Skip to content
R Rowset Logo Rowset
Sign in Create a backend
Docs / Configure agent accessCreate a backend →

Configure agent access

Rowset gives signed-in users a short copy/paste setup prompt for trusted AI agents. It includes the current instance's MCP URL, REST API base URL, SKILL.md instructions URL, the repo skill install command, and an API key for bearer-token auth. On a self-hosted deployment, these URLs are generated from that instance's configured SITE_URL.

The dashboard preview masks the API key. The copy button includes the real key, so treat the copied prompt like a password.

Copy/paste setup prompt

The docs show a masked example:

Set up Rowset for this user.

Rowset MCP URL: https://rowset.lvtd.dev/mcp/
Rowset REST API base: https://rowset.lvtd.dev/api/
Rowset API key: ***
Rowset skill: https://rowset.lvtd.dev/SKILL.md
Rowset skill install: npx skills add LVTD-LLC/rowset

Read the skill URL or install the repo skill before acting. Store the full Rowset API key in a private environment variable named ROWSET_API_KEY; do not print it in logs, screenshots, public chats, generated files, or final responses. Do not commit it, paste it back to chat, or save it in a tracked config file. Configure Rowset as a remote Streamable HTTP MCP server named rowset, using the Rowset MCP URL above and bearer-token env var ROWSET_API_KEY so requests send Authorization: Bearer <key>. For Codex/OpenClaw-compatible clients, the setup command is: codex mcp add rowset --url <Rowset MCP URL> --bearer-token-env-var ROWSET_API_KEY. If the client only supports custom headers, set Authorization to Bearer <key>; use X-API-Key only for REST clients that cannot send bearer tokens. After setup, discover the current MCP tools and API docs at runtime before invoking named tools. Then call get_user_info to verify authentication, get_rowset_capabilities to load the current Rowset feature guide, and get_all_datasets to discover available datasets, get_archived_datasets before restoring archived datasets, and search_datasets when you need filters. If auth fails, confirm the env var contains the full key, not only its prefix. Use search_rows for ranked row search across datasets, or get_dataset before dataset-specific row work so dataset context, schema, and relationships are in context. Use search_dataset_rows for ranked row search within one dataset. Use create_dataset when you need to create a dataset on the fly. Use update_dataset_public_preview when the user asks for a shareable read-only preview.

Sign in and use the dashboard copy button when you want the full prompt with the API key included.

Choose permissions

When creating an agent API key, choose the smallest permission level that fits the agent's job:

  • Read for inspection, exports, and reporting.
  • Read + write for agents that create or update datasets, rows, projects, relationships, or public preview settings.
  • Admin for trusted automation that needs to create other agent API keys through REST or MCP.

Installable skills

The canonical setup skill lives in the Rowset repo. The app also serves that same checked-in file as markdown at:

https://rowset.lvtd.dev/SKILL.md

Agents that support the skills CLI can install it with:

npx skills add LVTD-LLC/rowset

The source text is available at:

https://raw.githubusercontent.com/LVTD-LLC/rowset/main/.agents/skills/rowset/SKILL.md

The skill gives agents durable setup instructions for Rowset MCP and REST fallback. It tells agents how to discover the current tools and API docs instead of hardcoding an endpoint list.

The repo also includes two companion skills:

  • rowset-features for explaining the current Rowset feature surface
  • rowset-use-cases for concrete dataset patterns such as CRMs, task boards, feedback trackers, content pipelines, catalogs, and QA trackers

The app serves those skill files at:

https://rowset.lvtd.dev/skills/rowset-features/SKILL.md
https://rowset.lvtd.dev/skills/rowset-use-cases/SKILL.md

Agents and search tools can also read the generated Rowset overview:

https://rowset.lvtd.dev/llms.txt

For MCP, store the key in a private environment variable such as ROWSET_API_KEY, then configure the MCP client's bearer-token env var to ROWSET_API_KEY. That makes the client send Authorization: Bearer <key>.

If you are deciding whether a workflow should start with MCP or REST, read When should an AI agent use MCP instead of REST?. In short: use MCP for compatible agent sessions that benefit from discovery, and use REST for scripts, backend jobs, or constrained runtimes.

For Codex/OpenClaw-compatible clients, the concrete setup command is:

codex mcp add rowset --url https://rowset.lvtd.dev/mcp/ --bearer-token-env-var ROWSET_API_KEY

Set ROWSET_API_KEY in the agent's private runtime environment before running or syncing the client. The command records only the env-var name, not the raw key.

If a client only supports custom headers, set Authorization to Bearer <key>. Use X-API-Key only for REST clients that cannot send bearer tokens.

If the agent will use REST instead of MCP, follow How to connect an AI agent to the Rowset Dataset API for the handoff checklist: scoped key, private secret storage, dataset inspection, and by-index row operations.

Recommended agent behavior

  • Prefer MCP tools over browser automation.
  • Discover current MCP tools and schemas from the connected server before acting.
  • Load the current Rowset capability guide with get_rowset_capabilities.
  • For REST fallback, inspect the current API docs from the REST API base.
  • Verify setup with get_user_info.
  • Discover available datasets with get_all_datasets.
  • Find archived datasets with get_archived_datasets before restoring them.
  • Search for a specific dataset or project with search_datasets and search_projects.
  • Create new datasets with create_dataset when the user asks for an on-the-fly dataset.
  • Inspect one dataset with get_dataset before row operations. The response includes dataset context, semantic schema, and relationship summaries.
  • Read rows with list_dataset_rows, get_dataset_row, or get_dataset_row_by_index.
  • Search across datasets with search_rows when the relevant dataset is unknown or multiple datasets may contain the answer.
  • Search inside one dataset with search_dataset_rows when vector search is enabled and ranked matches are more useful than a paginated row list.
  • Modify rows with create_dataset_row, update_dataset_row, update_dataset_row_by_index, and delete_dataset_row only when requested.
  • Enable or disable read-only public previews with update_dataset_public_preview only when the user asks to share a dataset.
  • Archive mistaken datasets with archive_dataset, and restore them with restore_dataset when recovery is needed.
  • Archive inactive project groups with archive_project; this hides the project without archiving its datasets.
  • Use the Dataset API only if MCP configuration is unavailable and the user approves REST API authentication. The user can copy the API key from Settings.
  • Ask before destructive actions like archiving datasets or deleting rows.
  • Keep user data private and never print credentials into public logs or messages.

Related docs