Belay

AI coding agents are getting remarkably good at writing code. But they still make the same mistakes over and over.

A build fails because the agent forgot to run a check. A workaround gets rediscovered three sessions later. You explain the same project constraint again. The agent says “done,” only for you to discover that the typecheck is still broken.

The problem isn't that your agent can't solve the problem.

It's that it doesn't remember what mattered from the problems it solved before.

Belay is the experience layer for coding agents.

It works locally with the coding sessions you already have from Claude Code and Codex. Belay builds a timeline of that work, identifies patterns across sessions (loops, stalls, failed commands, repeated interventions) and keeps the evidence attached.

When Belay finds a lesson that looks worth keeping, it doesn't silently change your agent. It proposes a rule, hook, project instruction, or other piece of guidance based on the actual evidence. You review it first.

Once approved, that lesson can travel into the next session.

So instead of your agent starting every task from zero, it can start with the things you've already learned.

For example:

Your agent repeatedly marks a task complete without running the project's typecheck.

Belay notices the pattern across multiple sessions. It shows you the sessions, the failed commands, and where you had to intervene. It proposes a concrete fix (eg. such as a completion hook that runs the typecheck before the agent can finish).

You approve it.

Next time, the agent has the lesson.

That's the core loop:

Observe -> Find the pattern -> Preserve the evidence -> Propose a fix -> Review -> Carry the lesson forward.

Belay is deliberately local-first. Your session history, evidence, and guidance stay on your machine. There is no account required, and Belay does not upload your coding sessions to Doplex Labs.

Belay also doesn't replace your coding agent. It has no model of its own. It works alongside the agents you already use, including Claude Code and Codex, and optional analysis uses the provider you've already configured.

The goal isn't to make your agent autonomous.

It's to make your agent learn from the work you've already done.

Our first evaluation found that, on the evaluated tasks where a prior lesson was useful, Belay-assisted agents completed 11 of 16 tasks, compared with 7 of 16 with a concise human-written instruction and 6 of 16 with no additional guidance. This is an initial evaluation, not a claim of general performance; broader evaluation is underway.

Belay is free, local, and open source.

Teach the important lesson once.