MIMIR

What MIMIR MCP server does

The MIMIR MCP server exposes structured decision capabilities for agent workflows. Instead of generating a free-form response, MIMIR evaluates context against a supplied decision specification and returns a typed result. The available decision forms are choice, multi-choice, yes/no, verification, ranking, rating, and numeric estimation.

Results include the model's answer and calibrated probabilities. Depending on the operating mode and configured risk level, the result can also include evidence and a certificate supporting the applicable risk bound. The answer is still returned when the result is abstained or deferred; the status tells the caller whether it should act on that answer.

This makes the MIMIR MCP server suitable for tasks such as routing support requests, checking claims, selecting applicable labels, ordering candidates, assessing urgency, or estimating a value within a defined range.

How it works

An application supplies context as prose, structured fields, tables, or a combination of these. A context can also be a string, a list of strings, or a dictionary representing JSON state. The decision specification supplies the question and allowed options, or the relevant parameters for another decision type.

MIMIR supports three policy behaviors described in the documentation. Standard mode returns the calibrated highest-probability answer without applying finite-sample risk deferral. Threshold mode defers results below a caller-provided minimum confidence. Certified mode defers answers below the release threshold associated with the selected risk level and attaches a certificate when the answer passes.

The documented risk levels are 0.005, 0.01, 0.02, and 0.05. A lower risk setting establishes a stricter operating floor. A deferred result is intended for review, while an abstained result indicates that no listed option applies according to the model.

Setup and configuration

The local package requires Python 3.11 or newer. Install mimir-decisions[local] to use the local engine with CPU or CUDA support. Installing mimir-decisions without the extra provides the data models and HTTP client.

The first local model call downloads the pinned model revision from Hugging Face Hub, verifies its Sigstore signature, checks files against the manifest's SHA-256 values, and loads the model. The Python API can load the documented Mythologic/MIMIR-1 model with Mimir.from_pretrained.

For remote use, MimirClient connects to a MIMIR HTTP endpoint and accepts an API key. It presents the same interface as the local Mimir object. The client retries connection failures, timeouts, and selected HTTP errors with exponential backoff while honoring Retry-After.

The provided material confirms MCP support but does not specify an MCP launch command, transport configuration, or required environment variables.

Tools and capabilities

The MIMIR MCP server is built around typed decision tools. Supported specifications include:

• Choice: select one option identifier.
• Multi-choice: select the applicable option identifiers.
• Yes/no: return a boolean decision.
• Verify: classify a claim as supported, contradicted, or lacking sufficient information.
• Rank: order candidate identifiers from best to worst.
• Rate: select a level such as low, medium, or high.
• Estimate: return a number within configured bounds and provide a confidence interval.

The Python interface also supports batching through decide_many and asynchronous counterparts such as adecide and adecide_many. Contexts can include typed fields, passages, and tables; tables may be created from pandas or polars data frames. A decision can be converted into a callable tool with model.tool, allowing an application to define a named operation around a decision specification.

Limitations and notes

MIMIR only chooses from options supplied by the caller for option-based decisions, so the quality of the decision depends on the context and decision schema provided. The documentation notes that a short summary can produce a different result from a fuller context containing the original passage and typed fields.

Certificates apply to an exact configuration covering weights, Torch version, device, and hardware. On hardware not listed in the certificate, the first load performs the release's equivalence checks. Custom certification is available from labelled JSONL data through the mimir calibrate command, producing a policy file that can be supplied when loading the model.

The material describes MCP support but does not identify specific MCP client applications or provide server-specific installation instructions. It also does not state a project license.