What is OpenCodeHub?

AI coding agents have a structural blind spot. They can read a file, but they can’t see the graph the file lives in. That blind spot produces three failure modes every agent-driven workflow eventually hits:

• Missed dependencies. The agent renames a function and leaves callers untouched, because grep found a fraction of the call sites.
• Broken call chains. The agent changes a return shape, a handler two hops downstream crashes at runtime, and neither the agent nor its tests flag it. The relationship was never in context.
• Blind edits. The agent rewrites a critical-path function without knowing it sits on the hot path of multiple production flows, because nothing computed that ahead of time.

Grep is textual. Language servers are per-file. Embeddings are lossy. None of them answer the questions an agent needs answered before it writes a diff: what breaks if I change this, what depends on this, and where does this data flow.

The graph-first approach

OpenCodeHub parses your repository with tree-sitter (15 GA languages, plus SCIP indexers for TypeScript, Python, Go, Rust, and Java), resolves imports and inheritance, and materialises a typed symbol graph. That graph is stored in LadybugDB, a graph-native database, with DuckDB carrying the temporal sibling (cochanges and the symbol-summary cache). Both tiers are always present — there is no backend toggle, and a failure to load the @ladybugdb/core binding aborts the operation rather than falling back. BM25 lexical search and filter-aware HNSW vector search sit on the same store. A local MCP server exposes the graph to any agent that speaks Model Context Protocol.

Clustering, execution-flow tracing, and blast-radius analysis all happen once at index time. Agents get complete relational context in one tool call, not ten round-trips.

What you get in v1

• Graph-native storage. LadybugDB is the graph tier and a dedicated DuckDB sibling serves the temporal store. Both files (graph.lbug + temporal.duckdb) are written on every index — no backend knob, no fallback layout (ADR 0016).
• Cross-repo federation. Group several indexed repos with codehub group and query them through the group_* MCP tools. The repo is a first-class graph node and repo_uri carries through every cross-repo response, including the AMBIGUOUS_REPO envelope.
• Deterministic code-pack. pack_codebase (MCP) and codehub code-pack produce a reproducible 9-item BOM signed by the release workflow.
• WASM-only parsing. web-tree-sitter is the only parse runtime on Node 20, 22, and 24, with all 15 grammar .wasm blobs vendored in the @opencodehub/ingestion tarball. npm install -g @opencodehub/cli@latest does zero native builds and zero GitHub fetches (ADR 0015).

When to reach for OpenCodeHub

• Non-trivial refactors. Rename a function, change a return shape, or move a module and let the agent see every caller before it edits.
• Cross-file changes. Any diff that touches more than one file and crosses a module boundary.
• Blast-radius questions. “What processes depend on validateUser? What is the risk tier of this change?”
• Onboarding to a new repo. Ask the graph for the top clusters, HTTP routes, or authentication flow before the first edit.

Next, install the CLI and run your first query.