A freelancer that learns from bad reviews

What it does

CashClaw runs as a single local Node.js process that connects to the Moltlaunch marketplace, evaluates incoming tasks, quotes ETH prices, and completes deliverables through a multi-turn LLM tool-use loop. It handles the full client lifecycle—negotiation, execution, revision, and payment—then logs client feedback to refine future work. The README notes you can rip out the Moltlaunch integration and wire it to other marketplaces or your own clients.

The interesting bit

The self-improvement loop is unusually concrete. Between tasks, the agent runs scheduled study sessions—analyzing past ratings, researching its declared specialties, and simulating fake tasks—to generate structured knowledge entries. These insights are indexed with BM25+ search and decayed by age (30-day half-life), then injected into the system prompt when a relevant real task arrives, so the agent recalls what worked instead of just what happened last.

Key highlights

• 13 tools spanning marketplace actions, utility lookups, and optional paid API access via AgentCash
• Raw fetch() to Anthropic, OpenAI, or OpenRouter with an adapter translating between their tool-use formats—zero SDK dependencies
• BM25+ memory search with temporal decay, plus an active-recall tool the LLM can invoke mid-task
• Hot-reloading dashboard for monitoring, task review, operator chat, and config tweaks
• Atomic file writes to local JSON stores for config, knowledge, feedback, and logs

Verdict

Worth a look if you’re building autonomous agents that need to persist experience across sessions and operate in a closed feedback loop. Skip it if you want a general-purpose LLM framework without marketplace or payment plumbing.