At a glance
An early-stage, model-agnostic terminal coding agent that defaults to a hardened Docker container, supports direct host execution, coordinates bounded subagents, offers configurable multi-model counsel and two-round debate panels, connects MCP and optional browser tools, and separates session state from a provenance-aware durable memory store.
Good choice if
- Developers who want container isolation to be the default rather than an add-on
- Local-model and OpenAI-compatible workflows with explicit endpoint control
- Research-heavy coding tasks that benefit from bounded parallel subagents, multi-model counsel, and durable memory
Check before choosing
- The project is young and source-oriented; setup requires building its Docker image and managing configuration rather than installing a mature signed desktop product
- Direct mode executes tools on the host, while opting into the Docker socket gives the container control of the host Docker daemon and materially weakens isolation
- Durable memory is intentionally an injection channel: subagents can propose entries and some TTL-scoped writes become active without main-agent review, so untrusted work still requires the gate and memory auditing
See 3 more considerations
- Browser tools are optional and disabled by default; opening a URL in the host browser is explicitly outside the sandbox
- Mashūra debate can multiply external-model calls and shares successful first-round responses across providers for round-two critique; automatic approval can bypass the otherwise disclosed panel gate
- Repository tests are project-owned engineering evidence, not independent product-performance measurements