

mcpServers:
  - name: Local Worker
    type: streamable-http
    url: http://your-host.your-tailnet.ts.net:8787/mcp

Exposed tools

• delegate_task(task, workspace_root, model?, think?, max_steps?) -- runs the full loop. Returns finalText plus every steps entry (which tool it called, with what arguments, and the result) so you can audit before trusting the outcome.

• list_worker_config() -- current config, so a client can confirm which model/folders it's talking to before delegating.

Development

npm run dev   # build + start

Roadmap / not included yet

• Proxying to other existing MCP servers (Git, Context7) from inside the worker's own loop -- today it only has its own search_web plus the native file tools.

• A shell-command tool -- deliberately NOT included in v0.1 given the risk; if added, it needs its own allow-list of permitted commands, not free execution.

• Auth on the HTTP endpoint -- currently assumes it runs behind a trusted network (Tailscale/LAN), not exposed to the open internet.

License

Apache 2.0.