Axon builds a living knowledge graph of your project — every dependency, every pattern, every tacit rule. Then it compiles precisely the context an AI agent needs to write code that actually works.
# Analyze. Brief. Execute. Verify. $ axon init -p ./my-project ✓ Graph: 847 nodes, 2341 edges ✓ 12 constraints discovered $ axon session "Add rate limiting to /api/upload" ✓ Brief compiled — 2847 tokens, 14 files ✓ 3 forbidden moves identified ⟶ Executing via claude-code... ✓ Verification passed: tsc ✓ tests ✓ ✓ Attribution: 4 files modified, 0 regressions
Static analysis + git history fused into a typed graph. Every artifact, dependency, co-modification pattern, and ownership signal — queryable and always current.
Task-specific briefs assembled from the graph. Seed resolution, dependency slicing, budget enforcement. The AI gets exactly what it needs — nothing more, nothing less.
Briefs include constraints the AI must not violate: owned files, invariants, failure modes. Mistakes are prevented before they happen, not caught after.
Every execution is followed by automated verification — type checks, tests, custom commands. Failures trigger retry with enriched context, not blind repetition.
Learns unwritten rules from failures and patterns. Hypotheses are generated, validated by humans, and promoted to operational constraints that shape future briefs.
Claude Code, Codex CLI, or any future provider — same execution contract. The orchestrator picks the right tool for each sub-task, not a one-size-fits-all model.
Static analysis via ts-morph + git history builds a typed knowledge graph. Every function, import, dependency, co-modification pattern, and ownership signal is captured as nodes and edges.
Given a task description, the compiler resolves seed artifacts, traverses the graph with task-type-specific strategies, and assembles a CompiledBrief — token-budgeted, priority-ordered, with forbidden moves and invariants.
The brief is injected as structured context into a CLI provider session. The agent writes code with full awareness of constraints, dependencies, and failure patterns — not just the file it's editing.
Automated verification runs type checks and tests. Failures feed back into the TKE as learning events, generating hypotheses about what went wrong. The graph gets smarter with every task.
Every component communicates through typed Zod-validated protocols. The graph is the single source of truth.
Give your AI agents the context they deserve. Axon turns a raw codebase into an execution surface that can actually guide autonomous work.