melete

# Melete The open-source, always-on agent with its own computer: any model, your machine, you approve what matters. > Melete is in early beta. You can run it yourself today, and hosted Melete is > coming soon.  Hand Melete a loose end, like a refund you were promised or a reply you're still waiting for. It writes from your own address, waits for the answer, and follows up until it's settled. You approve the first message, and it follows up within the limits you set. ## What it does - **Keeps going while you're away.** A job can wait for a reply, a date or an event, then wake up and take the next step. - **Has its own computer.** Code and commands run in a cloud sandbox on [E2B](https://e2b.dev), [Modal](https://modal.com) or [Daytona](https://www.daytona.io), and a separate browser handles web forms. - **Works with your model.** Anthropic, OpenAI, Google, Fireworks, your ChatGPT account, or a model on your own network. - **Connects to your accounts.** Gmail, Google Calendar and Outlook connect by signing in with Google or Microsoft, any mailbox connects over IMAP, and any remote MCP server adds new tools. - **Asks before it acts.** You see the exact text of every message before it sends, and each action leaves a receipt you can undo while it still works. - **Remembers what you tell it, in the open.** You can see, correct or forget anything it knows, and a correction can teach it a better way to do the job. - **Shares a space.** Your household or small team can work in one space, and each person's jobs and messages stay private to them. ## How it compares | | Melete | Closed always-on agents | | --- | --- | --- | | Source code | Open source, Apache-2.0 | Closed | | Where it runs | Your computer or your own server | The vendor's cloud | | Models | Anthropic, OpenAI, Google, Fireworks, a ChatGPT sign-in, or a local model | The vendor's models | | Where you can use it | Anywhere you can run Docker, including the EU and the UK | The countries the vendor serves | | Outgoing actions | You approve the exact text, with a receipt and Undo | Set by the vendor | | Price | Free. You pay your model provider, or run a model yourself | A paid plan | ## See it work 1. **Connect your inbox.** Gmail and iCloud take an app password, Gmail and Outlook can also connect by signing in, and most other mailboxes connect over IMAP.  2. **See what's open.** Melete reads your mail and lists what each company owes you and what is overdue.  3. **Hand it one.** Ask it to chase the refund, and it drafts an email that quotes the date the company gave you.  4. **Approve the exact email.** You see who it's from, who it's to and every word, and it goes out when you allow it.  5. **It follows up until it's settled.** When the money doesn't arrive, it writes again with the reference, and tells you when the refund is back.  ## Run it yourself You need Docker, Bun and an API key for your model provider. ```bash git clone https://github.com/ychampion/melete.git && cd melete && bun install --frozen-lockfile read -rs FIREWORKS_API_KEY && export FIREWORKS_API_KEY; bun run deploy/scripts/configure.ts; unset FIREWORKS_API_KEY docker compose -f deploy/docker-compose.yml up -d --build --wait --wait-timeout 300 ``` The second line waits for you to paste your Fireworks API key and press Enter. The key stays hidden, `configure.ts` writes it into `deploy/.env`, and the line then clears it from your shell. For another provider, export its key instead and name the provider and model, for example `bun run deploy/scripts/configure.ts --provider anthropic --model <model id>` with `ANTHROPIC_API_KEY`. Then open http://localhost:3101 and create your account. To try a demo with a practice model first, use `bun run deploy/scripts/configure.ts --fake` as the second line. It needs no key, and you can switch to your own model later. On a server, you can skip the build and pull the published images instead: [Using prebuilt images](docs/DEPLOYMENT.md#using-prebuilt-images). `bun run melete check`, `bun run melete doctor` and `bun run melete status` judge an installation and name what to do next: [The melete command](docs/DEPLOYMENT.md#the-melete-command). [Deployment](docs/DEPLOYMENT.md) covers version requirements, Windows, remote servers, HTTPS, Tailscale, backups and removal. ## Set it up with your coding agent Claude Code, Codex, Cursor or another coding agent can install Melete for you, on this computer or on a server. It checks the machine, starts Melete, and tells you when to create your account. You type your own passwords and keys, into Melete or your own terminal, and the agent does the rest. Paste this: ```text Set up Melete on this machine using https://github.com/ychampion/melete/blob/main/SETUP-WITH-AN-AGENT.md ``` The [guide](SETUP-WITH-AN-AGENT.md) also covers the optional parts: a computer for the agent, voice, mail and calendar, and a public address for other assistants. For Claude Code there is a [skill](skills/README.md) you can install once. ## Connect your model Pass the provider to `configure.ts` with `--provider` and `--model`, and export the key it reads first: | Provider | Key it reads | | --- | --- | | `fireworks` (the default) | `FIREWORKS_API_KEY` | | `anthropic` | `ANTHROPIC_API_KEY` | | `openai` | `OPENAI_API_KEY` | | `google` | `GOOGLE_API_KEY` | | `openai-compatible` | `OPENAI_COMPAT_BASE_URL` and `OPENAI_COMPAT_API_KEY`, which can point at a model server on your own network | | `chatgpt` | No key. Once Melete is running, you sign in to your ChatGPT account under **Settings → Models**, or through its [sign-in routes](docs/DEPLOYMENT.md#signing-in-to-a-provider). | You can also connect a provider from the app, in Settings › Models: paste a key, test it, and choose the model, with no restart. Write the model's name the way the provider does. To change provider or model later, edit `MELETE_DEFAULT_PROVIDER`, `MELETE_DEFAULT_MODEL` and the key in `deploy/.env`, then restart the two services that use them: ```bash docker compose -f deploy/docker-compose.yml up -d --force-recreate --wait melete runtime ``` [Providers](docs/DEPLOYMENT.md#providers) explains each one, including signing in with ChatGPT. ## Connections and plugins - **Mail and calendars.** Add them in **Settings → Connections**. Gmail and iCloud take an app password, and other IMAP and CalDAV accounts take their password. With your own Google or Microsoft app set in `deploy/.env`, Gmail, Google Calendar and Outlook connect by signing in ([Google setup](docs/mail-calendar.md#setting-up-your-google-client), [Microsoft setup](docs/mail-calendar.md#setting-up-your-microsoft-app)). - **Any MCP server.** Add one over HTTP, or from a package or image, and a packaged server runs in its own locked-down container. A starter set of files, fetch, time and GitHub is listed at `GET /plugins`, and each installs with one request. - **A sandbox.** Connect an E2B, Modal or Daytona account as a **Sandbox** connection, and the agent's terminal runs there, with a receipt for each command. - **A browser.** The [browser worker](docs/browser-worker.md) fills forms in its own isolated browser. ## Security - Every message it sends waits for your approval, or falls under a limit you set for someone you trust. Approving a chase's first message also covers up to three follow-ups that repeat it to the same person, and you can withdraw that in **Settings → Rules**. - The passwords and sign-in tokens it stores are sealed with your installation's master key. - The agent's code runs in an isolated container that can reach only Melete's own service, which checks each action against what you allowed. - Each person on an installation has their own space, and their jobs, drafts and receipts stay private to them. - Every action leaves a receipt, and one that can be reversed shows an Undo button while it still works. - To report a vulnerability, follow [SECURITY.md](SECURITY.md). The [threat model](docs/THREAT-MODEL.md) has the details. ## More it can do - Teach it by correcting it, then see, pause or remove what it learned in **Settings → Memory**. - Start your day with a short brief of your calendar, your tasks and the decisions waiting on you. - Correct or forget anything it remembers, and a forgotten fact stays gone after a restore. ## Coming soon - Hosted Melete and our website, so you can try it in your browser. ## Documentation - [Deployment](docs/DEPLOYMENT.md): hosting, providers, sandboxes, Tailscale, backups and removal - [Upgrading](docs/UPGRADING.md): moving an installation to a later release - [Connectors](docs/CONNECTORS.md) and [mail and calendars](docs/mail-calendar.md) - [Melete in other assistants](docs/MCP-SERVER.md): adding Melete to ChatGPT, Claude or Hermes as a connector - [Browser worker](docs/browser-worker.md): the browser Melete drives, and taking over from it - [Memory](docs/MEMORY.md) and [learning](docs/LEARNING.md) - [Architecture](docs/ARCHITECTURE.md) and [threat model](docs/THREAT-MODEL.md) - [Building a client](docs/CLIENT.md): the API and how the app uses it - [Problem reports](docs/FEEDBACK.md): reporting a problem from the app, and pulling one by its id to fix it - [Contributing](CONTRIBUTING.md): setting up, testing and sending changes ## Licence and credits Melete is Apache-2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE). Melete's agent runtime is built on [Hermes Agent](https://github.com/NousResearch/hermes-agent) by Nous Research, customised for Melete.