MetaClaw

MetaClaw is a local proxy/agent runtime that sits in front of a user's personal agent (e.g., OpenClaw/CoPaw/etc.) and an OpenAI-compatible LLM API. It injects “skills” into prompts, can summarize sessions into new skills, and (optionally) runs asynchronous RL-style training (e.g., GRPO) on accumulated interaction data. It also includes a scheduler mode (madmax) to defer weight updates to idle/sleep/meeting windows and provides an Anthropic-compatible /v1/messages endpoint for Anthropic-native clients. A CLI (metaclaw) manages setup, config.yaml, starting/stopping, and daemon logging.

Evaluated Mar 29, 2026 (0d ago)
Homepage ↗ Repo ↗ Ai Ml ai-agents continual-learning meta-learning online-learning rl-training proxy memory openai-compatible anthropic-compatibility python cli
⚙ Agent Friendliness
41
/ 100
Can an agent use this?
🔒 Security
42
/ 100
Is it safe for agents?
⚡ Reliability
31
/ 100
Does it work consistently?

Score Breakdown

⚙ Agent Friendliness

MCP Quality
0
Documentation
55
Error Messages
0
Auth Simplicity
75
Rate Limits
10

🔒 Security

TLS Enforcement
35
Auth Strength
55
Scope Granularity
20
Dep. Hygiene
45
Secret Handling
55

Runs as a local proxy and uses API keys (upstream and proxy). TLS enforcement is not explicitly documented (examples show http://127.0.0.1), so network exposure depends on local-only deployment. No mention of fine-grained scopes, audit logging, or data minimization. RL/scheduler modes suggest accumulation and reuse of conversation data; the README does not specify retention, redaction, or training-safety controls. Dependency hygiene is unknown from provided snippet; manifest lists common libraries but no CVE/status information.

⚡ Reliability

Uptime/SLA
10
Version Stability
45
Breaking Changes
35
Error Recovery
35
AF Security Reliability

Best When

You want a locally hosted, agent-proxy workflow to continually evolve skills from live conversations, and you’re comfortable with CLI-driven setup and optional training dependencies/backends (Tinker/MinT/Weaver).

Avoid When

You need strong enterprise controls (RBAC, audit logs, strict policy enforcement) and a formally documented REST API with stable schemas; or you cannot permit the tool to patch other agents’ configuration files and restart local services.

Use Cases

  • Continual/online skill injection for an existing personal agent
  • Asynchronous training loops that run without interrupting interactive chatting
  • Building a meta-learning loop over real-world conversations
  • Integrating with multiple “claw” agent ecosystems via automatic config patching
  • Using an Anthropic-compatible interface layer for clients expecting /v1/messages

Not For

  • Production-grade managed SaaS usage without local operations/hosting
  • Environments that require strict guarantees about data retention or training safety controls
  • Teams needing a well-specified, stable public API contract (OpenAPI/SDK) for programmatic integration
  • Highly restricted environments where automatic filesystem/service patching is not allowed

Interface

REST API
Yes
GraphQL
No
gRPC
No
MCP Server
No
SDK
No
Webhooks
No

Authentication

Methods: API key for the underlying LLM provider (configured via metaclaw config / setup wizard) Proxy api_key configured for local client calls (example shows api_key: metaclaw)
OAuth: No Scopes: No

No OAuth or fine-grained scopes described. Authentication appears to be via API keys (local proxy key and upstream LLM keys).

Pricing

Free tier: No
Requires CC: No

MetaClaw is described as having no GPU cluster requirement; however, optional RL training may incur external service/compute costs depending on chosen backend.

Agent Metadata

Pagination
none
Idempotent
False
Retry Guidance
Not documented

Known Gotchas

  • Acts as a local proxy; local networking and port availability must be correct.
  • Service/config patching for many 'claw_type' agents may require permissions and OS-specific service control (systemctl/launchctl).
  • Anthropic-compatible endpoint (/v1/messages) behavior depends on correct routing/base URLs for Anthropic-native clients.
  • RL/training backend choice (tinker/mint/weaver) requires extra packages and credentials; misconfiguration may break training even if skills-only mode works.

Alternatives

Full Evaluation Report

Comprehensive deep-dive: security analysis, reliability audit, agent experience review, cost modeling, competitive positioning, and improvement roadmap for MetaClaw.

AI-powered analysis · PDF + markdown · Delivered within 30 minutes

$99

Package Brief

Quick verdict, integration guide, cost projections, gotchas with workarounds, and alternatives comparison.

Delivered within 10 minutes

$3

Score Monitoring

Get alerted when this package's AF, security, or reliability scores change significantly. Stay ahead of regressions.

Continuous monitoring

$3/mo

Scores are editorial opinions as of 2026-03-29.

5365
Packages Evaluated
21038
Need Evaluation
586
Need Re-evaluation
Community Powered