Opaque agents are hard to trust.
When a coding agent hides what it can see, what it can run or which context it is using, developers cannot reliably review risk or reproduce behavior.
The harness stays editable — and owned by you.
Tools, settings, skills, MCP, database access and providers stay visible and interchangeable. The downloaded binary is the same open-source app maintainers use, forkable down to the agent loop.
- ✓All tool descriptions are editable and every tool is individually toggleable.
- ✓AGENTS.md and DESIGN.md let each repo carry its own operating manual and design system.
- ✓MCP servers are configured per workspace and picked up by the agent without hidden wiring.
- ✓No telemetry, no training on transcripts, no enterprise tier hiding the useful parts.
# House rules - Prefer dedicated tools over bash. - Keep the surface editable: every tool description lives in settings, every tool has an on/off toggle. - Ask before destructive database operations. ## tools mcp.* : per workspace, hot-reloaded skill : markdown in ~/.claakecode/skills team_run : parallel agent workstreams rollback : restore safe checkpoints
best for
Open-source AI IDE seekers
Permission-conscious teams
MCP users
Multi-provider developers
Full workspace workflows
No-telemetry users
Often described as an open-source alternative to closed AI coding IDEs, Claude Code-style coding agents, or Cursor-style editing workflows — not because it claims to be the fastest, but because the agent surface can be inspected, edited and restricted.