--- title: "Agent setup" description: "MCP, tools, skills, client configuration, and the first task." group: "For agents" order: 0 --- ## MCP - Server: `https://ctx4.me/mcp`. - Transport: Streamable HTTP. - Human sign-in: browser OAuth, without issuing a Context4 API key. - [Resource discovery](https://ctx4.me/.well-known/oauth-protected-resource/mcp), [OAuth discovery](https://ctx4.me/.well-known/oauth-authorization-server). - [Console and accounts](https://console.ctx4.me/integrations). Start with this page. Open other material for the step you are working on. Documentation and an MCP configuration entry do not prove access to data. ## 1. Identify the client Identify the application from reliable environment context, not the model. If unknown, ask once: ChatGPT, Claude, Codex, Claude Code, Hermes, OpenClaw, or another MCP client? Cloud clients cannot reach localhost documentation or a local MCP server. Stop and explain the need for a reachable HTTPS deployment. Do not invent a domain or silently substitute the production server. ## 2. Skills and client configuration Check existing Context4 skills and tools. Reuse a suitable MCP connection. Do not overwrite another configuration or edited files without consent. For normal operation, install `context4` and `context4-setup`; add `context4-workflows` for tasks across systems. [File manifest and SHA-256](https://ctx4.me/docs/skills-index.json), [package descriptions](https://ctx4.me/docs/skills.md). Read the whole package before saving it, including references. If the client cannot persist skills, use this guidance in the current chat and state that there is no persistent installation. Open only the pair of links for your client: | Client | Connection | Skills | | --- | --- | --- | | ChatGPT | [Plugin](https://ctx4.me/docs/connect-chatgpt.md) | [Installation](https://ctx4.me/docs/skills-chatgpt.md) | | Claude web/desktop | [Remote connector](https://ctx4.me/docs/connect-claude.md) | [ZIP upload](https://ctx4.me/docs/skills-claude.md) | | Codex | [Direct MCP](https://ctx4.me/docs/connect-codex.md) | [Files](https://ctx4.me/docs/skills-codex.md) | | Claude Code | [HTTP MCP](https://ctx4.me/docs/connect-claude-code.md) | [Files](https://ctx4.me/docs/skills-claude-code.md) | | Hermes | [MCP/OAuth](https://ctx4.me/docs/connect-hermes.md) | [Installation](https://ctx4.me/docs/skills-hermes.md) | | OpenClaw | [MCP/OAuth](https://ctx4.me/docs/connect-openclaw.md) | [Installation](https://ctx4.me/docs/skills-openclaw.md) | If the chat cannot change client settings, show exact manual steps. The user confirms sign-in, workspace, and consent. Wait for MCP tools to appear; explain a restart or new chat if needed. Installing a skill does not create an OAuth grant. ## Other MCP clients For an unlisted client, use its native remote Streamable HTTP MCP configuration with the address above and OAuth discovery. Check skill storage against the actual client's documentation; a standard package contains SKILL.md and references. Do not guess configuration paths or CLI flags. If persistent skills are unsupported, use the guidance for this conversation and state the limitation explicitly. ## 3. Core tools The global `/mcp` endpoint provides five meta-tools. Use the actual client's schemas; see the [full reference](https://ctx4.me/docs/mcp-tools.md). | Tool | Purpose | | --- | --- | | `CONTEXT4_SEARCH_TOOLS` | Find an action for a service and task | | `CONTEXT4_GET_TOOL_SCHEMAS` | Get exact arguments for discovered actions | | `CONTEXT4_MANAGE_CONNECTIONS` | Inspect accounts or obtain a Connect Link | | `CONTEXT4_WAIT_FOR_CONNECTIONS` | Wait for an account to become ready | | `CONTEXT4_MULTI_EXECUTE_TOOL` | Execute the task with verified arguments | Do not guess tool slugs or use another product's tools. COMPOSIO aliases are permitted only in a session that actually lists them. ## 4. Connect a business account If the task or service is unknown, ask. Discover a suitable read, fetch its schema, and inspect connections. Ask the user to choose when multiple accounts are available. If authorization is needed, obtain a link through `CONTEXT4_MANAGE_CONNECTIONS`, show it to the user, and wait for actual readiness through `CONTEXT4_WAIT_FOR_CONNECTIONS`, respecting retry hints. Do not construct an authorization URL manually or request secrets in chat. ## 5. Confirm readiness with a read Perform one permitted read through `CONTEXT4_MULTI_EXECUTE_TOOL`. Inspect every result. A successful empty response means the query works but has no matching data. Pending status, an error, a discovered schema, or a displayed link does not yet prove readiness. If no read is available, explain the limitation; do not create a record as a test. Briefly report the connected service/account, what was checked, and a useful next task. Writes, deletions, and messages must stay within the user's instruction. ## As needed [Providers](https://ctx4.me/docs/agent-providers.md), [errors and MCP](https://ctx4.me/docs/ai-mcp.md), [connections](https://ctx4.me/docs/connections.md), [revoking access](https://ctx4.me/docs/access-and-disconnect.md), [all chapters](https://ctx4.me/docs.md). For server integrations: [API](https://ctx4.me/docs/api-reference.md), [SDK/proxy](https://ctx4.me/docs/composio-sdk.md). Do not load the full API reference for ordinary MCP setup. For a person: [open the HTML instructions](https://ctx4.me/docs/agent-setup). ## Data and terms Context4 is free. Operation content is proxied without persistent storage. Account and authorization data is processed separately. [Data handling](https://ctx4.me/docs/data-handling.md) · [Privacy](https://ctx4.me/privacy) · [Terms](https://ctx4.me/terms) · [Support and deletion](https://ctx4.me/support#delete).