Apache Maka Challenges Proprietary AI Agent Runtimes With Local-First Open Harness
Apache Software Foundation accepts its first AI agent-harness project, a vendor-neutral local-first runtime built for open models and reproducible evaluation.
PRO- Apache accepts Maka as its first AI agent-harness incubating project, a vendor-neutral local-first runtime.
- Runtime Event Log records every model message, tool call, result, and permission decision as append-only truth.
- Ships as macOS Apple Silicon desktop app, terminal TUI, non-interactive CLI, and Eval runner.
- Local tools include Read, Write, Edit, Bash, Glob, Grep, all gated by a permission engine.
- Project initiator is a PMC member on Apache Arrow, DataFusion, and Doris; incubator status here.
- Positioned as open harness for open-weight models against Terminal-Bench 2.0 and closed vendor CLIs.
The Apache Software Foundation just accepted its first agent-harness project into incubation. Apache Maka is a local-first AI desktop assistant and agent runtime that treats the execution trace itself as the source of truth, giving developers a vendor-neutral place to run open models against real files, shells, and tools without handing sessions to a hosted service.
The pitch is deliberately narrow. Every major lab now ships an official CLI agent, and the harness layer, the piece that binds a model to tools, permissions, and a loop, is quietly becoming proprietary. Maka positions itself as the open counterweight: a harness tuned for open-weight models that any project can audit, extend, or fork.
Log is the runtime
The core architectural idea is simple and unusual. Model messages, Tool Calls, Tool Results, and termination facts enter a Runtime Event Log. Sessions, UI, model context, and recovery are projections over that log. The transcript is not a side effect of the agent, it is the agent's state.
That design has practical consequences. Runtime Host owns Session, Turn, agent lifecycle, continuation, tools, and events. Eval owns only experiment semantics and results. Because everything runs through one execution authority and every event is durable, you can crash the process, resume a turn, branch a session, or replay a benchmark cell without losing the causal chain.
The team also draws a distinction that most agent frameworks blur: context is not history. Tool-result pruning and LLM compaction change what the next inference sees, while the recorded evidence in the log stays intact.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
