One project. Any AI.
Your project remembers, even when you switch tools.
For solo builders and teams using more than one AI, Phewsh keeps goals, decisions, current work, evidence, and handoffs in one portable record. Project-owned .intent/ files give Claude Code, Codex, Cursor, and supported tools the same starting truth.
No shared model memory · Provider-neutral · User-owned · Evidence-backed
phewsh shim on — then each supported tool you launch shows a phewsh status line: intent loaded, record current. Proof it’s working, before you type a word.Switch tools and you re-explain the goal, the decisions, and where you left off. Phewsh keeps five durable things in one record, so the next tool — or teammate — picks up without starting over.
At every handoff, Phewsh verifies the project files and lists what did not move — including the prior tool’s transcript. Your record lives in a .intent/ folder you own; syncing to Phewsh Cloud is optional. Remote MCP and provider plugins (ChatGPT, Claude connectors) are in developer preview.
Phewsh gives every project a shared intent — then keeps humans and AI agents aligned as the work moves across tools. Wherever you start, the loop completes the picture.
Ground any project in the 12 nodes of intent — purpose, constraints, risks, what “done” means. Plant the right seeds before any agent runs. No signup, no code.
Create intent →One install wires Claude Code, Codex, Cursor, and Gemini to the same project truth — your logins, your machine. No API keys required.
Install the CLI →The project thread for humans + AI agents: request in plain language, a teammate or agent claims it, evidence comes back, the record updates. Founder beta — free.
Enter Ion →Phewsh keeps recorded work visible — for you, and for supported AI tools you hand it to.
Answer four plain questions and you're aligned — whether you're thinking alone with no AI at all, or handing off to Claude, Codex, or Cursor. Supported adapters can read the same recorded answers; unrecorded conversation stays tool-local.
What you're building, and why. The goal, the constraints, the decisions already made.
What should happen next. A simple list you — or an AI tool — can pick up. Works with zero AI.
What's being done right now, and by which tool. Nothing happening in the dark.
What happened, what changed, what you learned — so the same mistake doesn't repeat.
Install once — supported AI tools start from the same project truth. See the CLI →
PHEWSH does not pretend one vendor can transfer another model’s memory. It gives the next harness the same project files and leaves an integrity-checksummed receipt that is re-checked against your checkout on pickup.
Project truth lives in versioned .intent/ files, not inside Claude, Codex, or Phewsh cloud. Change the harness; keep the project.
Start work, switch tools, hand off, or quit cleanly. The next pickup checks the receipt against the actual checkout before claiming continuity.
Open skills and native hooks adapt each tool around the same project. Remove Phewsh and the project files remain readable, useful, and yours.
No subcommands to memorize. Run it and it finds your AI tools, your projects — and the repos that don't have shared project truth yet.
curl -fsSL phewsh.com/install.sh | sh
One command. Nothing else changes until you say so.
Source on GitHub → · MIT · full CLI guide →
phewsh clarify
A guided walk through what you're building and why. Your answers become files you own — the project truth your installed adapters read.
phewsh claude
Or codex, gemini, cursor… The selected tool opens with a verified project brief — grounded before the first keystroke.
phewsh codex
Switch tools, hit a usage limit, come back tomorrow — when you hand off, the next AI gets recorded project state and an explicit loss list instead of a blank start.
Want to see it run first? Watch the 2-minute demo →
Free with your own keys, forever. Credits from $5 when you want the hosted engine — no subscription, they never expire. Start free →
See what's built on PHEWSH →