# Argos for AI agents: MCP server, CLI, skills, and REST API > Your agents use Argos end to end. Deploy a preview, read the diff, review the build, fix the flake: every pillar is reachable from the CLI, the MCP server and the REST API, with the permissions of the person running the agent. Canonical: https://argos-ci.com/ai-agents Documentation: https://argos-ci.com/docs/agents Argos does four things for a pull request (diff, review, stabilize, deploy), and an agent can take each of them with the same tools a human uses: the `@argos-ci/cli` npm package, the remote MCP server at https://mcp.argos-ci.com, and the REST API at https://api.argos-ci.com/v2. Argos itself does not use AI to decide anything: comparisons are deterministic pixel and text diffs, and agents use the resulting build data as review evidence. ## One agent, four jobs Each pillar page has a "For AI agents" section with the full set of commands. The first command an agent reaches for on each: - **Diff** (https://argos-ci.com/diff#agents): Any file, not just pixels: screenshots, Markdown, JSON, and more. The agent reads every diff as data (status, score, diff mask, baseline and current file): `argos build snapshots --json` - **Review** (https://argos-ci.com/review#agents): One place for humans and agents to approve what changed. The agent approves, rejects or comments, attributed to the person whose token it runs with: `argos review create --event approve` - **Stabilize** (https://argos-ci.com/stabilize#agents): Kill flakes and debug failures with full per-test history. The agent reads a test's flakiness and its recurring changes, then fixes the cause or ignores the change: `argos test changes --json` - **Deploy** (https://argos-ci.com/deploy#agents): Free preview URLs for your Storybook or static site on every PR. The agent ships its own preview with the command your CI runs, and the URL lands on the pull request: `argos deploy ./storybook-static` ## MCP server - Official remote MCP server: https://mcp.argos-ci.com (streamable HTTP transport, nothing to install or run locally). - Setup: `claude mcp add --transport http argos https://mcp.argos-ci.com` (Claude Code), `codex mcp add argos --url https://mcp.argos-ci.com` (Codex CLI), or add `{ "mcpServers": { "argos": { "url": "https://mcp.argos-ci.com" } } }` to your client's MCP config (Cursor; Windsurf uses `serverUrl`). Claude.ai, Claude for desktop and VS Code with Copilot add it from their settings. Guide: https://argos-ci.com/docs/agents/mcp-server#setup - Authentication: OAuth (the client opens a browser; you choose the organizations to share and the scopes to grant; revoke any time under Authorized applications) or a personal access token sent as `Authorization: Bearer `. Project tokens are not accepted by the MCP server. - Scopes: `projects:read` (projects, builds, tests), `reviews:write` (reviews, review requests, ignoring changes), `comments:write` (comments, notification subscriptions), `projects:write` (project configuration, automation rules), `media:read` and `media:write` (shared media), `account:admin` (team administration). - Tools are generated from the REST API: every operation a user can call is a tool with the same name, parameters and permissions. When the management API shipped, the server went from 37 to 71 tools with no client update (https://argos-ci.com/changelog/2026-08-10-management-api). Agents list builds and inspect diffs, approve or reject builds, request reviewers, read and post comments, investigate flaky tests and ignore their changes, configure projects and automation rules, administer a team, upload media, and read analytics and usage. - Keep your client's approval prompts on for write actions, and prefer read-only scopes when the agent only inspects builds. ## CLI and agent skills - Install the CLI: `npm i --save-dev @argos-ci/cli` (Node.js 22 or later). It runs wherever the agent has a terminal: locally, in CI, in a sandbox. - Install every skill Argos publishes: `npx skills add https://argos-ci.com`. Skills work with assistants that support them, such as Claude Code, Codex and Cursor. Activate one in a prompt: `Use $argos-pr-review to review this pull request with its Argos build.` - `argos-cli`: the CLI's commands, flags, authentication rules and output formats, flakiness commands included. - `argos-pr-review`: a complete pull-request review: find the Argos build, inspect the snapshots that need review, summarize what changed, compare it with the pull request intent, then approve or request changes. - `argos-upload`: when and how to share a screenshot or a screen recording by link, how to embed it so a pull request renders it, and how to read back the comments a human pinned on it. - Tokens: uploads, deployments and read-only commands (`build get`, `build snapshots`, `test list`, `test get`, `test changes`, `change list`, `project get`, `project deployments`, `media list`) work with a project token. Anything attributed to a user (submitting or dismissing a review, requesting reviewers, posting comments, ignoring a change, configuring a project, administering a team) needs a personal access token; `argos login` stores one locally. - Guides: https://argos-ci.com/docs/agents/agent-skills · https://argos-ci.com/docs/learn/review-workflow/review-builds-with-ai-agents · https://argos-ci.com/docs/learn/reliability-and-flakiness/fix-flaky-tests-with-ai-agents · https://argos-ci.com/docs/sdks-reference/argos-command-line-interface-cli ## REST API and discovery - REST API base URL: https://api.argos-ci.com/v2, bearer token in the `Authorization` header; a project token reads, a personal access token acts as a user. Reference: https://argos-ci.com/docs/api-reference - OpenAPI 3.1 description: https://api.argos-ci.com/v2/openapi.yaml - API catalog (RFC 9727): https://argos-ci.com/.well-known/api-catalog · MCP server card: https://argos-ci.com/.well-known/mcp/server-card.json · OAuth protected resource metadata: https://argos-ci.com/.well-known/oauth-protected-resource - How agents register and authenticate: https://argos-ci.com/auth.md - This site in Markdown: send `Accept: text/markdown` to a product page, or prefix its path with `/md` (for example https://argos-ci.com/md/review). Site map: https://argos-ci.com/llms.txt - Documentation in Markdown: append `.md` to any docs URL (for example https://argos-ci.com/docs/agents/mcp-server.md); indexes at https://argos-ci.com/docs/llms.txt and https://argos-ci.com/docs/llms-full.txt; read-only docs MCP server at https://argos-ci.com/docs/~gitbook/mcp - WebMCP: in a browser with WebMCP, every page of this site registers 12 read-only tools on `document.modelContext`: get_argos_overview, get_pricing, estimate_monthly_cost, get_page, search_docs, get_docs_page, get_quickstart, compare_with, get_security_compliance, get_latest_changelog, get_service_status, get_mcp_connection_info. ## For AI agents The commands from this page in one place: ```bash # Install the CLI and every Argos skill npm i --save-dev @argos-ci/cli npx skills add https://argos-ci.com # Connect an MCP client claude mcp add --transport http argos https://mcp.argos-ci.com codex mcp add argos --url https://mcp.argos-ci.com # One command per pillar (each pillar page has the full set) argos deploy ./storybook-static argos build snapshots --json argos review create --event approve # personal access token argos test changes --json # REST API and machine-readable site curl https://api.argos-ci.com/v2/projects/acme/app/tests//changes -H "Authorization: Bearer $ARGOS_TOKEN" curl https://api.argos-ci.com/v2/openapi.yaml curl -H "Accept: text/markdown" https://argos-ci.com/review curl https://argos-ci.com/llms.txt ``` ## Related - Diff (https://argos-ci.com/diff): Any file, not just pixels: screenshots, Markdown, JSON, and more. - Review (https://argos-ci.com/review): One place for humans and agents to approve what changed. - Stabilize (https://argos-ci.com/stabilize): Kill flakes and debug failures with full per-test history. - Deploy (https://argos-ci.com/deploy): Free preview URLs for your Storybook or static site on every PR. - Media sharing (https://argos-ci.com/media-sharing): screenshots and recordings uploaded from the CLI and posted on the pull request. ## Frequently asked questions ### Which AI clients and agents work with Argos? Any MCP client can connect to the Argos MCP server at mcp.argos-ci.com. The docs walk through Claude Code, Claude.ai and Claude for desktop, Cursor, VS Code with Copilot, Codex CLI, and Windsurf. Any agent with a terminal can use the CLI, locally or in CI, and the [agent skills](https://argos-ci.com/docs/agents/agent-skills) install into assistants that support skills, such as Claude Code, Codex, and Cursor. ### What can an agent do over MCP, and what limits it? The MCP server generates its tools from the REST API: every operation a user can call is a tool with the same name, parameters, and permissions. An agent can list builds and inspect diffs, approve or reject a build, request reviewers, read and post comments, investigate flaky tests and ignore their changes, configure projects and automations, administer a team, upload media, and read analytics and usage. What it may actually do is decided by the OAuth scopes you grant, such as projects:read, reviews:write, comments:write, or account:admin, and by your client's approval prompt on write actions. See [the available tools](https://argos-ci.com/docs/agents/mcp-server#available-tools). ### Which token does my agent need? Uploads, deployments, and read-only commands such as build get, build snapshots, test list, test get, and test changes work with a project token, the one your CI already has. Anything attributed to a user, such as submitting a review, requesting reviewers, posting a comment, ignoring a change, or configuring a project, needs a personal access token, because the action is checked against that user's permissions. Locally, argos login stores one for you. The MCP server takes OAuth or a personal access token and does not accept project tokens. See [project tokens and personal access tokens](https://argos-ci.com/docs/sdks-reference/argos-command-line-interface-cli#project-tokens-and-personal-access-tokens). ### Does Argos use AI to decide whether a change is acceptable? No. Argos compares screenshots with deterministic pixel diffing, built on the open-source odiff library, and compares non-image files as text. Same input, same result, and the exact pixels that changed are visible. Agents use that build data as review evidence; they do not replace your team's ownership of review decisions. See [how Argos detects visual differences](https://argos-ci.com/docs/learn/platform-fundamentals/how-argos-detects-visual-differences). ### Can an agent read this site and the documentation? Yes. Send Accept: text/markdown to a product page, or prefix its path with /md, and you get its Markdown twin. [llms.txt](https://argos-ci.com/llms.txt) maps the site and [auth.md](https://argos-ci.com/auth.md) explains how agents authenticate. Every documentation page exists as .md, indexed by /docs/llms.txt and /docs/llms-full.txt, and the docs expose a read-only MCP server at argos-ci.com/docs/~gitbook/mcp. In a browser with WebMCP, this site also registers read-only tools such as search_docs and get_pricing. See [AI-ready docs](https://argos-ci.com/docs/agents/ai-ready-docs).