$ proposal
~$605 redirectable
252 model-less dispatches on expensive models matched mechanical routing classes.
FIX → NEW HOOK: route mechanical subagent dispatches to cheaper models
REVIEW →copy brief
Turn every agent session
into a better next run.
Built because we got tired of guessing what actually works.
ax watches every session your coding harness runs, spots the mistakes it repeats, and turns them into small, repo-specific fixes you review and apply — one at a time.
install in 30 seconds
paste it - your agent installs ax, labels your skills, and tells you which ones to actually use✓ Copied - paste into your coding agent for the guided setup
the loop, in your terminal
Your sessions, working for the next one.
Every session lands in one local graph your agent can reach back into — to catch a repeat mistake, replay a clean run, recover the moment a test fails, or grade the harness itself. Same loop, whatever you run it in.
open the dashboard, fixes are waiting
Your next fixes, already mined.
The Improve deck turns your history into ranked proposals — spend to reroute, failures that recur, tools that keep breaking. Accept one, and ax measures your next 30 sessions to see if it actually worked.
$ what's next
20 actions waiting
Mined from your sessions — savings to route, fixes that recur, verdicts due.
$ proposal
26× recurring
Feature closure needs stronger same-file follow-up verification.
FIX → NEW SKILL: post-feature verification checklist
REVIEW →copy brief
$ failing tool
Fix write_stdin cluster
1,681 failures across 205 sessions — one flaky tool taxing every run.
exits [2, 1, 130, -1, 127…]
DETAILS →copy brief
+17 more in the registry below
$ experiments Past bets, measured
No bets placed yet.
Every accepted fix becomes a bet. ax watches your next 30 sessions — a trace bar that shrinks to zero is a confirmed win.
Run ax serve to see yours — Improve deck, Agent Wrapped, sessions, skill triage.
the pipeline · what feeds the graph
auto · pausedclaude transcripts
~/.claude/projects
codex sessions
~/.codex/sessions
git + hook fires
repo + ~/.claude/hooks
typed local graph
surrealdb · 127.0.0.1
4,773 sessions · 369,132 turns
FTS median 5.9ms
FTS median 5.9ms
interventionsranked · safety-contracted · brief-only128 applied → agents
every event → typed graph → ranked interventions. local SurrealDB at 127.0.0.1. see /features for the schema →six forms: skill · guidance · subagent · hook · automation · harness_check
open source
If it shapes your agent, you should be able to fork it.
AGPL-3.0. Local-first. Typed end to end. Every signal lives in files and a database you own — no cloud, no lock-in. Commercial license available if you need it.
$
gh repo clone Necmttn/axCloned ax into ./ax$
cd ax && bun install1 284 packages installed in 4.2s$
ax daemon startax dashboard → http://127.0.0.1:1738$
AGPL-3.0
Free & open · commercial license available
TypeScript
End-to-end, strictly typed
Local-first
SurrealDB on 127.0.0.1
Forkable
Hack the loop, ship a PR
for teams
Free for your loop. Evidence for your team.
Everything above runs on your laptop, yours to fork. When you need to see how a whole team ships with AI — what’s spreading, what’s stuck, and what should become standard practice — that’s ax for teams.