# Let AI clients call AskTable through MCP Expose AskTable's query and management capabilities to AI clients through MCP. 使用指南 · 模块: integrations · 应用开发者 · 部署形态：资料未说明 · 已确认产品版本：资料未说明 · 来源：现有文档 AskTable's MCP Server exposes data questions and management capabilities as tools through Model Context Protocol, allowing Claude, MCP Inspector, and other MCP-capable AI clients to call them directly. It is a single server named `asktable`, with 43 tools. MCP does not bypass AskTable permissions: the API-Key type and project permissions still determine what the client can call. ## Before you begin - An AskTable project and project API-Key. Use `asker` for questions only; use `admin` for management tools. - Creation location: left main navigation → “Settings” → “API-Key” → “Create API-KEY”; select the type in the dialog. - The client supports Streamable HTTP or stdio. Use HTTP in production. - Confirm the URL first; the three deployment types differ, as shown below. - UI entry: bottom avatar → “Documentation” → “MCP service,” which opens the URL for the current deployment. ## Steps **Confirm your MCP URL** | Deployment | URL | | ------------------------------------------------------------------------- | ----------------------------------- | | SaaS (cloud.asktable.com) | `https://mcp.asktable.com/` | | Private deployment, all-in-one image (in-container nginx proxies `/mcp/`) | `https:///mcp/` | | Private deployment, standalone `asktable-mcp` container | `http://:8690/` | The standalone container's HTTP transport accepts any path without requiring the `/mcp/` prefix. The all-in-one image exposes only `/mcp/` externally; nginx strips the prefix before forwarding to 8690. **Enter the URL and Key in the client** Streamable HTTP configuration (field names depend on your client): ```json { "mcpServers": { "asktable": { "type": "http", "url": "https://mcp.asktable.com/", "headers": { "Authorization": "Bearer YOUR_API_KEY" } } } } ``` The `Authorization` header is required. The server creates an independent client instance for each request and does not retain state across tenants. **Confirm that the client can list tools** After connecting, ask the client to list tools; you should see 43. They are grouped by domain below: | Domain | Tools | Count | | ------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | ----- | | datasource | `datasource_list`, `datasource_get`, `datasource_create`, `datasource_update`, `datasource_delete`, `datasource_upload_file`, `datasource_test_connection` | 7 | | meta | `datasource_meta_get`, `datasource_meta_export`, `meta_sync`, `meta_import` | 4 | | training | `training_pair_list`, `training_pair_create`, `training_pair_update`, `training_pair_delete` | 4 | | field | `field_set_visibility`, `field_set_description`, `field_set_ai_value_index` | 3 | | table | `table_set_description` | 1 | | role | `role_list`, `role_get`, `role_create`, `role_update`, `role_delete` | 5 | | policy | `policy_list`, `policy_get`, `policy_create`, `policy_update`, `policy_delete` | 5 | | chat | `chat_ask`, `chat_followup` | 2 | | conversation | `conversation_list`, `conversation_get`, `conversation_delete` | 3 | | data\_agent | `data_agent_list`, `data_agent_get`, `data_agent_create`, `data_agent_update`, `data_agent_delete` | 5 | | project | `project_get`, `project_preference_get`, `project_update`, `project_preference_update` | 4 | List tools accept `page` and `size` (`size` is capped at 200); omitted values use backend defaults. **Ask with chat\_ask, then follow up with chat\_followup** Have the client call `chat_ask` to ask a question. For multiple turns, use `chat_followup` to continue the previous conversation. If the client provides `progressToken`, these tools send progress notifications between each poll. Delete tools (`*_delete`) require `confirm` to be `true`; omitting it is treated as a parameter error. **Use stdio for local debugging** For standalone debugging or MCP Inspector, start stdio in the `mcp/` directory: ```bash bun src/main.ts --stdio ``` In stdio mode, the Key comes from environment variables: ```bash export ASKTABLE_API_URL='https://' export ASKTABLE_API_KEY='YOUR_API_KEY' ``` The server uses `ASKTABLE_API_URL` to target AskTable's API, defaulting to `https://api.asktable.com`. In the all-in-one image it is set to `http://127.0.0.1:8688`, using localhost within the same container. **Secure credentials before production** Client configuration files are often synchronized or backed up. Use the client's secret management to inject `Authorization` in production rather than leaving the Key in plaintext in a configuration repository. If a Key leaks, immediately delete and recreate it in “Settings” → “API-Key.” ## Key constraints | Item | Value or default | Where to change it | | ----------------------------------------- | ------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | | Server name and tool count | Single `asktable` server, 43 tools | No UI switch | | Cloud URL | `https://mcp.asktable.com/` | Bottom avatar → “Documentation” → “MCP service” | | Private deployment (all-in-one image) | `/mcp/` | See [Private deployment (Docker)](https://docs.asktable.com/en/deploy/docker) | | Private deployment (standalone container) | `http://:8690/` | Container `ports` and `ASKTABLE_API_URL` | | Authentication | `Authorization: Bearer `, required | Client configuration | | `ASKTABLE_API_URL` | Default `https://api.asktable.com` | MCP container environment variables | | List pagination `size` | Maximum 200; omission uses backend defaults | Tool arguments | | Delete tools | Must explicitly pass `confirm: true` | Tool arguments | | Audit logs | One JSON line per call to stderr; only the first 16 characters of the Key's sha256 are kept | Container logs | | Production transport | Streamable HTTP; stdio is for local debugging only | Startup options | ## Troubleshooting | Symptom | Identifying condition | Resolution | | -------------------------------------------------------- | ---------------------------------------------------- | ------------------------------------------------------------------------------------------------- | | Client reports `InvalidRequest` | Server maps HTTP 401 or 404 to this error | Check that `Authorization` has the `Bearer ` prefix, the Key is valid, and the URL is correct | | Client reports `InvalidParams` | Server maps HTTP 422 to this error | Correct parameter types or required fields according to the tool schema | | Client reports `InternalError` | Upstream returned another status code | Check `httpStatus` preserved in the error object, then correlate by time in AskTable service logs | | Calling `*_delete` reports a parameter error | Arguments lack `confirm: true` | Explicitly pass `confirm: true` | | Standalone container is unreachable | `http://:8690/` cannot be reached | Confirm the container port is published and `ASKTABLE_API_URL` points to an accessible backend | | `/mcp/` returns 404 in the all-in-one image | Port 8690 is not running in the container | Check the in-container `mcp-server` process and nginx proxy configuration | | `meta_sync` immediately returns `{"status":"triggered"}` | `wait` was omitted | Expected behavior; to wait for results, pass `wait`, `poll_interval_ms`, and `max_attempts` | | List tools return fewer entries than expected | `size` was omitted, using backend default pagination | Explicitly pass `size` (up to 200) and `page` | | Client never receives progress | Client did not provide `progressToken` | Progress notifications are sent only when the client provides `progressToken` | ## Next steps - Operate AskTable directly in a terminal: [Command-line tool](https://docs.asktable.com/en/integrations/cli) - Handle requests and responses yourself without an AI client: [OpenAPI](https://docs.asktable.com/en/integrations/openapi) - Run a separate MCP container in a private deployment: [Private deployment (Docker)](https://docs.asktable.com/en/deploy/docker)