Connect your agent to AIsa

Add the AIsa MCP server to Claude Code, Cursor, Codex, or other agents, sign in with browser OAuth, and start using tools. Use the CLI only when MCP cannot run.

AI agent reading this? Go to Agent instructions and run the setup yourself.

AIsa provides a remote MCP server at https://mcp.aisa.one/mcp. Authentication uses browser OAuth, so there is no API key to copy.

1. Add the server

   Claude Code

   claude mcp add --transport http -s user aisa https://mcp.aisa.one/mcp
   

   For Claude Desktop or ChatGPT: add a custom connector with https://mcp.aisa.one/mcp in the app’s connector settings.

2. Sign in

   Your MCP client opens a browser window; click Allow to complete OAuth. In Claude Code, run /mcp, select aisa, and authenticate. In Cursor, click Needs login in the MCP panel.

3. Ask your agent

   Paste this into your agent chat:

   Use AIsa to check my account, then find three recent public updates about a company of your choice with source links. Ask me before any paid call.
   

Prefer to let the agent do the setup too? Paste Set up AIsa using https://aisa.one/docs/agent-quickstart.md. and it follows the instructions below.

Agent instructions

Set up AIsa. Run the matching commands yourself. Do not ask the user to run them — except when a sandbox or a denied write blocks you: then relay the exact steps and wait.

Reuse a working connection: an authorized AIsa MCP server at https://mcp.aisa.one/mcp with the tools search, list_categories, get_details, use, and batch_use, or an official aisa skill plus valid CLI auth. Explicit choice or a still-authorized connection wins; do not migrate or reinstall automatically. Do not change unrelated model settings. Describe installs and credentials as in this agent environment.

Default to native remote MCP at https://mcp.aisa.one/mcp (Streamable HTTP, OAuth) whenever this client can add that connector — user-local or cloud, with or without a terminal. A working shell is not a reason to choose the CLI.

Use the CLI only when this client has no MCP mechanism at all, for non-interactive automation (CI, cron, scripts) that cannot complete OAuth and already holds a key, or by explicit user choice. These gates and the Node rule in Use the CLI take precedence over any installed aisa skill; the skill covers follow-up operations only.

MCP (default)

Connect https://mcp.aisa.one/mcp (Streamable HTTP, OAuth). No CLI, npm, npx, or Skill install is required; already-installed guidance may be used. The client owns browser sign-in and tokens. Use this client’s own MCP mechanism, not a hand-edited config file of another tool:

• Claude Code: claude mcp add --transport http -s user aisa https://mcp.aisa.one/mcp
• Codex: codex mcp add aisa --url https://mcp.aisa.one/mcp (Codex starts OAuth on its own)
• Cursor / VS Code: a url entry for https://mcp.aisa.one/mcp in the MCP config

If AISA_API_KEY is already set in this environment and the user did not ask for OAuth, the client may send Authorization: Bearer $AISA_API_KEY instead of OAuth (Claude Code: --header; Codex: --bearer-token-env-var AISA_API_KEY). Never print the key.

Add the connector, then let the user finish OAuth in the browser (skip this when the client sends the existing AISA_API_KEY). If this client supports MCP but you cannot add the connector from here — unknown capability, sandbox, or permission denied — do not switch to the CLI: ask the user to enable or add the connector, then wait. If the client must reload or start a new session to load the connector, tell the user; verify the five tools afterwards, and do not switch to the CLI meanwhile.

Grok Bot + Plugins: Settings → Plugins → Add → auth if asked → @ attach. If a native custom MCP setup exists, enter that URL and complete OAuth. If neither is available, ask the user to enable or add the connector. grok.com (grok.com/connectors) is a different product.

Then work with those five tools: search (describe the task) → get_details when you need the full contract or a quote → use with the operation_id and arguments, or batch_use for up to 20 calls. search, list_categories, and get_details are free; use and batch_use are billed per call. Take operation IDs and arguments from search or get_details; do not invent IDs or prices. A quote does not execute and is not approval. A missing, failed, or partial quote is not free, not a batch total, and not a cap; neither is an estimated cost. Pass the approved amount as max_price_usd: the server rejects a call whose estimate is higher before any spend. It checks the estimate only and does not cap the final charge, so for metered operations approve only when the endpoint’s documented maximum, or a bound you can prove from its limits, fits the remaining approved budget. If no such bound exists, tell the user and do not execute under a fixed budget. Re-quote if the operation, arguments, or scope change. Stop if the quote or documented maximum exceeds the remaining approved budget.

CLI (only when MCP cannot run here)

Follow Use the CLI: its gate, Node.js rule, skill install, and sign-in. Do not work around a missing prerequisite or a denied write.

Finish

AIsa is connected. only after valid authentication (OAuth, or the existing AISA_API_KEY sent as a Bearer token) plus a successful free protected call — MCP use with operation_id account, which returns the key, balance, and usage — or aisa balance (CLI). Tool metadata or a 401 is not a protected call. A missing connector is not connected. Then continue a named task, or ask What would you like to do first? CLI follows the aisa skill; MCP uses those tools. Setup is not spend approval.

Do not jump to HTTP to bypass missing credentials, exhausted credit, or the same API outage. Do not make paid calls to test availability. Do not extract client-owned MCP OAuth tokens.

Only if CLI and MCP are both unavailable, HTTP works, and a valid AIsa key already exists: read the direct HTTP section of Use the CLI (endpoint docs, pricing, approval XML) before a paid proposal, cost approval, or paid request. Extra models, REST, keys, or skills: AIsa docs.