Mimir: AI Agent That Operates APIs by Intent

Every API endpoint becomes one safe, callable operator. GET /orders becomes get_orders(). You point Mimir at an OpenAPI spec, or just paste the docs, and the model reads it, normalizes it, and commits it as a set of typed, validated tools an agent can run.

Boundaries live in the runtime, not the prompt. Permissions, approval gates, and destructive-write flags are enforced in code. You can't jailbreak a policy that was never a sentence. The agent is able to act, unable to overreach.

The model never sees the key. Credentials are vaulted and injected at execution time. The LLM operates the system without ever holding the secret to it.

Every action is signed and logged. Runs are stamped to an operator and written to an audit trail, so "the AI did it" is always answerable with who, what, and when.

The part I'm most proud of: I engineered it to run on a local 7B model, not a frontier API. That forced real discipline:

Small-model-first architecture

Code enforces, not prompts

Eval-driven development

Structured output over brittle parsing

A non-negotiable rule: the AI understands the docs and normalizes them; code only validates against known types

No hard-coding one company's naming conventions, because the next API won't share them. It's a deliberately-engineered system, not a wrapper around a chat box. And it's executed real refunds end-to-end.

Frontend: Next.js 16 · React 19 · Tailwind v4 · TypeScript

Backend: FastAPI · Postgres · LiteLLM → Ollama

A custom design system built for this project. Carbon, bone, and one acid accent. Archivo Black. Flat brutalist. It's meant to feel like an instrument panel, not a landing page.