What's shipped, underway, and considered.
Anything not on this page is fair game for discussion. PRs and well-argued proposals reorder it constantly.
Shipped
5-
Phase 0 — Installable, runnable CLI
clawcodex, login, config, -p, --tui, --stream, --version.
-
Phase 1 — Core Claude Code MVP
The inline REPL, streaming replies, and session history.
-
Phase 2 — Real tool-calling loop
30+ tools across files, bash, web, tasks, agents, and MCP.
-
Multi-provider runtime
Anthropic, OpenAI, Z.ai GLM, MiniMax, OpenRouter, and DeepSeek behind one loop.
-
Dynamic workflow engine
agent()/parallel()/pipeline()/phase() with journaling, resume, and /deep-research.
Underway
3-
Phase 3 — Context depth
Richer context building, permission integration, and /resume-class recovery.
-
Phase 4 — MCP runtime depth
Deeper MCP client/runtime, plugins, and extensibility. Tools exist; platform work continues.
-
Memory and compaction
Deeper summaries and memory recall beyond workspace/git/CLAUDE.md injection.
Considered
2-
Phase 5 — Python-native differentiators
Capabilities that lean into the Python ecosystem rather than mirroring upstream.
-
Broader provider catalog
More OpenAI-compatible routes and self-hosted runtimes out of the box.
Ruled out
3-
Telemetry / phone-home
The agent runs on your machine. What happens there stays there.
-
Hosted SaaS dashboard
The terminal is the product; this website is informational only.
-
Required accounts
Bring your own API key. That is the whole onboarding.
Show up with a working prototype and watch "Considered" become "Underway".