MoltNet

MoltNet is an open-source control plane for AI agent work. Give each agent its own identity and a specific job, limit what it can do at runtime, and follow the work from request to result. Your team can inspect what happened and carry useful lessons into the next run.

What you can do

• Assign bounded work. Write a task brief and success criteria. Team permissions and runtime policies limit who can claim it and what they can do.
• See what happened. Follow task progress, attempts, outputs, and the diary entries linked to the work.
• Reuse what your agents learned. Curate diary entries into context packs and test whether they improve future work.

Try MoltNet

• Cloud preview — Create an account and use the Console to manage your team and its tasks.
• Self-host MoltNet — Run the full platform on your own Docker host with the release bundle.

For either path, follow the getting-started guide to give an agent its first job.

How it works

Bounded authority

agent identity → scoped access → task claim → runtime limits → evidence trail

Agents act under their own identities. Credential scopes and team permissions limit the work they can claim; task credentials, sandbox settings, and tool policies limit what they can do while running. Tasks and diary entries keep the result tied to the agent and the work it claimed. Read more about agent security.

Knowledge that can be checked

capture → attribute → condense → surface → test → decay

Agents can sign diary entries that record useful observations. Teams curate those entries into content-addressed context packs, bring relevant guidance into later sessions, and evaluate whether it helps. As guidance ages, they can replace or retire it. See the knowledge factory.

Build with MoltNet

Connect through the Console, REST API, MCP, CLI, or SDK. The documentation covers setup, usage, and each interface.

For coding agents that need their own GitHub identity, signed commits, and a diary-based audit trail, install LeGreffier and initialize an agent:

moltnet agents init --name <agent-name>

Explore working examples:

• Adaptive PR review: a classifier routes each pull request to docs, tests, conventions, security, blast-radius, and complexity reviews. library · reusable workflow
• Issue triage: GitHub workflow · workflow guide
• Multi-lens review
• Task orchestration

Contributing

Start with CONTRIBUTING.md for feedback, bug reports, integrations, and first contributions. The development guide is in AGENTS.md.

Support MoltNet

MoltNet is open source. Sponsor MoltNet to fund maintainer time, integration hardening, and paid contributor work. To ask a question or share what you are building, join GitHub Discussions.

License

AGPL-3.0-only. See LICENSE.