> Documentation index: https://rift.sh/llms.txt # Use Rift with your agent Rift needs an agent with a shell, a network connection and a disk that lasts as long as its work. Setup is the same for most agent products. Give the agent the [prompt](https://rift.sh/index.md#agent) with an invite token. Each product keeps a standing instruction in a different place, and each sandbox allows different things. An agent that runs as you on a machine where you already use Rift works with your key and sees every folder in `~/rift`. To give it a key of its own, make an invite and tell the agent to join with `rift init --invite --home ` and to add `--home ` to every `rift` command. `` is a directory you choose for that agent. If you paste the prompt on such a machine and name no directory, the agent stops, tells you it would be working with your key, and waits for you to choose. The standing instruction is short. ``` Keep your work inside a folder that rift ls lists. A folder it lists as work/ is at ~/rift/work. To reach a folder you cannot see, run rift request --reason "". ``` For an agent that joined with `--home `, the same line carries the directory. ``` Keep your work inside a folder that rift ls --home lists, under the folder that rift status --home prints. To reach a folder you cannot see, run rift request --reason "" --home . ``` ## Claude Code Paste the prompt into a session. Put the standing instruction in `CLAUDE.md` in the project, or in `~/.claude/CLAUDE.md` to apply it to every project on that machine. `~/rift` is outside the project, so start Claude Code with `--add-dir ~/rift` or approve the access when it asks. Written from [memory files](https://code.claude.com/docs/en/memory), read in October 2026. ## Codex These steps are for Codex on your own machine. Codex in the cloud is not covered. Codex runs commands in a sandbox whose default mode has no network access and writes only inside the project and temporary folders. Install Rift and run `rift init` yourself, outside Codex, because the installer needs the network and writes outside the project. Codex then works with your key. Then allow network access and add `~/rift` as a writable folder, and put the standing instruction in `AGENTS.md` in the project. Written from [sandboxing](https://learn.chatgpt.com/docs/sandboxing) , [approvals and security](https://learn.chatgpt.com/docs/agent-approvals-security) , [AGENTS.md](https://learn.chatgpt.com/docs/agent-configuration/agents-md), read in October 2026. ## OpenClaw OpenClaw runs on your own machine and runs commands on that machine by default, so the prompt needs no changes. Put the standing instruction in `~/.openclaw/workspace/AGENTS.md`, which it loads at the start of every session. Written from [running commands](https://docs.openclaw.ai/tools/exec) , [the workspace](https://docs.openclaw.ai/concepts/agent-workspace), read in October 2026. ## Grok Bot Your account has one cloud computer with a command line, and every Bot on it uses that computer. The computer hibernates when idle. xAI treats packages you install by hand as replaceable, so Rift may need installing again after an update. xAI advises against pasting passwords or one-time codes into chat. Its pages describe taking control of the computer for a sensitive step and do not say whether you can type commands there. If you can, run the install and the join from the prompt yourself. If you cannot, paste the prompt with the token, and make that invite under a name you use for nothing else. Put the standing instruction in the Bot's description, under Edit Profile, which is where xAI says rules that should stay true belong. Because your Bots share one computer, they share one Rift key and one `~/rift` folder. Share a folder with that key only if every Bot on the account may read it. A Team Bot that answers in Slack channels runs on a different computer and needs its own invite. Written from [the computer](https://docs.x.ai/grok-bot/computer-and-apps) , [security](https://docs.x.ai/grok-bot/security) , [Bots](https://docs.x.ai/grok-bot/bots) , [Team Bots](https://docs.x.ai/grok-bot/team-bots), read in October 2026. ## Buzz Buzz is a workspace of channels that people and agents share. An agent in it is Claude Code, Codex, goose or another agent that speaks the Agent Client Protocol, running on the machine where you started it. Set Rift up on that machine, or send the prompt to the agent as a Buzz message. The standing instruction goes in that agent's own instruction file. Agents that run as the same user on one machine share one Rift key and one folder. Written from [Block's repository](https://github.com/block/buzz) , [Block's announcement](https://engineering.block.xyz/blog/buzz), read in October 2026. ## dots A dot can work on a computer of yours that you connect to it. Install Rift on that computer first, connect it from the dot's profile, and tell the dot to keep its work in a folder under `~/rift` there. The computer has to stay online with the ChatGPT app open while the dot uses it, and a dot connects to one personal computer at a time. In a workspace, an owner has to allow local computer access. A dot also has a cloud computer of its own. In a workspace, an owner has to allow cloud computer use and cloud network access. OpenAI's pages do not say whether Rift's installer runs there. They say only that the computer can keep its state between uses. OpenAI's pages say to give a preference like this one in conversation. Custom rules, in settings, only control when a dot may take an action. So tell the dot the instruction, ask it to save it to its notes, and ask it to repeat what it saved. Written from [dots](https://learn.chatgpt.com/docs/dots) , [computers and apps](https://learn.chatgpt.com/docs/dots/computers-and-apps) , [controls](https://learn.chatgpt.com/docs/dots/controls) , [admin guide](https://learn.chatgpt.com/docs/enterprise/dots-admin-guide), read in October 2026. ## Claude Tag Claude Tag runs each Slack thread in a new sandbox that is released a few minutes after the reply. Its network goes through a proxy that carries HTTP and HTTPS only, to hosts on an allowlist. Rift's traffic between machines is not HTTP, so it cannot cross that proxy, and a key made in the sandbox is gone a few minutes after the reply. There are two ways to use them together. - Run Claude Tag's sessions on a self-hosted environment, with Rift installed in the runner's image and the runner's own network. Keep Rift's key on a volume that outlasts the runner, because a runner can restart with a fresh disk. Claude Tag cannot use Access bundles in self-hosted sessions yet. - Keep Rift on a second agent on a machine of yours, and have it pick up what Claude Tag posts or pushes. Written from [how it works](https://claude.com/docs/claude-tag/concepts/how-it-works) , [agent identity](https://claude.com/docs/claude-tag/concepts/agent-identity) , [self-hosted environments](https://code.claude.com/docs/en/self-hosted-environments), read in October 2026. ## With other tools Guides for using Rift beside a sandbox, a memory service, a protocol or another tool. - [E2B](https://rift.sh/docs/use-with/e2b.md) - [Daytona](https://rift.sh/docs/use-with/daytona.md) - [Modal](https://rift.sh/docs/use-with/modal.md) - [Sprites](https://rift.sh/docs/use-with/sprites.md) - [Vercel Sandbox](https://rift.sh/docs/use-with/vercel-sandbox.md) - [Mem0](https://rift.sh/docs/use-with/mem0.md) - [Supermemory](https://rift.sh/docs/use-with/supermemory.md) - [Zep](https://rift.sh/docs/use-with/zep.md) - [Rivet Actors](https://rift.sh/docs/use-with/rivet-actors.md) - [MCP](https://rift.sh/docs/use-with/mcp.md) - [A2A](https://rift.sh/docs/use-with/a2a.md) - [ANP](https://rift.sh/docs/use-with/anp.md) - [SLIM](https://rift.sh/docs/use-with/slim.md) - [Tailscale](https://rift.sh/docs/use-with/tailscale.md) - [GitHub](https://rift.sh/docs/use-with/github.md)