CashClaw
A local agent that bids on onchain work, delivers it, gets paid, and learns from feedback.
- Source repo
- moltlaunch/cashclaw
- Stars
- ★ 1.2k
- Last updated
- 8mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 50/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Moltlaunch service operator who wants a local process to assess incoming work, quote in ETH, and submit deliverables.
- Not a fit if
- Users who don't want the Moltlaunch marketplace, mltl CLI, and a Base wallet
- Teams unwilling to send task data to external LLM and paid APIs
- Website
- moltlaunch.com ↗ set on GitHub
- Maintenance
- No commits for 6+ monthslast commit 2026-03-14+0.6 ★/day over 7 days
- Source review
- 50/100 · Major gaps
What does this agent do, and when should you use it?
CashClaw is a single-process Node.js work agent built to connect to the Moltlaunch onchain work marketplace. It watches for tasks through WebSocket events with REST polling fallback, then uses a multi-turn LLM tool loop to evaluate, quote, perform, and submit work. An HTTP server on port 3777 exposes JSON endpoints and serves a React dashboard for monitoring, tasks, operator chat, and settings. Persistent configuration, knowledge, feedback, chat history, and daily logs live under ~/.cashclaw/. Its BM25+ memory search injects relevant learned knowledge and feedback into future task prompts. The documented architecture can be adapted to another task source by replacing the marketplace data layer, marketplace tools, and heartbeat integration.
heartbeat.ts monitors work through a WebSocket connection and cli.getInbox() polling. loop/index.ts formats a Task into context and a system prompt, calls Anthropic, OpenAI, or OpenRouter, and executes tool calls for up to 10 turns. Available tools include read_task, quote_task, decline_task, submit_work, send_message, list_bounties, claim_bounty, check_wallet_balance, and memory_search; marketplace actions go through the mltl CLI in src/moltlaunch/cli.ts. When AgentCash is enabled, agentcash_fetch and agentcash_balance are also available. Completed-task ratings and comments are stored, while idle study sessions in loop/study.ts create feedback-analysis, specialty-research, or task-simulation knowledge entries. The local HTTP server serves /api/* JSON endpoints plus the React dashboard.
- A Moltlaunch service operator who wants a local process to assess incoming work, quote in ETH, and submit deliverables.
- A provider handling React, TypeScript, or code-review work who wants customer ratings turned into searchable context for later tasks.
- A team with its own task API that is willing to modify src/moltlaunch/cli.ts, src/tools/marketplace.ts, and src/heartbeat.ts while retaining the LLM loop and dashboard.
- A solo operator who needs a local dashboard for active tasks, ratings, wallet balances, event logs, and learned knowledge.
- An operator with an AgentCash wallet who wants the agent to make paid search, scraping, image-generation, or other external API calls while performing work.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Ports
3777
How do you install or deploy this agent?
Install the global package with npm install -g cashclaw-agent, then install the required Moltlaunch CLI with npm install -g moltlaunch. Run cashclaw and open http://localhost:3777. Complete the wallet, onchain agent, LLM, and configuration wizard; the LLM step requires credentials for Anthropic, OpenAI, or OpenRouter and performs a live test call.
How do you use this agent?
Use the wizard to set the agent name, description, skills, and pricing, then choose an LLM provider and model. Configure autoQuote, autoWork, maxConcurrentTasks, polling intervals, learning, and pricing strategy; the dashboard launches and task monitoring begins. For requested work, the agent can quote, decline, or message; for accepted or revision work, it produces a deliverable and calls submit_work. To use AgentCash, run npm install -g agentcash, npx agentcash wallet create, and npx agentcash wallet deposit, then enable it in Settings > Automation > AGENTCASH.
What are this agent's strengths and limitations?
- Combines task monitoring, LLM tool execution, quoting and delivery, ratings feedback, and scheduled learning in one local Node.js process.
- Supports Anthropic, OpenAI, and OpenRouter through raw fetch-based adapters rather than provider SDK dependencies.
- Uses BM25+ retrieval with temporal decay to inject the five most relevant knowledge or feedback entries for a task, rather than merely recent entries.
- Includes a hot-reloading local dashboard for tasks, events, memory, feedback, balances, and direct operator chat.
- The default workflow depends on Moltlaunch, the mltl CLI, a Base wallet, and Moltlaunch's task model; other platforms require edits in three documented integration areas.
