Your first query
All three commands below run locally against the index at .codehub/.
None of them open sockets once the index is built (add --offline to
analyze to also prove zero network during indexing).
Hybrid search: query
Section titled “Hybrid search: query”codehub query fuses BM25 lexical search with HNSW vector search (when
embeddings are present) to find symbols related to a natural-language
concept.
codehub query "auth token refresh"Expected shape (abridged):
process: auth.refresh_token — 7 steps, 4 files 1. refreshToken (src/auth/handler.ts:42) 2. validateRefreshToken (src/auth/validate.ts:18) 3. ...
process: oauth.rotate_session — 5 steps, 3 files 1. rotateSession (src/oauth/rotate.ts:12) 2. ...Results are grouped by process (an execution flow the indexer
detected during clustering), not a flat symbol list. Use --limit,
--bm25-only, or --granularity file to change the shape.
360-degree context: context
Section titled “360-degree context: context”codehub context <symbol> returns callers, callees, and the processes
the symbol participates in.
codehub context PaymentProcessorExpected shape:
PaymentProcessor (Class, src/payments/processor.ts:24) callers: - checkout.charge (src/checkout/charge.ts:88) - subscriptions.renew (src/subscriptions/renew.ts:56) callees: - validateCard, stripeClient.createCharge, auditLog.write processes: - payments.checkout (12 steps) - payments.renewal (8 steps) ACCESSES edges: - ProcessorConfig (read), AuditLog (write)Pass --json to get the structured envelope that the MCP context
tool returns.
Blast radius: impact
Section titled “Blast radius: impact”codehub impact <symbol> reports how many symbols, files, and
processes depend on the target, plus a risk tier (LOW / MEDIUM
/ HIGH / CRITICAL).
codehub impact validateUser --depth 2Expected shape:
impact(target = validateUser, depth = 2) direct callers: 14 transitive callers (depth 2): 37 affected processes: 3 risk tier: HIGH confidence: 0.82--direction up restricts to dependents (who calls me), --direction down restricts to dependencies (who do I call), and --direction both
(the default) returns both in one response.
- MCP tools overview for the full server capabilities (29 tools across exploration, federation, scan, HTTP, and meta).
- Using with Claude Code to let the agent run these tools for you.