roadmap
What works, and what's next.
Karya is built in milestones, and each one ends in something that runs. The landing page describes the product at launch. This page is the honest version of where it is today.
Works today
- ✓Terminal UI on Bubble Tea: streaming, markdown, syntax highlighting, diffs
- ✓Modes in the TUI: normal, accept edits, plan (shift+tab)
- ✓Threads grouped by project, prompt history, queued messages
- ✓@ file mentions, ! shell commands, $EDITOR prompts, notifications, themes
- ✓Live streaming from Gemini behind the provider interface
- ✓Protocol v0 types, in-memory transport, request/response peer with tests
To launch
M0Contractin progress
- Protocol v0 package and in-memory transport
- Fake core with scenario files
- TUI rewired onto the protocol
- Dependency-rule test: the TUI never imports the core
M1Core loopplanned
- Gemini tool-call assembly and the session loop
- read, list and search tools
- JSONL session log and resume
- karya run for headless tasks
M2Edits and the sandboxplanned
- write and edit with diff approvals
- plan, default and accept-edits modes
- bash through sandbox-exec and bubblewrap, network off
- Record-before-run crash safety and interrupt
M3Budgets and undoplanned
- Step, token, cost and time limits
- Loop detection and verify commands
- Shadow-git /undo, /rewind and /redo
- Visible compaction and the end-of-task report
M4Breadthplanned
- Anthropic and OpenAI-compatible providers, switching mid-session
- Network proxy, profiles and fetch
- MCP tools and AGENTS.md
- Config files and karya doctor
M5Launchplanned
- Project identity and drift notes
- Export, import and end-to-end encrypted sync
- The full acceptance suite and a fixture benchmark
After launch
Interactive commands (PTY)Parallel tool callsBranching sessionsHooks and skillsSub-agentsEditor clientsGitHub ActionContainer and remote executorsNative Windows sandbox