--- title: Instructions for coding agents description: A short Kaviris section for AGENTS.md, so any coding agent knows how to deploy your project. --- # Instructions for coding agents First connect your agent, once, by following its quickstart: [Claude Code](quickstart/claude-code.md), [Cursor](quickstart/cursor.md), or the [command line](quickstart/cli.md) for any other agent that can run commands. You, the person, do the sign-in. An agent cannot do it for you. For Claude Code there is also a [deploy skill](https://docs.beta.kaviris.com/skills/deploy/SKILL.md): the whole loop of deploying, reading refusals, fixing and rolling back ([how to install it](quickstart/claude-code.md#give-it-the-deploy-skill-optional)). ## Tell the agent about Kaviris in your project Most coding agents read a file of instructions in your project: `AGENTS.md` for nearly all of them, `CLAUDE.md` for Claude Code, `GEMINI.md` for Gemini CLI. Add the section below to it, and your agent knows how to deploy to Kaviris: over MCP when it can sign in, otherwise with the [kaviris CLI](quickstart/cli.md). In your project folder: ```sh curl -fsSL https://app.beta.kaviris.com/cli/agents.md >> AGENTS.md ``` Use `CLAUDE.md` or `GEMINI.md` instead of `AGENTS.md` for those agents. Or copy it from here: ```markdown ## Deploying with Kaviris This project is hosted on Kaviris (https://docs.beta.kaviris.com/llms.txt). - If you have the Kaviris MCP server (https://api.beta.kaviris.com/mcp) and its sign-in works, use its tools. Otherwise use the `kaviris` CLI. - Install the CLI if `kaviris version` fails: `curl -fsSL https://app.beta.kaviris.com/cli/install.sh | sh` - Sign in: `kaviris login --json`, or `kaviris login --device --json` when there is no browser here (a cloud VM, SSH). Show the user the `url` and `user_code` it prints and wait; they sign in with their passkey. - Deploy: commit, then `kaviris deploy --json`. Exit code 2 means refused or waiting for approval: each refusal has a `fix`; apply it, then deploy again. - Secrets: never put values in code or in the chat. `kaviris secrets set NAME --declare` reads $NAME or .env. - The app listens on $PORT on 0.0.0.0. - `apps delete` and `rollback` print a summary and a `confirm_token`: show the summary and repeat with `--confirm ` only if the user agrees. - `pending_approval` (with an `approve_url`) means a person must approve in the portal with their passkey: deleting or restoring, linking GitHub, support access, members, policies, and contract changes the org marks. Give the user the link; never try to approve it yourself. ``` Tools that call HTTP APIs rather than run commands (custom GPT actions, Copilot Studio) import the OpenAPI description of the REST API: https://api.beta.kaviris.com/v1/openapi.json.