Happy vs Pi Durable#

Happy and Pi Durable both keep an agent's work through a crash or a restart. They solve it for different people. Pi Durable is a TypeScript library for building your own agent app, on your server, on Cloudflare, or on Vercel. Happy is an app you use today: a desktop app and phone apps for coding agents, running Claude, Codex, Grok, Gemini, and more on the subscriptions you already pay for.

Side by side#

Facts about Pi Durable come from Earendil's launch post, its npm package, and the Cloudflare and Vercel docs, checked on November 4, 2026. It is experimental, and its API can change in minor releases.

Both save each step before moving on, let one process own the storage at a time, run many conversations at once, and let a client steer a running turn or queue the next message.

Where Pi Durable is better#

• It runs anywhere JavaScript does. Its SQLite and JSONL storage use no Node APIs, so the same harness runs in Node, Bun, or a Cloudflare Durable Object. Happy Agent runs on a machine you keep on.
• Bring your own storage. The storage interface is small, and it ships a conformance suite to test your backend against. Happy Agent stores everything in its own SQLite database.
• Fork a conversation. A conversation can branch from any point in another's transcript without copying it. Happy can't fork a conversation yet.
• Small enough for an agent to read. Earendil, October 1, 2026: "The entire source code, without tests, is about 15,000 lines."
• Build any agent. Coding agents are one use. Earendil's examples are Slack bots and threads, each thread a fork of the channel.

Where Happy is better#

• Nothing to build. Install Happy Desktop and start a session. Pi Durable gives you the harness; the app, the clients, and the auth are yours to write.
• Your subscriptions. Happy Agent uses the Claude Code, Codex, and Grok sign-ins already on your machine.
• A desktop and a phone app. Both watch and steer the same sessions. The phone goes through Happy's relay, end-to-end encrypted.
• A sandbox and review. Commands run in an OS sandbox, and in Auto a separate agent reviews anything that would cross it. See Permissions & Sandbox.
• A team server. In team mode each member signs in as themselves, through WorkOS or your own identity provider, and works in the same projects. See Multiplayer & Teams.
• Updates mid-turn. Install an update while an agent is working and the same turn continues on the new version. See Updates Mid-Turn.

What survives a crash in Happy#

Happy Agent is a daemon on your machine, and the apps are only clients. Close the window or the phone app and your agents keep working.

Happy Agent writes every message, tool call, and result to its SQLite database as it goes, inside a transaction, and only then moves on. After a crash it reopens the database and picks up each agent that still owed work. Idle agents stay on disk until something asks for them.

• Messages. A client names each message with its own id. Sending the same id again returns the message that already exists, so retrying a send whose response was lost is safe.
• Tools. A tool that isn't safe to repeat runs at most once. If a crash cuts it off, it is answered as interrupted and never run again.
• Answers. Only finished blocks are stored. A half-written answer is dropped, and the model is asked to continue from where it left off.
• Subagents. They are sessions too, saved the same way. One that finished can take a follow-up with its full context.
• Waits and scheduled messages. An agent waiting until a time, or a message scheduled for later, survives the restart.
• Workspaces. A workspace is a folder on disk, or a Git worktree for parallel work, so the agent's edits are where it left them.

On macOS and Linux, Happy Agent restarts itself after a crash; on Windows it doesn't yet. On a server it runs as a systemd or launchd service.

We checked the agent loop with machine-checked proofs in Lean: accepted messages are never lost or duplicated, an unsafe tool never reruns after a crash, and a restarted agent can finish its work and settle. The proofs are over a model of the loop, not the TypeScript itself.

Pi Durable on Cloudflare and Vercel#

Cloudflare. The Agents SDK's PiHarness stores Pi's transcripts and tasks in a Durable Object's SQLite database. If the object is evicted mid-run, an alarm restarts it and the run continues. You choose how clients connect: WebSockets, HTTP, or RPC. Each Agents SDK release supports one Pi minor version, and Cloudflare warns that after moving to Pi 1.5, rolling back leaves conversations unreadable, so "roll forward to fix a bad deploy instead."

Vercel. durable-harness runs Pi, OpenCode, fx, Claude Code, and Codex as durable agents on the Workflow SDK. Each session is a workflow run, so a crash resumes the turn from its event log on another machine. Claude Code and Codex run only in the session's own Vercel Sandbox. It is on version 0.0.x, Apache-2.0, and its README says the API may change between releases. To self-host it, use its Postgres backend.

Which one#

To start with Happy, see the Quick Start or download it.