mimir-decisions

Decisions your agents can act on. MIMIR is a non-generative decision model: give it a context, a question and the options, and get back a typed answer with calibrated probabilities, the evidence behind it, and a certified verdict on whether to act or escalate. No text generated. Nothing to parse. Nothing to hallucinate.

It beats Laya and GLiNER2.5-Decide head-to-head on six of ten tasks — by 54.7 points on Banking77, 42.3 on MASSIVE, 36.3 on typed decisions — and where it cannot back an answer, it abstains instead of guessing.

pip install "mimir-decisions[local]"        # CPU engine
pip install "mimir-decisions[local-gpu]"    # CUDA engine
pip install mimir-decisions                  # data models and HTTP client only

Python 3.11+. Documentation: https://abderahmane-ai.github.io/mimir/

Why MIMIR

Most agents route, classify, and verify using a general-purpose language model: slow, expensive, and impossible to audit. MIMIR is built for structured decisions. It runs on ONNX Runtime in milliseconds, returns calibrated probabilities with every answer, and issues a mathematical certificate — a formal guarantee that its realised error rate stays at or below the risk level you ask for, measured on held-out data.

• No generation. Answers are drawn from the options you supply, not synthesised. The model cannot hallucinate an answer that wasn't on the list.
• Calibrated confidence. Probabilities are not softmax scores; they are calibrated to match realised accuracy on held-out data.
• Certified deferral. When confidence falls short of the certified threshold, the decision defers rather than guessing. The coverage and the error rate of taken decisions are proven.
• One typed contract. Seven decision types — choice, multi-choice, yes/no, verify, rank, rate, estimate — all returning the same result shape, over any context.
• Portable. The same Python interface works locally on CPU or GPU, over HTTP, and over MCP. Framework adapters exist for eight agent SDKs.

Quickstart

from mimir import Mimir

model = Mimir.from_pretrained("Mythologic/MIMIR-1")
result = model.choose(
    "My card was charged twice for the same order.",
    "Which team should handle this ticket?",
    options={"billing": "Billing: payments, refunds", "security": "Security: account access"},
)

result.status         # Status.DECIDED, Status.ABSTAINED or Status.DEFERRED
result.answer         # an option id, or None when no option applies
result.probabilities  # calibrated probability of each option id
result.certificate    # the certified threshold the decision was checked against

answer is the model's prediction. status is the policy's verdict:

• DECIDED — the answer is an option and is certified at the requested risk level.
• ABSTAINED — no listed option applies, and that is certified.
• DEFERRED — not certified; result.deferral.reason is below_threshold, out_of_distribution, or no_certified_threshold.

The first call downloads the model from the Hugging Face Hub at the revision this package version pins, verifies its Sigstore signature, checks every file against the manifest's SHA-256, and loads it.

Decision types

from mimir import Context, Field, Passage, Rate, Table

context = Context(
    passages=[Passage(title="Ticket #4412", text="The export has failed every night this week.")],
    tables=[Table.from_rows([["2026-03-02", "failed"]], header=["date", "status"])],
    fields=Field.from_json({"customer": {"plan": "enterprise", "seats": 240}}),
)
result = model.decide(context, Rate("How urgent is this?", ["low", "medium", "high"]), risk=0.01)

A context can be a string, a list of strings, a dict read as a JSON state, or a Context of typed passages, tables, and fields. Table.from_dataframe(frame) reads a pandas or polars DataFrame. decide_many batches multiple decisions, and every method has an async counterpart (adecide, adecide_many, …).

Certification

decide takes a risk level certified by the loaded policy (model.info().risk_levels). A decision is taken only when its calibrated confidence clears a threshold certified on held-out data to keep the realised error rate at or below that risk with 95% confidence, and when the context passes the out-of-distribution gate. decide_uncertified returns the raw model answer with no policy applied.

A certificate covers one exact configuration: model files, variant, ONNX Runtime version, execution provider, and options. On hardware not listed in the certificate, the first load runs the release's equivalence set and requires every decision to match. To certify thresholds on your own labelled data:

mimir calibrate labelled.jsonl --risk 0.01 --confidence 0.95 --out policy.json

model = Mimir.from_pretrained("Mythologic/MIMIR-1", policy="policy.json")

Remote use

from mimir import MimirClient

remote = MimirClient("https://mimir.internal", api_key="...")
remote.choose("...", "Which team?", options=["billing", "security"])

MimirClient has the same interface as Mimir, so all code, decision tools, and framework adapters accept either. It requires only the base install. Connection errors, timeouts, and 429/502/503/504/529 responses are retried with exponential backoff that honours Retry-After.

