Tokenmizer

What Shweta-Mishra-ai/tokenmizer MCP server does

Shweta-Mishra-ai/tokenmizer MCP server connects MCP clients to TokenMizer’s graph-backed session memory. It supports checkpointing and resuming sessions, graph and savings statistics, file analysis, and decision tracing. The graph records typed items such as decisions, tasks, files, errors, and goals rather than reducing the session to a plain summary.

A decision that is replaced remains represented as superseded, allowing the reasoning history to be followed later. The resume output is a compact projection containing active decisions, unfinished work, and unresolved errors. This makes the server relevant to long-running coding or research sessions where preserving rationale matters as much as preserving the latest state.

How it works

TokenMizer runs locally as a proxy between an OpenAI-compatible client and an LLM provider. Applications send chat requests to http://localhost:8000/v1; the proxy processes the request, updates its graph, applies compression and caching stages, and forwards the request to the configured provider. SQLite stores the graph, checkpoints, and ownership data.

The MCP process communicates with the proxy through TOKENMIZER_URL, normally set to http://localhost:8000. Checkpoint, resume, statistics, and decision-reasoning tools require the proxy to remain running. File analysis can run without the proxy.

The Shweta-Mishra-ai/tokenmizer MCP server is therefore an MCP interface to the shared graph, not a separate memory store for each client. A session saved from one supported tool can be resumed from another client using the same graph.

Setup and configuration

Install the Python package with:

pip install "tokenmizer[anthropic,cache]"

For the Anthropic configuration shown in the README, set TOKENMIZER_ANTHROPIC_API_KEY, then start the proxy:

export TOKENMIZER_ANTHROPIC_API_KEY=sk-ant-...
tokenmizer serve

Configure an MCP client with the tokenmizer-mcp command and point it at the running proxy:

{
  "mcpServers": {
    "tokenmizer": {
      "command": "tokenmizer-mcp",
      "env": { "TOKENMIZER_URL": "http://localhost:8000" }
    }
  }
}

The README documents configurations for Claude Desktop, Claude Code, Cursor, VS Code, Zed, and Codex CLI. If the executable is not on PATH, it gives python -m tokenmizer.mcp.server as an alternative command. Ollama can be used locally by setting the provider to ollama; the README also documents Docker deployment and additional provider configuration separately.

Tools and capabilities

The Shweta-Mishra-ai/tokenmizer MCP server provides these six tools:

• checkpoint_session saves a session into graph memory.
• resume_session loads a compact continuation block.
• get_graph_stats reports graph statistics.
• get_savings_stats reports token savings.
• analyze_file processes a file for relevant session information.
• why_decision follows a decision’s supersession chain and returns the recorded reasons and evidence for each step.

The optional session_id request field enables memory for a named project or session when using the OpenAI-compatible API.

Limitations and notes

The proxy must be running for checkpointing, resuming, statistics, and decision reasoning. File analysis is the exception and does not require the proxy. The README describes the system as local, but model-provider access still depends on the selected provider: Anthropic requires its configured API key, while Ollama can run locally.

The MCP configuration exposes the tools through a client, while ordinary model requests still use the OpenAI-compatible HTTP endpoint. The supplied material does not specify operating-system-specific requirements beyond the documented client configuration paths, nor does it provide a complete list of provider environment variables.