Project intelligence for coding agents

Your agent has the code.Give it the decisions.

OneCue brings the decisions your project already made into your coding agent’s context — before it starts making changes.
One relevant cue, with its source. Or silence.

See it in action Read the docs
npx skills add onecueai/onecue-skill
Hooks for Claude Code + Devin · the skill loads wherever Agent Skills are read

01 / THE MOMENT IT MATTERS

The decision reaches the task.

One relevant cue. Or silence.
Explore an illustrative agent session.

Terminal examples
onecue / project-memory
AGENT SESSION~/your-project

❯ Add checkout to the booking flow.

↳ Read project memory product-scope.md

↳ Check relevance to this task

◎ ONECUE

This product does not collect payments.

Evercall confirms bookings and takes messages. It does not collect payments.

✓Ask before changing the decision — offer booking confirmation only.

❯ Ready for the next task

Demo sessionIllustrative playback · local skill

02 / HOW IT WORKS

Local first.
Built for the agent you already use.

Hooks do the remembering. The skill does the recall. Everything stays in your repo and on your machine.

01 / CAPTURE

Your sessions already contain the decisions.

Hooks log prompts and tool receipts locally; the skill saves explicit decisions as readable Markdown in .onecue/memories/.

Shipped · Claude Code + Devin

02 / SURFACE

One cue, or silence.

On each prompt — after the task is stated, before the first edit — the most relevant decision arrives with its source. Nothing relevant means nothing is said.

Shipped

03 / EVERYWHERE

The skill travels.

Loads wherever Agent Skills are read — Codex, Cursor, Gemini CLI, OpenCode. Automatic injection needs a hook protocol; Claude Code and Devin have one today.

Skill shipped · MCP planned

Start with one project. One command installs the skill; the next session starts writing to .onecue/.

Install OneCue

03 / PRACTICAL QUESTIONS

Asked, honestly.

Including the limits: a decision that was never recorded can’t be recalled, and a surfaced cue is evidence, not an order.

Why not just AGENTS.md or CLAUDE.md?
Those files are one hand-maintained blob; they can’t tell which decision matters for this task, and they go stale quietly. OneCue keeps decisions as separate records and surfaces the one that conflicts with what the agent is about to do — with the source it came from.
What does it actually read?
Sources you point it at: project docs and specs, plus sessions the local hooks captured and records under .onecue/. It does not scan your whole disk and does not upload anything.
What if a decision is outdated or conflicts?
The cue carries its source so the agent can check it, and it’s framed as a suggestion to verify, not an instruction. Deliberately changing an old decision is a normal flow — the agent is expected to ask, exactly like in the example above.
When does the context actually arrive?
On each prompt where a hook exists — after the task is stated, before the first edit. Session start alone isn’t enough when the task is still unknown; per-prompt injection is the timing OneCue claims.
Which agents work automatically?
Claude Code and Devin get hooks that inject a cue on each prompt. Codex, Cursor, Gemini CLI, and OpenCode load the skill wherever Agent Skills are read, and the agent recalls manually.
Does my code leave the machine?
OneCue’s storage and hooks are local. What the coding agent itself sends to its model provider is your existing setup — OneCue doesn’t change it either way.