caura-ai/caura-build-fleet

git clone https://github.com/caura-ai/memclaw-build-fleet.git && cd memclaw-build-fleet
python -m venv .venv && source .venv/bin/activate
pip install -r pipeline/requirements.txt
cp .env.example .env && python pipeline/run_pipeline.py --dry-run

Usage

• •Requires a free MemClaw account (MEMCLAW_API_KEY + MEMCLAW_TENANT_ID) plus an LLM that supports OpenAI-compatible function calling (cloud or local via Ollama, e.g. llama3.1).
• •Elevate Manager and Code Review agents to trust_level 2 once per tenant via the two curl PATCH commands, or they 403 on stats/list/insights.
• •Run python pipeline/run_pipeline.py (--skip-manager, --loop, --reset, --json-output for variants).
• •Query fleet memories from Claude Code: claude mcp add --transport http --header 'X-API-Key: mc_...' --scope user memclaw https://memclaw.net/mcp.

Key features

• ✓Hybrid recall fusing vector similarity + keyword match + knowledge graph traversal in a single call.
• ✓Per-agent allowlist of MCP tools enforcing least-privilege (the Manager is read-only and never receives memclaw_write).
• ✓Contradiction detection via memclaw_insights and cross-agent citation in Code Review verdicts.
• ✓Fleet namespacing (fleet_id) and row-level scope_agent ACL for per-agent memory isolation.
• ✓Plain Python with adapters for CrewAI, LangGraph/LangChain, and AutoGen/AG2, plus a REST compatibility mode.

Best for

Teams building or learning multi-agent pipelines that need governed shared memory, constraint propagation, and auditable data isolation across specialists.

Caveats

• ⚠Depends on the hosted MemClaw platform (free key required; enhanced stats/insights need elevated trust or a managed account).
• ⚠Code Review verdicts and pipeline health are LLM-dependent and not reproducible across runs/models — BLOCK is not a pipeline failure.
• ⚠At default trust, memclaw_insights and memclaw_stats only see the calling agent's own memories (the Manager sees 2 bootstrap seeds, not the fleet).
• ⚠Tenant/agent IDs normalize underscores to hyphens — a 403 on an otherwise correct key usually means the wrong ID format.