MetaClaw
A local proxy that turns personal-agent conversations into skills, memory, and optional scheduled LoRA learning.
Per-dimension scores and reasoning
Evidence: The agent runs as a transparent proxy, requiring access to user configs, API keys, and local files, but least privilege is not explicitly addressed. User confirmation: installation and configuration require user interaction, but auto-configuration and auto-install of plugins may occur without explicit consent. Data flow transparency: documentation describes proxy interception and skill injection but lacks detail on data flow and storage. Sensitive data handling: API keys and OAuth tokens are stored in local config, but encryption is not mentioned. Dependency security: dependencies are not pinned, posing potential vulnerabilities. External effects: the agent modifies config files of multiple external agents and restarts services, with broad impact. Rollback: uninstall command exists, but rollback for config changes is not documented. Source attribution: third-party components' origins and licenses are not clearly stated. Deductions: lack of security details and explicit least-privilege statements.
Evidence: Documentation and code are largely consistent, but there are inconsistencies such as 'OpenClaw' vs 'openclaw' casing. Dependency availability: dependencies are not pinned, risking incompatibility. Failure messages: no detailed error handling or troubleshooting guide. Deductions: unpinned dependencies and insufficient failure documentation.
Evidence: Target audience is clearly personal agent users, covering multiple agents and modes. Capability boundaries: supported agents and modes are documented, but limitations are not explicit. Trigger precision: conditions for skill injection and RL training are clear. Environment fit: supports multiple OS and agents, but system requirements are not detailed. Deductions: capability boundaries and system requirements could be more detailed.
Evidence: Information architecture is clear with README, config reference, and docs. Install notes are detailed with multiple methods. Naming stability: version numbers are clear, but naming conventions are not specified. Examples and FAQ: examples and demos provided, but FAQ missing. Known limitations: not explicitly listed. License: MIT license is clear. Versioning/changelog: news section provides version updates, but no formal CHANGELOG. Maintenance responsibility: maintainers or contribution guidelines not specified. Deductions: missing FAQ, known limitations, and formal changelog.
Evidence: Output usability: CLI commands and config are clear. Marginal value: unique features like skill injection and RL training provide high value. Cost-benefit: no GPU required reduces cost, but RL training requires external services. Deductions: cost-benefit not quantified.
Evidence: Claims of technical report and ranking are made but not verifiable from provided sources. Cross-source corroboration: no other sources provided. Fact-inference separation: documentation mixes facts and inferences, e.g., 'No GPU required' is an inference. Deductions: lack of verifiable evidence and clear separation of facts and inferences.
- The agent automatically modifies configurations of multiple external agents and restarts services, which may affect existing setups; test in a sandbox first.
- API keys and OAuth tokens are stored in local config without mention of encryption; users should protect config files.
- Dependencies are not pinned, posing supply-chain risks; consider using lock files or pinned versions.
What does this agent do, and when should you use it?
MetaClaw is a local proxy layer for personal agents including OpenClaw, CoPaw, IronClaw, PicoClaw, ZeroClaw, NanoClaw, NemoClaw, Hermes Agent, and OpenAI-compatible clients. It retrieves and injects skills on each request, then summarizes sessions into new SKILL.md files; when enabled, it also stores and retrieves user and project memory across sessions. Its RL mode asynchronously scores live conversation samples with a PRM and runs LoRA fine-tuning through Tinker, MinT, or Weaver. The default auto mode defers weight updates to sleep, idle, or calendar-meeting windows to reduce disruption during active use. Its deployment boundary is a local HTTP proxy plus configuration, skills, and optional memory data under `~/.metaclaw`; model and training calls use network services.
