Verifiable intelligence for AI software engineering.

Try AxiomCode Graph on GitHub

Everything starts with a formally grounded code graph.

Request Demo

Works with your favorite agents

Cursor
Claude Code
Codex
Antigravity
Windsurf
GitHub Copilot
Gemini CLI
OpenCode
Cline
Amp
Cursor
Claude Code
Codex
Antigravity
Windsurf
GitHub Copilot
Gemini CLI
OpenCode
Cline
Amp

AxiomCode Graph

Better code startswith a graph.

A formally grounded map of your codebase. The foundation for agents and engineers to understand, plan, and act.

Give coding agents the callers, dependencies, and affected code behind a task, so they can make changes with better context.

the changecode it depends ontests that call itunrelated, skipped

Missed by AI agents alone: 67%0% with AxiomCode
More impact sites found: 85%Including indirect calls
Lower cost, fewer rounds: 48%4 tool calls, not 10

Context

Start from the task, not from a symbol.

One question, put to a type-aware call graph and answered five times over — every chain verified, and anything the graph cannot vouch for widening the answer instead of shrinking it.

claude — hono — zsh

dev@MacBook hono % claude

Tip: Use git worktrees to run multiple Claude sessions in parallel.

⏵⏵ auto mode on(shift+tab to cycle) · esc to interrupt · ← for agents

axiomcode graph hono · typescript · 8,657 call edges · 366 files

97.9%

of the suite ruled out
4,086 of 4,172 tests never run

38 / 11

methods / files that reach the change
bearerAuth among them

< 1s

per answer
all five calls, 0.16–0.80s

Five questions about one change, measured against hono — 373 source files, 137 test files. Every timing is that run's own. With 121 unresolved calls inside the impacted set, the selected tests remain a lower bound.

Explore AxiomCode Graph features

01 · UNDERSTAND HOW IT WORKS

Ask how the code works. Get the answer from the code.

Ask a question and the graph finds where the answer lives, then traces the flow through it, so your agent explains the real path instead of guessing from file names.

exploring a new codebase: up to 50% fewer tool calls than an agent searching alone

excalidraw · TypeScript

$ axiomcode context "how does StaticCanvas draw elements"

where the work is (nearest first):

hop 0 components/canvases/StaticCanvas.tsx -> StaticCanvas

hop 0 element/src/renderElement.ts -> drawElementOnCanvas

hop 2 renderer/staticScene.ts -> _renderStaticScene

$ axiomcode path StaticCanvas renderElement

StaticCanvas → renderElement: nearest at 5 hop(s)

3 call(s) + 2 containment hop(s) (a callback it defines, then calls from it)

verified: every printed hop is an edge in the graph

HOW AN ELEMENT GETS ON SCREEN

StaticCanvasStaticCanvas.tsx:33
The canvas component. It defines an effect that runs after each render.
useEffect callbackStaticCanvas.tsx:44
After each render, asks for a repaint.
renderStaticScenestaticScene.ts:535
Holds repaints to one per animation frame.
_renderStaticScenestaticScene.ts:274
Sets up the canvas, then loops over the elements in view.
forEach callbackstaticScene.ts:398
Runs once for each visible element.
renderElementrenderElement.ts:967
Draws one element with its opacity and offset.

Test selection, measured.

Bugs with all triggering tests selected

Benchmark

Measured on
real bugs.

On 748 held-out Defects4J bugs, AxiomCode included every triggering test for 94.9% of bugs while selecting a median of 0.98× as many tests as Defects4J’s run.

Explore the results, the method, and what a formally grounded code graph makes possible.

Read release article

Born in a Harvard CS lab