Phewsh coordinates the AI tools you already use around one project-owned record — plain .intent/ files plus a five-tool MCP contract at mcp.phewsh.com. Every door below reaches the same record. Unrecorded conversation nuance still does not transfer.
Labels are literal. Live works today · Custom setup works today with manual configuration. Nothing here is “coming soon.”
The fastest way to understand Phewsh is to watch the record itself.
In your browser. Free account.
Open phewsh.com/ion and sign in. You’ll see real projects with a revision number, decisions and handoffs attributed to the tool that made them (Claude Code, OpenAI, the CLI…), and the work waiting on a human. Ion reads the real ledger only — it never invents sample activity.
The record itself is plain files: every project’s truth lives in a .intent/ directory in its own repository. Remove Phewsh and the record stays readable.
For anyone in a terminal with Claude Code, Codex, Cursor, Gemini CLI, and friends. Open source (MIT).
phewsh claude, phewsh codex, phewsh cursor… launch your installed tools with the same project brief. phewsh seq writes native context files (CLAUDE.md, AGENTS.md, …) from one verified core. Installed harnesses are detected automatically.
When do I need phewsh serve? Only when the browser workspace should see this machine: it runs a loopback-only worker so Ion’s “Check local node” can report your registered projects and installed harnesses. Everything else works without it. Work begins only after a human explicitly claims the task — the worker never grants remote execution.
Four doors into the same ledger. Pick the ones you actually use — each is independent, and each is removable.
The Phewsh plugin, inside Claude Code. Directory listing pending
Installs from our public marketplace today; the Anthropic Plugin Directory listing is submitted and under review.
You get five skills (resume-project, plan-with-context, finish-session, create-handoff, reconcile) and the five-tool MCP connection. The key is stored in your OS keychain.
For chatting with Claude outside the terminal. Plan must support custom connectors
A different door than the plugin: the plugin serves Claude Code, this connector serves claude.ai and Claude Desktop.
https://mcp.phewsh.comCustom connectors are a Claude plan feature, not available on every tier. Phewsh is not yet in Anthropic’s curated Connectors Directory; that listing has separate requirements. If your connection is refused, email hello@phewsh.com — client origins are allowlisted deliberately, never wildcarded.
Developer connector. Public app listing pending
The Phewsh app is submitted to OpenAI and under review. Until it’s approved, connect it yourself as a developer connector — the same server the listing will use.
https://mcp.phewsh.com — OAuth is auto-discovered; sign in and approve on the Phewsh consent page.ChatGPT then reads your bounded project snapshot and can record proposed decisions and handoffs into the same ledger — attributed, revision-guarded, visible in Ion within seconds.
Two options for OpenAI’s terminal agent.
Simplest: run phewsh codex from the CLI above — Codex launches with the project brief, no MCP config needed.
Direct MCP: add https://mcp.phewsh.com as an MCP server in ~/.codex/config.toml with an Authorization: Bearer <your MCP key> header. Key from phewsh.com/api. Codex then speaks the same five tools.
One record, many doors. The project truth is yours — plain files in your repo, one ledger in your account — and every AI tool starts from the same page.