For macOS developers who use Claude Code

A coding agent that remembers what worked on your project.

Q: What is Commonshell Pilot?

Commonshell Pilot is a local-first AI coding agent for macOS by commonshell. It runs the Claude Code CLI in session tabs and keeps learnings from earlier sessions so the next session starts further along.

Q: Does Commonshell Pilot send my code to commonshell?

No. Tasks, learnings and project files stay on your Mac under ~/.pathfinder. Anonymous usage counts never include prompts, code, or file and project names. Claude Code talks to Anthropic under your own account.

Commonshell Pilot runs Claude Code in tabs on your Mac. It records what worked, what failed and why, and gives that back to the next session on the same project. Your learnings stay on your machine.

Download for macOS (coming soon) See how it works

Pricing and the download link are announced here when they are ready.

api-server

web-checkout

docs

~/code/api-server · Claude Code · recalled 3 learnings

› worked: run migrations with APP_ENV=test make test

› failed: npm ci with a stale lockfile. Instead: npm install first.

› research reused: Stripe webhook signature check, fetched 2026-10-02

Illustration of the session tabs and recalled learnings. Not a screenshot.

Facts, not promises

Key facts

macOS 13+Apple silicon and Intel

Local-firstLearnings and files stay in ~/.pathfinder

Selenium IDE testsRun in Chrome you can watch

Zero repeat lookupsin our tests, for research already on file

The problem

Every new session starts from zero.

A coding agent doesn't know what the last session learned unless someone wrote it down. So the same mistakes come back, and the same answers get looked up again.

The test command that needs one environment variable set, found again for the third time this month.Repeated failure

The same API detail fetched from the web on Monday and again on Thursday.Repeated research

The project's conventions read from scratch at the start of every session.Repeated discovery

How it works

Three steps, then it keeps going.

You work the way you already do. Pilot handles the memory in the background.

Open a tab in your project

Press ⌘T and pick a folder. Each tab runs Claude Code there, like a terminal tab. Tabs come back after a restart.
Learnings are saved as you go

When something works, fails or is decided, it is recorded with its date and reason, so a session that closes while idle is not lost.
The next session starts informed

Learnings for the project are recalled when a session opens, within a fixed size, so the agent starts with the fix instead of the bug.

What it does

Three things it does well.

Each one is something you can check yourself in the app.

It remembers the fix, not just the bug

Each learning records the approach, the outcome and why. A failed approach can carry what to do instead.

Recalled learnings are capped at 1,800 characters, the least trusted first dropped, so they never bury the task.

Browser tests you can watch

Tests are Selenium IDE files, so they open in the Selenium IDE extension and run in Chrome. You see every step and can take over at any point.

Run one with pf test, or re-run a saved test from the Tests screen.

Research you can trust, and date

A checked web answer is reused with its fetch date and source. A newer answer supersedes the old one; nothing is silently deleted.

In our tests, a proven answer caused zero repeat lookups. We have not yet measured time saved; see the research note.

Everything else

The rest of the app.

Task plansGoals break into tasks with status and attempts. Plans are checked before they are saved.
Project files beside the chatBrowse, edit and diff the session's folder. Git changes are marked.
Phone accessPair your phone to reach your sessions over an encrypted link to your Mac. Alerts come when a session is waiting for you.
Team memoryShare chosen learnings with a team through an append-only log. Private learnings never leave your Mac.
Keyboard first⌘1–9 jump to a tab, ⌘⇧[ and ⌘⇧] step through them, ⌘, opens Settings.
Permissions you controlEach permission request an agent makes is graded and shown with its reason. Risky requests wait for your answer.

Privacy

Your code and your learnings stay on your Mac.

Tasks, learnings and project files are stored under ~/.pathfinder on your Mac.
Learnings are private unless you choose to share them with a team.
Anonymous usage counts (for example, how often the app opens) never include prompts, code, or file and project names. Settings → About lists every event, and you can switch counting off.
Your Anthropic API key is read from the environment or the macOS keychain, never written to disk by the app, and never synced.

Requirements

What you need installed.

macOS 13 or laterApple silicon or Intel

Claude CodeInstalled with Anthropic's installer, signed in with your own account. Not bundled.

Google ChromeNeeded only for browser tests

Docker (optional)For the learning graph. Without it, memory still works from the local ledger.

Settings → Setup checks all of this on your Mac.

Questions

Frequently asked.

What is Commonshell Pilot?

A local-first AI coding agent for macOS, made by commonshell. It runs the Claude Code CLI in session tabs and keeps learnings from earlier sessions, so the next session starts further along.

Does it send my code to commonshell?

No. Tasks, learnings and project files stay on your Mac. Anonymous usage counts never include prompts, code, or file and project names. Claude Code talks to Anthropic under your own account.

Is Claude Code included?

No. Install Claude Code with Anthropic's installer and sign in with your own account. Commonshell Pilot's Setup screen checks for it.

Which Macs does it run on?

macOS 13 or later, on Apple silicon or Intel. Windows and Linux are not supported yet.

How is it different from Claude Code on its own?

Claude Code does the work. Commonshell Pilot adds the memory around it: what worked, what failed, researched answers with their dates, and tests you can watch. It doesn't replace Claude Code.

Read the research before you decide.

We're measuring whether memory makes coding agents faster and more reliable. We publish the method and the results, including the ones that show no gain so far.

Read the research note Download for macOS (coming soon)