mcpcodeserver
Let the model write TypeScript that calls your other MCP servers, instead of making one tool call at a time.Read full description
zbowlingLABSCO SUMMARY
The agent executes; you steer.
Approval before each tool call and per-tool switches are the controls that make a local model with real tools comfortable to run — and hot reloading is what makes it usable while you are still writing the server.
What it is
An interactive terminal application for driving MCP servers from a local model. It supports the three MCP primitives — tools, prompts and resources — and gives you the controls to steer while the agent executes.
What you get
- Human-in-the-loop approval: review and approve each tool execution before it runs
- Agent mode for iterative tool execution when a model requests several calls, with a configurable loop limit and a choice at the limit — continue, wrap up, or abort
- Several MCP servers at once over STDIO, SSE and Streamable HTTP, with enable and disable per tool or per whole server during a session
- MCP prompts with argument collection, preview and safe rollback, and MCP resources for reading files, documents and structured data
- Streaming responses with Plain, Markdown, Both or Markdown-blocks display modes, thinking mode for models that expose reasoning, and reasoning effort levels from minimal through xhigh
- Fifteen or more model parameters to tune, including the context window size, plus an editable system prompt
- Server hot-reloading during development, so a change to your server does not mean restarting the client
- History you can view, export to JSON and re-import, per-query performance metrics with duration timings and token counts, and Ctrl-independent abort by pressing
aduring streaming
Requirements
Python 3.11+, Ollama running locally with at least one model pulled, and the uv package manager. Install with uv tool install --upgrade ollmcp or pip install --upgrade ollmcp, add a server with ollmcp mcp add, then run ollmcp. Servers can also come from script paths, URLs, a JSON configuration file, or Claude Desktop's config via --claude-desktop, which is no longer automatic. OpenAI-compatible providers such as OpenRouter and DeepSeek are selectable with --provider, with the key read from --api-key or $OLLMCP_API_KEY; keys from the environment variable are never written to the config file. Non-Ollama providers are described as experimental.
Setup effort
Follow the published steps — uv tool install --upgrade ollmcp
USE CASES
01
Running tools against a local model
Ollama is the default and needs no key, so the model, the tools and the conversation all stay on your machine.
02
Keeping a tool from firing without you
Approve each execution before it runs, and turn individual tools off for the session when you do not want them offered at all.
03
Developing an MCP server
Hot-reload the server after a change instead of restarting the client and rebuilding the conversation.
04
Stopping a runaway tool loop
Agent mode has a configurable loop limit, and when it is reached you choose to continue, wrap up or abort rather than watching it spin.
05
Keeping a record of a session
View the full history, export it to JSON for backup or analysis, and import a previous session to carry on.
BEFORE YOU INSTALL
What to know first
Requires
- Python 3.11+
- Ollama running locally, with at least one model pulled (ollama pull)
- uv
INSTALLATION
Connect your client
Server setup
TERMINAL
# Install ollmcp via uv
uv tool install --upgrade ollmcp
# or via pip
pip install --upgrade ollmcp
# Add an MCP server (example: playwright stdio server)
ollmcp mcp add playwright -- npx @playwright/mcp@latest
# Run the client (check optional flags with `ollmcp --help`)
ollmcp # once running, use /help for interactive commands