After metaclaw setup, MetaClaw writes configuration to ~/.metaclaw/config.yaml; metaclaw start launches a local proxy, normally at http://127.0.0.1:30000/v1, and can configure the selected personal agent. The proxy accepts OpenAI-compatible requests, while Anthropic-native clients such as NanoClaw can use /v1/messages; it retrieves SKILL.md files from ~/.metaclaw/skills/ according to skills.retrieval_mode and injects them into prompts. After a session, it creates skill summaries; with memory.enabled, it extracts and retrieves episodic, semantic, preference, project_state, and working_summary memories. With RL enabled, each conversation turn becomes a training sample, a PRM scores it asynchronously, and the Tinker, MinT, or Weaver backend selected by rl.backend runs GRPO LoRA training; optional OPD obtains per-token probabilities from a teacher’s OpenAI-compatible /v1/completions endpoint.
- An OpenClaw user who wants routine conversations to become reusable skills without manually maintaining a prompt library.
- A team using an OpenAI-compatible model API that wants to add skill retrieval and cross-session memory in a local proxy layer.
- A NanoClaw developer who needs the learning pipeline through an Anthropic-compatible
/v1/messagesendpoint without changing client-side calling conventions. - An experimenter with Tinker, MinT, or Weaver credentials who wants to run GRPO LoRA fine-tuning from real conversation traffic.
- A user who wants model updates deferred to sleep, idle, or calendar-meeting time rather than interrupting active work.
What are this agent's strengths and limitations?
- Provides a transparent local OpenAI-compatible proxy plus a
/v1/messagespath for Anthropic-native clients, covering several personal-agent integrations. - Injects skills on every turn and automatically summarizes sessions into individual SKILL.md files, with template or embedding retrieval.
- Decouples serving, reward modeling, and training; auto mode can defer RL updates to sleep, idle, or meeting windows.
- Adds structured cross-session memory for preferences, project state, and recent activity alongside skills.
- The RL path depends on external training services and credentials; Tinker is the default reference route, while MinT and Weaver require separate compatibility packages.
- RL weight hot-swapping can pause the agent for several minutes; auto mode reduces when this happens but does not remove that behavior.
- Automatic integrations modify or register local configuration for the selected personal agent and may restart its gateway or service.
- The supplied documentation does not establish pricing, data-retention policy, or guaranteed recovery behavior for PRM and training-service failures, so external API and training costs need separate evaluation.
How do you install or deploy this agent?
For the lightweight skills mode, run pip install -e .. For RL, skill evolution, and calendar scheduling, run pip install -e ".[rl,evolve,scheduler]". Then run metaclaw setup, select a personal agent, authentication method, and LLM provider, and enter the API base, API key, and model ID; API-key setup supports Kimi, Qwen, OpenAI, Volcano Engine, or custom providers. The first working invocation is metaclaw start; for skills only, use metaclaw start --mode skills_only. RL additionally requires rl.enabled plus credentials for a Tinker, MinT, or Weaver backend and a PRM.
How do you use this agent?
Run metaclaw start for the default auto mode; it waits for the local proxy to become healthy and then configures the selected personal agent. If no supported auto-configuration target is appropriate, set claw_type to none and point the client to base_url: http://127.0.0.1:30000/v1, api_key: metaclaw (or proxy.api_key), and your chosen model ID. Anthropic-native clients can set ANTHROPIC_BASE_URL=http://127.0.0.1:30000 and ANTHROPIC_API_KEY=metaclaw. Use metaclaw status to inspect proxy, mode, and scheduler state, and metaclaw stop to stop the instance.
How does this agent compare with similar options?
For online training, MetaClaw can use Tinker, MinT, or Weaver: Tinker is the default reference path, while MinT and Weaver each require a separately installed compatibility package plus their own rl.backend, endpoint, and credentials.
FAQ
Can I use it without RL?
metaclaw start --mode skills_only provides the proxy, skill injection, and post-session skill summarization without a GPU, Tinker, or another RL backend.Do I need to host a model locally?
Will it change my existing agent configuration?
claw_type selects a supported agent, metaclaw start automatically writes or updates that agent’s local configuration and restarts the relevant service. With none, you wire the client to the proxy yourself.Is long-term memory on by default?
memory.enabled to be set to true. Once enabled, relevant stored memories are injected into later requests.