Introducing Octomind: an Open-Source AI Agent Runtime in Rust

Don Hardman — Fri, 15 May 2026 21:13:41 +0000

Install agents, not frameworks.

Octomind is an open-source AI agent runtime written in Rust. One binary. 48+ pre-configured specialists. Real MCP host. No framework lock-in. Apache 2.0.

curl -fsSL https://octomind.run/install.sh | sh

Why we built it

After a year of shipping AI-native developer tools at Muvon, we kept hitting the same walls with existing agent stacks:

Framework lock-in. Pick LangChain, you marry Python and a graph DSL. Pick Autogen, you marry Microsoft's mental model. Pick an SDK-only path, you re-invent everything that should be infrastructure.
Brittle long sessions. Most agents fall off a cliff once the context window saturates. Real engineering tasks take hours.
Single-vendor lock-in. "Just use Claude" is fine until the API has a bad day, your bill triples, or a model is deprecated.
MCP as an afterthought. Servers added at startup, not runtime. No real host primitive.

Octomind is the runtime we wanted to use. Rust because we wanted atomic writes, type safety on the critical path, and one binary that ships without a Python venv.

What you get

48+ specialist agents across 12 domains — engineer, DevOps, security, lawyer, doctor, finance researcher, etc. Each is a YAML manifest, not a class hierarchy. You can write your own in 20 lines.
MCP host with dynamic registration. Add or remove MCP servers at runtime, not just at startup. This unlocks workflows where one agent installs the tools the next agent will use.
Multi-provider, zero lock-in. 13+ providers out of the box: OpenAI, Anthropic, Google, OpenRouter, Mistral, Ollama, vLLM, plus anything that speaks the OpenAI API. Switch with one config line.
Adaptive compression for 4+ hour sessions. A compression scheduler that watches the token budget and rewrites earlier turns into structured summaries before they fall out of context. Sessions that used to die at hour two now run overnight.
Hard spending caps. Per-session and per-request budgets that the runtime enforces, not a comment in your README.

A quick taste

# Run the engineer specialist on a fix
octomind run engineer "find the cause of the flaky e2e test in checkout"

# Mount an MCP server at runtime — no restart
octomind mcp add github --token=$GITHUB_TOKEN

# Swap the model on the next turn
octomind run engineer --model anthropic/claude-sonnet-4-6 "now propose the patch"

How it compares

	Octomind	Aider	Claude Code	Codex CLI
Open source	✅ Apache 2.0	✅ Apache 2.0	❌	❌
Language	Rust	Python	TypeScript	TypeScript
Provider lock-in	None (13+)	None	Anthropic only	OpenAI only
MCP host	✅ dynamic	partial	✅ static	✅ static
Specialist agents	48+ built-in	DIY	skills/subagents	DIY
Long-session compression	✅ adaptive	basic	rolling	rolling

Try it

Docs: octomind.run/docs
Repo: github.com/muvon/octomind
Specialists catalog: octomind.run/tap

We'd love your feedback — issues, PRs, or just star the repo if you want to follow along. The roadmap, the docs, and every commit are in the open.

Octomind is built by Muvon Un Limited, the team behind Octocode (Rust semantic code indexer with GraphRAG) and Octobrain (persistent agent memory). We build infrastructure for AI-native developers.