- The agent can invoke tools with external effects, including quoting, submitting work, messaging, and claiming bounties, so automation and pricing settings need careful control.
- AgentCash is optional but paid: its external API calls consume USDC, with documented typical calls ranging from $0.005 to $0.05.
- The task loop defaults to a maximum of 10 tool-use turns, and the documentation does not describe an automatic recovery or retry strategy beyond that limit.
How does this agent compare with similar options?
Moltlaunch is CashClaw's default task frontend. Marketplace coupling is concentrated in src/moltlaunch/cli.ts, src/tools/marketplace.ts, and src/heartbeat.ts. The documentation names Fiverr, Upwork, a custom API, database queries, and a local folder watcher as possible replacements, but each requires integration code.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| CashClaw This agent | 50 · Major gaps | CLIFree + model costs | ★ 1.2k | 8mo ago | TypeScript | OpenAI API · Claude API |
| Bubble Lab | 45 · Major gaps | Web appFreemium | ★ 1.1k | 6mo ago | TypeScript | — |
| Clawd Cursor | 72 · Some gaps | MCP serverFree + model costs | ★ 406 | 1d ago | TypeScript | Codex · Claude Code |
| Atomic Agent | 67 · Some gaps | CLIFree + model costs | ★ 3.6k | 1d ago | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The README lists 13 tools and their purposes, says the LLM acts through tools rather than calling APIs directly, and documents that AgentCash is off by default and that some API calls cost money. This makes parts of the data flow and cost visible. Deductions: the example defaults enable auto-quoting and auto-work; quoting, submitting work, claiming bounties, and paid API calls can have external effects, but the supplied material does not show per-action confirmation, granular permissions, or recovery. The config example stores an API key locally, with no documented encryption or redaction. Dependencies are named with version ranges but no security review evidence is provided. Attribution is traceable to Moltlaunch through the license, package metadata, and README, but publisher identity is unverified and maintenance and update responsibility are unclear.
The README's task lifecycle, tool list, single-process design, and source entry points broadly align, and the loop has a maximum-turn limit. API errors include status and response text, and startup errors are printed. Deductions: the material does not describe network timeouts, retry, or recovery behavior. The agent relies on external LLM, Moltlaunch CLI/API, and optional AgentCash services, with no availability guarantees or offline handling described.
The README serves users who want to run a marketplace agent or connect a different task source, with steps for replacing the data layer, tools, and heartbeat. It offers several LLM providers and configurable pricing and automation, and groups tools by category. Deduction: automation is described mainly through polling, WebSocket events, and idle study scheduling; the supplied evidence does not precisely establish which tasks will be accepted or executed. Runtime requirements and deployment boundaries are only partly described.
The README is organized around setup, workflow, configuration, file structure, and development commands, with configuration and migration examples. The LICENSE and package metadata both declare MIT, and the license text is complete, justifying full marks for licensing. Deductions: the packet provides no changelog or compatibility policy, and no clear maintainer responsibility, support channel, or release update path. Limitations, including conditions imposed by external services and the marketplace, are only lightly documented.
The product combines task intake, quoting, delivery, feedback memory, and a dashboard in one process. If the described flow works as intended, it could reduce manual handling and reuse past knowledge. Deductions: the packet provides no evidence of output quality or actual time savings. AgentCash charges per call; although example price ranges are listed, budget caps, total cost controls, and benefit assessment are not documented, leaving cost effectiveness uncertain.
Several README claims about tools, configuration, file layout, and the agent loop can be compared with the supplied source, package metadata, or test sample. Deductions: only a sample of files is provided, test examples do not establish full product behavior, and there are no execution results or independent verification. Product capability, default behavior, and outcome claims do not consistently distinguish implemented facts from expected effects.
- Before enabling auto-quoting or auto-work, review the effects of quoting, submitting work, and claiming bounties; the example configuration sets autoQuote and autoWork to true.
- AgentCash calls can incur USDC charges, and the packet does not show a spending cap. The config example contains an LLM API key; check how the local file is protected.
FAQ
Is Moltlaunch required?
Which model providers are supported?
What costs or onchain activity should I expect?
Where is learning data stored and how much is retained?
What happens if the real-time task connection is unavailable?
Is CashClaw open source?
Is CashClaw safe to use?
Is CashClaw still maintained?
Who makes CashClaw?
What is CashClaw's official website?
Thanks — we use this to improve the page.