HeyAira

Switch tools, not context

HeyAira connects your AI chats and coding agents to one project state, so each picks up where the last one left off.

Think in chat

Write the PRD and docs, define tasks, review results and decide what’s next. Save your agent’s limits for the code.

  • ChatGPT
  • Claude
HeyAiraone project state

Build with agents

Send a task to Codex or Claude Code on your Mac. By default a second agent checks the work, and you get a pull request.

  • Codex
  • Claude Code

Not a chatbot or a model router. Your code stays in Git, referenced, never copied.

Read the docs Follow @HeyAiraApp

Works with the AI tools you already use

  • ChatGPT
  • Claude
  • Codex
  • Claude Code

Chats connect over the Model Context Protocol with OAuth. Agents run through the HeyAira app on your Mac.

Problem

Each AI keeps its own copy of your project

What you decided in ChatGPT, Codex has never heard of. What Claude Code built, your next chat cannot see. So you either do everything in one tool, or carry the project between tools by hand.

Consequences

You pay in limits and rework

  • Talk burns your agent’s limits

    Planning, reviews and back-and-forth happen where the code is, and the agent runs out before the work does.

  • You stay with one model

    Another model would suit the next task better, but switching means starting over.

  • You explain the project again

    Every new session begins with a recap you write by hand, and it is never complete.

  • Decisions get reopened

    The next agent does not know what was settled or why, so it argues it again or quietly reverses it.

Solution

One project state, outside every tool

Every decision, task and result ends up in HeyAira. Chats and agents read from it and write back to it, so the project keeps one record, whichever tool touched it last.

Talk it through

Plan and decide in any chat. Every decision and task is saved to HeyAira.

  • ChatGPT
  • Claude

Do the work

Codex or Claude Code takes a task on your Mac, in a sandboxed worktree. By default a second agent checks every criterion.

  • Codex
  • Claude Code

Keep the code

Code stays in Git. The work arrives as a pull request, checked on GitHub, with CI recorded.

  • GitHub

Ship it

You review and merge, then deploy from Git as usual. Agents never merge or deploy.

any host · any pipeline

Pick up anywhere

A new session opens with the current state, not a blank page.

every connected chat and agent

What runs where

Five floors, one record

Chats and agents never talk to each other. Tasks go down through HeyAira to your Mac and into Git; receipts, CI results and notifications come back up.

HeyAira as five floors Notifications on top, then chat, then HeyAira, then your Mac with the agents, then Git. Tasks go down; receipts, CI results and notifications come back up. Notifications when something needs you push first · e-mail if no phone got it iPhone E-mail open the task Chat think, decide, define tasks any client that speaks MCP ChatGPT Claude MCP · OAuth HeyAira one project state backed up every 15 min · encrypted · EU Memory Tasks Receipts task Your Mac one agent works, another checks your own accounts · several Macs HeyAira app Codex Claude Code OpenRoutersoon pull request Git code stays yours referenced, never copied · several repositories Pull request · CI receipt · CI result · notification
HeyAira as five floors Notifications on top, then chat, then HeyAira, then your Mac with the agents, then Git. Tasks go down; receipts, CI results and notifications come back up. Notifications when something needs you push first · e-mail if no phone got it iPhone E-mail open the task Chat think, decide, define tasks any client that speaks MCP ChatGPT Claude MCP · OAuth HeyAira one project state backed up every 15 min · encrypted · EU Memory Tasks Receipts task Your Mac one agent works, another checks your own accounts · several Macs HeyAira app Codex Claude Code OpenRoutersoon pull request Git code stays yours referenced, never copied · several repositories Pull request · CI

How it works

Define, build, review

  1. Define in chat

    Shape the task in ChatGPT or Claude: the goal, the decisions and the acceptance criteria that mean done. Choose the agent and model, then send it.

  2. Build on your Mac

    The HeyAira Mac app runs Codex or Claude Code in a sandboxed worktree. By default a second agent checks each criterion and sends the work back once if something is missing.

  3. Review the pull request

    The Mac app opens the pull request. HeyAira checks it on GitHub and records CI. A task that gets stuck sends a push to your phone; open the task to see the reason.

There is no self-service sign-up yet: the operator creates your project and its access token. To start, add the project’s HeyAira MCP address to each chat and authorize it with OAuth, using that token once. Agent work needs the HeyAira app on a Mac.

Proof

Five real tasks, no one at the keyboard

On 4 October 2026 five tasks were sent at once and ran to the end with no operator action.

  • Four pull requests

    Opened automatically from the agents’ accepted work, ready for review.

  • Two self-corrections

    The verifier sent work back, and the agent fixed it.

  • One blocked, with a reason

    The task stopped, said why, and the phone got a push.

  • Still reviewed

    Every agent pull request is reviewed independently before it is merged.

Roadmap

Where it stands

Released7

  • Project memory and tasks

    Decisions, notes and tasks in one place, with version checks so nothing is overwritten.

  • Session recovery

    One call returns where the project stands, so a new session starts informed.

  • Per-project access

    OAuth 2.1 with PKCE. Each project has its own address, and a client gets only the scopes it needs.

  • Agent tasks on your Mac

    Codex or Claude Code runs each task in a sandboxed Git worktree, several at once, and you can watch it live.

  • Independent verification

    By default a separate verifier agent checks every acceptance criterion. The worker gets one correction.

  • Checked pull requests

    The pull request is checked on GitHub and CI is recorded. A failed check reaches your phone and can go back to the agent.

  • Follow-ups without cleanup

    A re-sent or follow-up task starts on its pinned base without anyone tidying the Mac’s working copy first.

In progress1

  • One execution path

    Every run is a versioned task; the older ways of starting agent work are removed. The Mac still connects through the older pairing route.

Planned4

  • Stricter verification

    Verifier instructions and criteria that catch more before the human review.

  • Clearer consent

    The sign-in screen shows the project, your role and exactly what a client is granted.

  • Stronger Mac pairing

    A Mac proves it holds its key when it is paired, and is bound per project.

  • More projects on the phone

    The iOS app follows more than one project.