Where agents
build agents.

Build complex agentic systems, not just agents. An ontology-grounded, auditable substrate in one Rust binary, and an agentic OS grown on it as proof.

meclaw is a framework for composing complex agentic systems, not just agents: fast, secure, ontology-grounded, auditable and made for agents to build with, in one Rust binary. Use it to try a new harness structure in an afternoon, to grow an agentic OS of your own, or to build the next thing before it has a name.

meclaw-os is the reference implementation: a complete agentic OS, grown on that substrate and nothing else.

Three things that make a meclaw agent unique. A voice cell that listens and speaks on a wire protocol built for it. A web cell that gives the assistant a display of its own, where it shows you instead of telling you. And a vault that keeps your secrets secret, with every cell inside a kernel sandbox.

meclaw talks to any OpenAI-compatible endpoint and to MCP servers.

quick start

One command

bash
curl -fsSL https://github.com/mmeyerlein/meclaw/releases/latest/download/start.sh | sh

It asks for an OpenRouter key, grows the assistant and asks it one question. Set OPENROUTER_API_KEY first and it does not ask.

proofs

The three proofs

/hard-shell

An attack becomes a routed message, without a key

One cell grows into three, and not one of them is an llm cell. The kernel sandbox refuses the metadata endpoint, and the refusal is a typed event on the route.

read the walkthrough →
/meclaw-os

Seventeen cells out of one POST

An empty seed, four templates, four edges, one POST /colony/mutations, and nothing restarts. It is the same endpoint the builder uses.

read grow.json →
/never-forgets

An answer out of months-old memory

Tell it something in January, ask in March. The memory hive is a subsystem no harness has.

read the walkthrough →

security

Three templates, all shipped

Every cell runs inside a kernel sandbox.
No operation of vault@1.3.0 returns a secret.
With access@2.5.0 a credential leaves the vault only sealed.

Security →

blog

Writing

  • 3 September 2026 · Marcus Meyer

    Where agents build agents

    Every agent framework ships you the same thing: a loop. Call the model, run a tool, feed the result back, until some condition you wrote says stop.

all posts →