Mecatl

Mecatl is Stacklok's open-source, cloud-native agent harness: the loop, tools, permissions, hooks, delegation, and service boundaries for running AI agents as production workloads on infrastructure you operate. Unlike a terminal coding agent, it treats the harness as a server problem — durable sessions, append-only event logs, leases, and a Kubernetes-native runtime (mecak8s) with Redis-backed state and disposable replicas.

The core claim is seam separation: the agent loop (engine/) is provider-neutral, client-neutral, and environment-neutral, and everything else — model provider, state store, filesystem, UI — plugs in through explicit ports (engine/port/: LLMProvider, SessionStore, SessionLease, PermissionPolicy, HookRunner, EventLog, Compactor). Shipped binaries include mecated (server), mecatui (terminal client that can host or connect), mecatequi (one-shot headless automation), mecak8s (K8s runtime), plus a Go SDK for embedding the engine.

Security posture is a first-class design axis: deny-dominant permissions with approval flows, read-parallel/mutate-serial tool dispatch, delegated runs whose derived capabilities only narrow, an audit trail with durable attribution, and an opt-in local microVM execution backend that keeps provider credentials on the host while filesystem and shell tools run inside the VM. The README is candid that cross-process cryptographic identity and tenant isolation are still open design work.

The architecture docs are unusually deep — a 2,200-line docs/architecture.md, per-subsystem guides, frozen ADRs, a domain model with conformance-test adapters for every port — and the repo is itself developed with Claude Code (.claude/agents/ and .claude/skills/ ship reviewer subagents and skills).