A code graph that resolves calls like a compiler, not like grep

What it does

It parses a repository and builds a knowledge graph of functions, calls, and files, then exposes that map to AI agents via MCP tools, a skill, and hooks. Agents query the graph instead of grepping, so they can see a task’s full scope — implementation, downstream effects, affected tests — before editing. The pitch: prune 99.9% of the codebase so the agent’s context goes to the task, not the search.

The interesting bit

Most code-graph builders use tree-sitter and match calls to declarations by name. AxiomCode resolves each call from the receiver’s type, the way a compiler does — which matters through interfaces, overrides, generics, and callbacks, where name-matching produces invented or missing edges. Since agents follow edges several hops deep, one wrong link silently loses everything beyond it. Unresolved calls stay explicitly marked rather than dressed up as facts.

Key highlights

• On 748 held-out Defects4J bugs, its selected tests included every bug-revealing test for 94.9% of bugs (best tree-sitter builder: 61.5%), F1 of 72.4.
• Call-target resolution scored against compiler ground truth: 96.5% on 33,257 Java calls, 88.8% on 9,829 TypeScript calls.
• Java, TypeScript, and Python are stable; JavaScript and C# engines are in beta, plus HTML/CSS/XML/YAML/Gradle support.
• Ships as a prebuilt binary for macOS, Linux, and Windows; runs fully offline, no Soufflé or compiler needed (one exception: Linux arm64 compiles native modules at install).
• Claims up to 50% fewer tool calls to explore a new codebase in their benchmarks.

Caveats

• The headline benchmarks cover Java and TypeScript only — the languages where a compiler provides independent ground truth; Python accuracy is unquantified.

Verdict

Worth a look if your agent routinely ships half-finished refactors because it missed a caller two files away. If you work in a single small file, grep still works fine.