clay-good/OpenLore

npm install -g openlore && openlore install   (auto-detects agent, wires orient(), registers MCP server, builds index)
No API key needed for analyze/drift/mcp/init and all governance tools

Usage

• •Ask your agent orient("add a payment method") to start already knowing the relevant functions, callers, specs, tests
• •Before commit: use blast_radius, certify_public_surface, change_impact_certificate, verify_claim, and openlore enforce --hook
• •openlore prove --estimate projects whether the tool pays for itself on your repo (no key)

Key features

• ✓One-time static analysis (tree-sitter) builds a live SQLite knowledge graph (call structure, types, tests, decisions, IaC) — app code, IaC, and decisions on one graph
• ✓73 MCP tools: orient(), search_code, analyze_impact, trace_execution_path, select_tests, find_dead_code, blast_radius, change_impact_certificate, certify_public_surface, verify_claim, record_decision, and more
• ✓Governance guardrails: breaking/non-breaking verdicts with named consumers, differential sensitive-boundary checks, confirmed/refuted/unverifiable claim verification with citations, one commit gate (advisory by default)
• ✓Epistemic Lease tells the agent when context has gone stale; self-invalidating memory that survives renames; federation (cross-repo) and SCIP export
• ✓Measured: -26% round-trips on deep tasks in large repos; ~430µs p50 orient on a 15k-node graph; 5500+ tests; no telemetry

Best for

Large, unfamiliar, or polyglot codebases (code + IaC) where agents need deterministic structural context and pre-commit safety verdicts without any LLM in the retrieval path.

Caveats

• ⚠Static-analysis only: reflective dispatch, computed eval/DI registries, and cross-language bridges may be silently absent from the graph
• ⚠Best on deep/large-repo tasks; on small familiar repos it can add overhead (the README publishes the losses next to the wins)
• ⚠API key only needed for optional spec generation (openlore generate); everything else works offline deterministically