What is ContextIQ?

ContextIQ is a local, single-file AST code graph that gives AI coding agents token-efficient context. Instead of letting an agent read whole files (and re-read them every session), ContextIQ indexes your repository into a local SQLite graph of symbols + call / import / inheritance edges, then serves a token-budgeted "context pack" of only the symbols relevant to a task.

Naming: the product is ContextIQ. Its CLI is tokengraph_all.py, and the MCP server registers under the id tokengraph — those identifiers are kept stable so existing wiring keeps working.

The core idea

Without ContextIQ	With ContextIQ
Agent opens whole files to find relevant code	Agent asks for a task and gets only the relevant symbols
Re-reads the codebase every session	Graph persists in .tokengraph/graph.db across sessions
Risk of stale understanding after edits	Graph re-indexes incrementally before every query
Savings are invisible / unprovable	Every retrieval appends to a ledger; gain reports tokens + dollars saved

A context pack contains: relevant symbols (full body when small, signature when large), their callers / callees / base classes as signatures, matching indexed source chunks, and a list of anything dropped for budget so the agent can request it by name. Near-duplicate pieces are removed automatically.

What makes it different

• One file, zero required dependencies to run the CLI (tokengraph_all.py).
• Deep parsing for 25+ languages with a full call / import / inheritance graph, plus regex indexing for 30+ more. See Languages.
• Model-agnostic and offline — it emits context packs, never calls an LLM, so it works with any model, cloud or local. See Local LLMs.
• The full loop, not just retrieval — ask → validate → judge → verify closes the gap from finding context to trusting the answer.
• Grounded generation — detect house conventions, scaffold convention-matched files, and verify-plan before you act.
• Provable savings — a realized-savings ledger and a gain dashboard.

Next steps

• Start here: Quick start — install and run the real workflow in minutes.
• Daily driver: Retrieval · Validate · Judge · Verify
• Wire your editor: MCP server
• Reference: CLI · When to use what
• Proof: Savings · Benchmark