Ground → your workflow
Use Ground from your coding agent
Run a local stdio MCP server or CLI over Ground’s existing API, then inspect and save the same completed export in your project.
Start from your dataset’s export controls. Guides are optional; you do not need to visit this page to download your data.
From Ground to your destination
- From the dataset page, choose Export options → Copy agent prompt and paste it into your configured agent. The prompt identifies the completed export and explicitly reuses it without new generation. First-time setup only: install from this release’s source checkout using Python 3.10+ and the agents extra. API-key access requires an eligible Ground plan. Keep GROUND_URL and GROUND_API_KEY in the MCP process environment.
- Run ground mcp-config to print a stdio mcpServers entry using the exact installed Python interpreter. Add that entry to your client’s documented MCP configuration. It contains no key. GUI clients may not inherit shell variables; configure their process environment through the client’s supported mechanism.
- For Claude Code, a local stdio registration can run: claude mcp add --transport stdio ground -- ground mcp. Start Claude Code from the configured environment. Other clients can use the generated stdio entry where supported; client-specific marketplace installation has not been verified in this slice.
- Ask: ‘Use Ground export YOUR_EXPORT_ID. Inspect its schema, save a SQLite fixture, then check row counts and joins.’ The local server lists datasets/exports, reports status/schema and saves files/notebooks/fixtures. Saved files go under GROUND_OUTPUT_DIR, default ./ground-data; use an unused filename.
- The default toolset reuses existing exports. To expose new export generation, explicitly enable --allow-generation. New generation uses the same API permissions, plan limits and credits as Ground. Start once and poll the returned job ID; network uncertainty is not a reason to submit another billed request.
First-time agent setup
Run this from the Ground source checkout for this release. Package distribution is pending. Keep credentials outside code and notebooks.
python -m pip install -e './sdk/python[pandas,agents]' # Python 3.10+; set GROUND_URL and GROUND_API_KEY securely. ground mcp-config
Code workflow
# CLI: find the existing export and inspect its schema ground jobs ground schema YOUR_EXPORT_ID ground sqlite YOUR_EXPORT_ID --output fixtures.sqlite # MCP: start in a terminal-based client’s configured environment ground mcp # Optional: expose NEW, billed export generation ground mcp --allow-generation
Verify the result
Verify MCP initialization, tool discovery, a schema lookup and saving the fixture, then query its tables. The CLI emits JSON and returns a nonzero exit code on failure. MCP protocol smoke checks are part of this release’s test pack; actual client installation still needs separate acceptance. The optional use-ground skill lives at integrations/skills/use-ground/SKILL.md.
When something needs attention
Server unavailable: use the interpreter path emitted by mcp-config and install the agents extra there. Missing credentials: configure the MCP process environment; never paste keys into the conversation. Existing filename: choose another. 401/403: inspect credentials/entitlement. Hosted OAuth MCP, ChatGPT connectors, official client plugins and marketplace listings remain planned; a local stdio server does not complete them.