the entire crux codebase is now a single rust binary. 6.6MB, zero runtime dependencies. the python scripts, javascript plugins, bash wrappers — all gone. replaced by 47 rust source files that compile to one executable.

what shipped

66 MCP tools via rmcp over stdio. every crux capability exposed as a tool that any MCP-compatible client can call. claude code, opencode, cursor, windsurf — they all connect the same way, through .mcp.json.

66 slash commands — one for every tool. type /crux- in claude code and see the full menu. /crux-impact, /crux-recover, /crux-bg-run, /crux-overview.

4 LLM providers — ollama (local), anthropic, openai, openrouter. auto-selects based on available credentials. the 8B adversarial audit can now fall back to claude or gpt when ollama isn't running.

background agents — crux bg "run all tests and fix failures" spawns a detached process. crux bg list shows what's running. crux bg results <id> shows what happened.

terminal agent — crux agent launches an interactive REPL. picks the best available LLM, loads project context, maintains conversation history.

worktrees — crux worktree create experiment makes an isolated git branch workspace. work in it, then crux worktree done --merge or --discard.

codebase awareness — persistent project model with file roles (source/test/config/docs), module boundaries, language breakdown. /crux-overview shows the architecture at a glance.

auto-session logging — every MCP tool call automatically updates the session timestamp and writes to the interaction log. no AI cooperation needed. infrastructure over instructions.

the numbers

  • 66 MCP tools (64 real, 2 figma stubs waiting on API credentials)
  • 200 tests (180 unit + 18 integration)
  • 47 rust source files
  • 28 of 31 build plans converged
  • 0 open github issues
  • 4 LLM providers
  • 24 modes
  • single binary, 6.6MB

what's next

the three remaining NOT STARTED plans are IDE completion (017), sandbox execution (018), and voice coding (021). these need external integrations (LSP protocol, container runtime, whisper) that are longer-term work.

the autonomous evolution cron runs every 4 hours — prioritize work, check github issues, generate content. the system improves itself.