Remit

What the remit MCP server does

The remit MCP server gives AI agents access to Remit, a statically checked language for workflows that call tools and models. Before execution, Remit analyzes which host-approved capabilities a program may use, the maximum number of calls to each capability, and the worst-case spend. It also tracks whether untrusted inputs, private data, or model output can reach sensitive arguments such as payment details or email recipients.

Rejected flows produce checker errors. A flow that requires human review can instead be tied to approval for its exact arguments. The runtime broker applies the same rules while the workflow runs and records a trace that can be replayed.

How it works

The remit MCP server is available locally through remit mcp, using stdio, or over HTTP with remit mcp --http. The hosted deployment is available at the MCP endpoint documented by the project. A client sends source-oriented requests to the MCP tools, while the underlying checker returns diagnostics and an authority manifest.

Remit programs can also be checked and run from the command line. Offline fixtures support testing without live operations. If execution stops for approval, the command-line flow can approve the digest for the exact call and resume the run with its run ID.

Setup and configuration

Install the project from its repository in a Python virtual environment, including the MCP extra. The documented quick start also installs pytest and pydantic, then runs the test suites and example workflows. A local client configuration starts the server with the remit mcp command.

For Claude Code, the configuration supplies remit as the command and mcp as its argument. The hosted mode can be started with remit mcp --http; the README notes optional API keys and usage metering for that mode. The project also documents how to deploy a copy of the website and MCP server.

Tools and capabilities

The remit MCP server exposes these tools:

• remit_guide returns the guide or complete language specification.
• remit_check reports checker errors and the authority manifest.
• remit_format produces canonical source formatting.
• remit_authority_diff checks whether an edit expanded authority.
• remit_examples provides complete example programs.
• remit_run_fixtures runs offline test fixtures.

Together, these tools support authoring, review, policy inspection, and local workflow testing. The authority-diff operation is useful for checking whether an agent's edit changes what a program is allowed to do.

Limitations and notes

Remit is described as an early open-source release. The language, checker, runtime, MCP server, demo, and five example applications are implemented and tested, but the project is not presented as finished. It does not include built-in sandboxing, so operators must run it inside an isolated environment they control.

The README also identifies the absence of a module system, concurrency, and an editor language server. Reported benchmark results use small samples, one model family, and no human trials; they should be read as project evidence rather than a guarantee for every workflow. The project is licensed under Apache-2.0.