Get started
MIT licensed · CLI v0.4.0 on npm

Your agent keepsre-deciding what youalready decided.

OneCue remembers why your project chose Postgres over Redis, why the retry loop was rejected, why that incident happened — and puts the one relevant decision in front of your coding agent before it changes the code. One cue per prompt, with its source. Silence when nothing applies.

Try it free →How it works

Free, local-first · decisions live in your repo as Markdown

$ curl -fsSL https://onecue.sh/install.sh | sh

or npm i -g onecue-cli · skill only: npx skills add onecueai/onecue-skill

0.4.0
CLI on npm
6agents
Claude Code, Devin, Codex, Cursor, Gemini CLI, OpenCode
1cue
per prompt — or silence
How it works
See it work

The dead end your agent never re-enters.

One prompt, two outcomes. The difference is a decision your project already paid to learn.

claude — client-project
Without OneCue
"Make the queue faster."

The agent reads queue-worker.ts, finds no Redis usage, and proposes moving the queue to Redis — the approach already rejected after Redis reconnects duplicated queued jobs. You re-explain. Again.

With OneCue
"Make the queue faster."
decision · queue-storage — Keep Postgres SKIP LOCKED. Redis reconnects duplicated queued jobs. match 0.81 · source attached

The cue surfaces before the edit, and the agent tunes SKIP LOCKED batch size instead.

Scripted illustration of the local skill. Real store, real recall — the numbers are illustrative, not benchmarks.

Features

A decision store your agents can walk.

One canonical store, automatic capture and recall. Here is what you get to work with.

Local-first store

Decisions live in .onecue/decisions/ — git-visible Markdown committed with your repo. Diffable in PRs, reviewable like code, yours.

Hooks for Claude Code + Devin

Lifecycle hooks write and consult the store automatically. Tell the agent "remember this" and the skill captures it; recall happens on the next prompt.

Decision graph

--supersedes retires old decisions so recall surfaces only the current head; --conflicts-with warns instead of silently picking a side.

onecue doctor

Verifies runtime, store, hooks and decision-file integrity. Exits non-zero on warnings, so it fits CI.

Works where Agent Skills are read

Codex, Cursor, Gemini CLI and OpenCode load the same skill. One store, every harness your team runs.

Precision-first recall

onecue recall returns the one cue this context would surface — or silence. The store is never dumped into the prompt.

Trust

Runs on your machine. Reads like code review.

Built for teams running agents on client repos: nothing phones home, and every decision is a file you can diff.

Local-first

The store never leaves your machine — no account, no cloud. Your agent's own model provider still sees your code; OneCue adds no second hop.

Git-visible Markdown

Decisions are plain files in your repo. They travel with the code, review in the PR, and load in whatever agent reads the project.

MIT licensed

The whole CLI and skill are MIT on GitHub. You need not trust this page — read what the code does.

CI-ready

onecue doctor exits non-zero on warnings. Wire it into CI and a broken setup fails the build, not the next session.

v0.4.0 on npm · MIT · hooks for Claude Code + DevinRead the source →
Roadmap

Live today. More in development.

The free local skill and CLI are available now. These pieces are planned, not shipped:

  • In development

    OneCue Cloud

    Shared project intelligence across machines and teammates. The launch list collects an email on the Cloud page.

  • In development

    Background verification

    A loop that re-checks saved decisions against changing code, reconnects evidence, and refines the cue before it reaches the next task.

FAQ

Frequently Asked
Questions

Common questions about how OneCue works.

Memory tools maximise recall — they save a lot and replay it. OneCue maximises precision: the recorded decisions that bear on the change — usually one, a few when it spans several — surfaced only when a change would break them, or silence. It keeps the why, not a transcript, and it works across agents rather than one.
What was chosen, why, what it rules out, and when to revisit — kept as a small readable record in .onecue/, not a memory dump. When a task conflicts with one, that record is surfaced with its source.
Those are one hand-maintained blob that goes stale quietly and can’t tell which line matters for this task. OneCue keeps decisions as separate records and surfaces only the one that conflicts with what the agent is about to do.
Yes. Decisions live in a git-tracked .onecue/ folder, so they travel with the repo and load in whatever agent reads the code. Commit them like any other file; a second machine or a different agent gets the same decisions.
Claude Code injects a cue on each prompt through hooks. Codex, Cursor, Gemini CLI, and OpenCode load the skill wherever Agent Skills are read, and recall on request.
OneCue’s store and hooks stay local — no account, no cloud. Your coding agent still sends code to its own model provider; OneCue doesn’t change that and doesn’t add a second hop.
A cue is evidence, not an order. It arrives with its source so the agent can check it, and deliberately changing an old decision is a normal flow — the agent is expected to ask first.

Try it on one repo — two minutes.

Run the installer in a client project, ship a day's work, and watch the next prompt arrive with the decision already attached. onecue uninstall removes only the hooks — your decisions stay.

Install OneCue →