MateClaw

The deciding trait is where it meets people. Agents reach your colleagues through the chat systems they already have open, which means the thing you built gets used by staff who would never install a command-line tool, and the request that used to arrive as a ticket arrives as a message the agent can act on.

That also tells you what it is not. There is no editing session here, no diff to approve, no repository in the middle. Pick it for internal workflows. Pick a coding agent for code.

Swapping the engine swaps the failure modes. A provider contract that lets the native graph engine or an external harness drive the same employee means two different loops, two different tool implementations and two different sets of bugs behind one configuration switch, and the documentation describes the interface rather than what differs across it.

What it does right is stream everything back. Thinking, text, tool calls, usage and completion all cross the boundary as events, so the child process is at least observable from the parent rather than being a black box that returns an answer.

1. Long-running agent work is a distributed systems problem wearing a new hat, and this is the first row on the board to name the primitives that literature settled on decades ago. A lease prevents two workers claiming the same goal, an attempt counter bounds retries, and a cooldown stops a tight failure loop.

2. Reconciliation after restart means durability was designed rather than discovered. 3. No evaluation is offered, and the claims here are structural, so none is owed.

A project this ambitious at this age is a bet on one group's stamina. Five surfaces, a plugin SDK and a governance layer is a roadmap a funded team would take two years over, and it has been shipped in months, which usually means either exceptional focus or a lot of surface that has never met a user.

Moat: none, and self-hosting means there is no meter to grow into one. Likely path: it finds a commercial sponsor, or it thins out to the parts its authors actually run. Position: watch it, and do not put a business process on it this year.

Somebody wrote this for my department rather than for a demo. Sixty people cost nothing because it runs on our own hardware, the health endpoints plug into monitoring my team already operates, and an approval gate on the dangerous actions means the compliance conversation has an answer instead of a promise.

What is missing is the paperwork behind it: no vendor, no support contract, no directory integration named, and per-channel error isolation is a design note rather than a service level. Approved with conditions: my platform team runs it, and the audit export is proven before any regulated process touches it.

Failover across a provider chain is the feature I keep hand-rolling. Naming three local runtimes inside that chain means my own machine is a peer rather than a fallback nobody tested, and when it stalls the request moves on instead of the session dying. My own servers attach as tools through the protocol.

The Java plugin SDK is the other half: extensions are compiled code with real access rather than descriptions handed to a model. Permissive licence, one JAR, nothing phoning anywhere. I can own this outright.