Decision tools

from mimir import Choice

route_ticket = model.tool(
    "route_ticket",
    Choice("Which team should handle this ticket?", options=["billing", "security"]),
    description="Route a support ticket to the team that owns it.",
)
route_ticket("My card was charged twice")
route_ticket.input_schema, route_ticket.output_schema

Tools can also be declared in a YAML file, which the HTTP and MCP servers load:

tools:
  - name: route_ticket
    description: Route a support ticket to the team that owns it.
    decision:
      type: choice
      question: Which team should handle this ticket?
      options: [billing, security]

Tool-call checks

A tool-call check decides, against rules you write, whether an agent's pending tool call may run. A certified yes allows it, a certified no denies it, and anything else escalates to a person.

check = model.tool_call_check(
    ["Refunds above 500 dollars need a manager's approval."], tools=["issue_refund"]
)
outcome = check("issue_refund", {"order": "4412", "amount": 900})
outcome.permission    # Permission.ALLOW, Permission.DENY or Permission.ESCALATE
outcome.reason        # one sentence for the agent or the approver

Agent frameworks

Each adapter turns decision tools into the framework's native tool type and wires a tool-call check into that framework's own approval hook.

from agents import Agent
from mimir.integrations.openai_agents import as_function_tool

agent = Agent(name="support", tools=[as_function_tool(route_ticket)])

Every framework also reaches MIMIR through its own MCP client. examples/ has a native, an MCP, and a checked agent for each framework, plus a Vercel AI SDK agent in TypeScript.

HTTP server

pip install "mimir-decisions[local,server]"
MIMIR_API_KEYS=key-one,key-two mimir serve --host 0.0.0.0 --tools tools.yaml

Concurrent requests are batched. With keys in MIMIR_API_KEYS, every route except the probes requires Authorization: Bearer <key>. A server with no keys listens only on loopback unless started with --allow-no-auth. The OpenAPI 3.1 document is openapi.json.

MCP server

Each configured tool becomes an MCP tool that takes only a context; --generic-tools adds mimir_choose, mimir_verify, mimir_rank, and mimir_rate. A deferred decision is a normal result telling the agent to escalate.

uvx --from "mimir-decisions[local,mcp]" mimir-decisions mcp --tools tools.yaml               # stdio
MIMIR_API_KEYS=... mimir mcp --http --host 0.0.0.0 --tools tools.yaml       # Streamable HTTP at /mcp
mimir mcp --tools tools.yaml --remote https://mimir.internal                  # forward to a server
mimir serve --mcp --tools tools.yaml                                          # HTTP API and /mcp together

In Claude Code:

claude mcp add mimir -- uvx --from "mimir-decisions[local,mcp]" mimir-decisions mcp --tools /path/to/tools.yaml
claude mcp add --transport http mimir https://mimir.internal/mcp --header "Authorization: Bearer ..."

Claude Desktop, Cursor, and VS Code take the same command or the same URL and header in their MCP configuration. The server is registered in the MCP Registry as io.github.abderahmane-ai/mimir.

Containers

docker run -p 8000:8000 -e MIMIR_API_KEYS=... -v mimir-models:/models ghcr.io/abderahmane-ai/mimir:1.0.0-cpu
docker run --gpus all -p 8000:8000 -e MIMIR_API_KEYS=... -v mimir-models:/models ghcr.io/abderahmane-ai/mimir:1.0.0-cuda

Images carry the runtime, never the model weights. On first start, the model is downloaded at the revision the package version pins, verified, and cached in /models. To run from that cache with no network access, append serve --host 0.0.0.0 --model-cache /models --offline.

Images are signed with Sigstore by the release workflow:

cosign verify ghcr.io/abderahmane-ai/mimir:1.0.0-cpu \
  --certificate-identity https://github.com/abderahmane-ai/mimir/.github/workflows/release.yml@refs/heads/main \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

Command line

Integrity

Releases are loaded from a pinned Hugging Face revision. Before any model file is read, the manifest's Sigstore signature is verified against the abderahmane-ai/mimir release workflow, every file is checked against the manifest's SHA-256, and the ONNX graph is checked against its operator allowlist and signature. No pickle is used anywhere.

Migrating

mimir.compat.systemone.v1 converts Jev /v1/systemone requests and responses, and mimir.compat.laya.v1 exposes load(...).predict(state, questions) in Laya 0.3.20's shape. See the migration guides for step-by-step instructions.

License

The mimir-decisions package is licensed under Apache 2.0. The MIMIR model weights are distributed under their own license on the Hugging Face Hub.