Work · Pre-release

The AI engine that remembers.

Mithra is a locally-runnable, open-weights, multi-model AI engine and agentic CLI — in the same category as Claude Code, OpenAI Codex CLI, and Gemini CLI, but defined by the one thing they lack: persistent, brain-like memory. Where frontier CLIs start every session from zero, Mithra learns topics, summarizes them, chains related ones, and recalls the right context on demand. It runs on your machine, on open weights you own — no vendor lock-in, no deprecation, nothing sent to a third party. Named for Mithra, the ancient Persian keeper of covenants and memory.

Persistent memorySpectrum multi-layer memory recalls project facts across sessions, not just the current context window.
Local-first & privateOpen-weight models served locally — Ollama on CPU, vLLM on GPU. Runs offline; code never leaves your machine.
OpenAI-compatibleDrops into existing tools and SDKs unchanged. The mithra CLI is just one client of the local gateway.

Talk to us about Mithra →

How it’s built

A composed engine, not one giant model.

Spectrum memory

The differentiator. A persistent multi-layer store that learns topics, summarizes them, chains related ones, and recalls the right knowledge on demand — the continuity frontier CLIs don’t have. Recall happens before generation; distilled learning happens after. The model fleet stays stateless; continuity lives here.

Model router

A mixture-of-models at the orchestration layer: classify each request — code, chat, reasoning, vision — and route it to the smallest sufficient open model behind one interface. Simpler and more debuggable than in-model expert routing, and it runs on modest hardware.

Local providers

One provider interface over serving runtimes: Ollama today for CPU-only machines, vLLM when a GPU is available. Open weights mean you own the models forever and run fully offline — no API to be deprecated out from under you.

Agentic CLI & MCP

A bounded tool loop — shell, file read/write, web fetch, sandboxed code exec — with MCP support for external tools, behind an OpenAI-compatible gateway. The mithra command gives an instant REPL and one-shot mode; existing SDKs talk to the same endpoint.

Why it exists

Own your engine. Keep your context.

No amnesiaContinuity across sessions via Spectrum memory, instead of re-explaining your project every time you open a new window.
No lock-inOpen weights and local serving. APIs can’t be deprecated from under you, weights can’t be hidden, and nothing is sent to a third-party endpoint.
Measured, not hypedEvery route, recall, and tool action produces traces. Model, prompt, memory, and router changes ship only when evals prove they helped.
Pre-releaseThe memory core, engine spine, agentic CLI, and multimodal seams are accepted; vLLM, fine-tuning, and packaging are the remaining hardening phase. Building deliberately, correctness first.

References

Model Context Protocol (MCP)The open tool protocol Mithra bridges so external tools plug into the agent loop.
Ollama — local model servingThe CPU-friendly runtime Mithra uses by default to serve open-weight models on your own machine.
vLLM — high-throughput GPU servingThe GPU serving path Mithra abstracts behind the same provider interface.