New session.
Same old mistake.

Belay debriefs your Claude Code and Codex sessions, then carries the lesson into the next one.

curl -fsSL https://getbelay.vercel.app/install | bash
Free. Local. 30 sec install.

One failure, across a week of sessions

typecheck · orbital-api · hover a point
Monday"done", build is red Tuesday"done", build is red Wednesdayyou catch it in review Guidance proposed stop hook · typecheck evidence pattern guidance review next run ready for your agent to review Marked done. The typecheck never ran. Done again. 14 type errors left behind. You caught it in review. Third time this week. Same cause three times. Belay drafts one rule. The three sessions, with the proof attached. One failure signature, not three separate bugs. Stop hook: run the typecheck before saying done. You approve it, edit it, or drop it. The next session starts with the rule in place.
Monday"done", build is red Tuesday"done", build is red Wednesdayyou catch it in review Guidance proposed stop hook · typecheck ready for your agent to review evidence pattern guidance review next run Marked done. The typecheck never ran. Done again. 14 type errors left behind. You caught it in review. Third time this week. Same cause three times. Belay drafts one rule. The three sessions, with the proof attached. One failure signature, not three separate bugs. Stop hook: run the typecheck before saying done. You approve it, edit it, or drop it. The next session starts with the rule in place.

Tap a point on the rope

Across the sessions, the agent said done while the typecheck still failed. Belay groups the repeated pattern and keeps the proof attached.

For the next run, your agent receives a proposed rule and verification command. You review it before it has any authority.

Built for the agents you already use.

Claude Code Codex Cursor Antigravity

Install once.
Start with the sessions you already have.

Your history,
in one place.

Belay finds the coding sessions saved on your machine and loads them into a local timeline. You don't need to start a new run to find something useful.

Loading your session history
Claude Code
Codex
Cursor
Antigravity
~/.belay/timeline.html

Your sessions. On your machine.

Existing history loaded. New sessions added as you work.

Find what kept
going wrong.

Belay flags loops, stalls, and failed commands across your history. Open a session to see the work that went nowhere, and where you had to step in.

Billing webhooks Session timeline
Edited six files across the API packages/api/src/billing/*
Ran the unit tests. All passed. pnpm test billing → 48 passed
Marked the task complete. Never ran the typecheck. pnpm typecheck → 14 errors, left behind
You stepped in. "the build is red again"
Third session this week to end this way. Now your agent can see that.

See what it cost you.
And what to say instead.

Your own agent debriefs each session: where your prompting or scoping cost time and money, and the exact message to send next time.

Billing webhooks · Debrief Written by your Claude Code

You named the files on your fourth message. Before that, the agent searched the whole repo three times.

13:38–14:02 searching for the files 14:02 edits and tests
Scoping about 24 min · roughly $6
What you did Asked to “fix the webhook retry bug” with no file names. The agent grepped three times and read 40 files before you pasted the paths.
What an expert would have done Named the two files and the failing test in the first message. You knew where the bug lived.
Say this instead The retry bug is in packages/api/src/billing/webhooks.ts and retry.ts. The failing test is webhooks.test.ts. Fix it there, no need to search the rest of the repo.
1 of 2 things to change · 1 to keep Stays on this machine

Have your agent
fix the cause.

Ask your agent to look at the failed sessions. It reads Belay's evidence through MCP and proposes a real fix: a hook, a rule, a project instruction. You review it before anything changes.

Your coding agent Connected to Belay
Why do my sessions keep ending with a broken build?

Three times this week I marked work done without running the typecheck. Nothing in this project stops me. I can add a hook that runs it before I'm allowed to finish.

.claude/settings.json
+ "Stop": [{ "hooks": [{ "type": "command",
+   "command": "pnpm typecheck" }] }]
Try it

Carry the lesson
into the next session.

Belay prepares reviewed guidance with the evidence, scope, and verification command attached. Your next agent starts with the lesson that matters instead of rediscovering it from scratch.

Mission Pack · typecheck before done
1

relevant lesson ready
for the next session

Original evidence attached
Scope and limits included
Verification command included

Preprint results

Lessons from past work helped on new tasks.

Tasks completed when a prior lesson was useful, across Claude Code and Codex.

11/ 16Belay
7/ 16Human-written instruction
6/ 16No additional guidance

This evaluation measured approved lessons delivered to agents; session debriefs have not been evaluated yet. Broader evaluation is underway; full methodology and agent-level results will follow in a research preprint.

Read the white paper Launch note

Your Belay data stays
on your machine.

Belay stores session history, evidence, and guidance in an encrypted local database and does not upload it to Doplex Labs. No account required. Read the trust details.

Local collection. Local storage.

Belay reads the session files already on your machine and keeps its records there. It works offline.

Your agent stays your agent.

Belay has no model. Optional analysis uses your configured Claude Code or Codex provider, under the account and policies you already use.

Your work keeps moving.

Belay never blocks an agent action. If Belay stops, your agent keeps going.

Belay for Teams · Enterprise

Coming soon.

One Belay across every machine at your company.

  • Fleet-wide patterns
  • One fix for everyone
  • Local-first, like solo

Make the next run
a little wiser.

Less going in circles. More getting somewhere. For you, and for your agent.

curl -fsSL https://getbelay.vercel.app/install | bash
Free. Local. 30 sec install.