Caveman
Products
Agent toolkitlive

Cavekit

One file, one loop, zero sub-agents

A Claude Code plugin for spec-driven development over one durable SPEC.md. Three commands run the core loop; four optional commands add research and review when the change earns it. Test failures enter §B, while recurring bug classes harden §V. Frozen since August 2026: still installable and working, no longer in active development.

Product demo / Cavekit

interactive spec loop
SPEC.mddurable · survives context reset

# §V invariants

✓ email is validated before any DB write

✓ tokens are JWTs; org id from claim only

/ck:spec2 invariants tracked

Capability ledger

06
  1. 01Durable SPEC.md that persists across context resets
  2. 02Caveman encoding — ~75% fewer tokens than prose
  3. 03Every test failure becomes a §B bug entry; bug classes become §V invariants
  4. 04Core loop: /ck:spec, /ck:build, /ck:check
  5. 05Optional: /ck:grill, /ck:research, /ck:review, /ck:deepen
  6. 06Nine installed skills; no sub-agents, dashboards, or orchestration binaries

Product ledger

02
Language
JavaScript
License
MIT

Product index

08
01Caveman SkillThe MIT skill: output compression for Claude Code & 30+ agents.Compression / live02Caveman ProxyRecoverable local context compression for agents you already use.Compression / live03Caveman Agent SDKTypeScript controls for what each agent run can spend and what its context costs.Agent toolkit / live04Caveman BrowseToken-efficient browser automation: compressed a11y snapshots, byte-exact recovery.Agent toolkit / live05CaveGemmaCaveman compression baked into Gemma's weights.Compression / live06Caveman CloudThe byte-safe LLM gateway.Cloud / in development07Caveman CodeTerminal coding agent. Half the tokens.Agent toolkit / live08CavememPersistent memory your agents recall over MCP.Agent toolkit / live