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

Install agents, not frameworks.

, 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

Try it

• 
  Docs:
  


• 
  Specialists catalog: , the team behind (persistent agent memory). We build infrastructure for AI-native developers.
  Vollständiger Original-Bericht

  Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf dev.to.

  ↗ Original-Artikel auf dev.to lesen