# DataChef

DataChef is an agent-first data integration, super-query, Smart Forms, pipeline, API explorer, and security workbench.

## For AI agents
- Public summary: https://chef.threatco.io/llms.txt
- Full agent guide: https://chef.threatco.io/llms-full.txt
- OpenAPI discovery: https://chef.threatco.io/openapi.json
- MCP discovery: https://chef.threatco.io/.well-known/mcp.json
- MCP endpoint: https://chef.threatco.io/api/mcp
- Agent console: https://chef.threatco.io/agents

## Connect
1. Open https://chef.threatco.io/agents as an authenticated DataChef operator.
2. Create a scoped MCP token. Start with `mcp:read` and `mcp:query`.
3. Configure your agent client for Streamable HTTP JSON at https://chef.threatco.io/api/mcp.
4. Send `Authorization: Bearer <mcp-token>` with every MCP call. The MCP endpoint itself is bearer-token protected so agents do not need a browser OAuth session.

## Recommended Super Query flow
1. Call `datachef.sources.discover`.
2. Call `datachef.query.plan` with an objective and selected source bindings.
3. Call `datachef.query.run` with the returned query text and sources.
4. Call `datachef.query.profile` or `datachef.query.explain` for evidence and citations.
5. Call `datachef.query.save_recipe` only after the query is useful and bounded.

## Recommended Smart Forms flow
1. Call `datachef.forms.inspect_source` with a dataset, connector, API service, or recipe.
2. Call `datachef.forms.generate` to get a schema-validated form definition.
3. Call `datachef.forms.validate` before saving or publishing externally produced definitions.
4. Call `datachef.forms.check_drift` before running older published forms.
5. Call `datachef.forms.regenerate` with `confirm=true` when source drift requires a new draft.
6. Call `datachef.forms.publish` only after operator approval.
7. Call `datachef.forms.run` for bounded read/search interactions; mutation actions require confirmation.

## Recommended Display flow
1. Use DataChef query, source, or form tools to gather real evidence first.
2. Use `datachef.display.card` for summaries and metric strips.
3. Use `datachef.display.table` for bounded row displays.
4. Keep display blocks grounded in prior tool results; do not invent rows, counts, or labels.

## Available MCP tools
- `datachef.display.card` (mcp:read, read) - Return validated Adaptive Card-style JSON for summary, metric, or table displays. Use this when an agent needs structured UI instead of markdown.
- `datachef.display.table` (mcp:read, read) - Return a bounded Adaptive Card-style table JSON block from columns and rows.
- `datachef.operator.delete_api_service` (mcp:mutate, destructive) - Delete an API service registry entry. Requires confirm=true.
- `datachef.operator.delete_connector` (mcp:mutate, destructive) - Delete a connector and its linked datasets. Requires confirm=true.
- `datachef.operator.delete_dataset` (mcp:mutate, destructive) - Delete a DataChef dataset. Requires confirm=true.
- `datachef.forms.check_drift` (mcp:query, read) - Compare a saved Smart Form definition with the current source shape and return added, removed, and changed fields.
- `datachef.forms.discover` (mcp:read, read) - List Smart Forms, source coverage, modes, published links, intelligence scores, and available definition-driven form workflows.
- `datachef.forms.generate` (mcp:query, execute) - Generate a schema-validated Smart Form definition from source inspection, with grouped fields, components, actions, critique, and intelligence score.
- `datachef.forms.inspect_source` (mcp:query, execute) - Inspect a dataset, connector resource, API endpoint, or query recipe and return fields, samples, risks, lookup candidates, and repair suggestions.
- `datachef.forms.publish` (mcp:mutate, mutate) - Publish a Smart Form as a locked version with DataChef URI and live route.
- `datachef.forms.regenerate` (mcp:mutate, mutate) - Regenerate a saved Smart Form draft from the current source shape while preserving published versions.
- `datachef.forms.run` (mcp:execute, execute) - Run a read/search/execute Smart Form action. Mutation actions require confirm=true and existing action permits.
- `datachef.forms.validate` (mcp:query, read) - Validate a Smart Form definition for component registry, sections, safe action bindings, and mutation confirmation boundaries.
- `datachef.query.explain` (mcp:query, execute) - Explain rendered query, source bindings, execution mode, bytes/rows scanned, warnings, and citations.
- `datachef.query.plan` (mcp:query, read) - Create a deterministic query plan with source bindings, query text, risks, variables, and the next runnable call.
- `datachef.query.profile` (mcp:query, execute) - Run a bounded query and return field profile metrics, chart hints, warnings, truncation state, and citations.
- `datachef.query.run` (mcp:query, execute) - Execute SQL/KQL/recipe/federated queries over DataChef datasets and live connector sources.
- `datachef.query.save_recipe` (mcp:mutate, mutate) - Save an approved query plan as a DataChef query recipe with variables and time-window defaults.
- `datachef.sources.discover` (mcp:read, read) - List datasets, connectors, API services, recipes, schemas, capabilities, row limits, and supported query modes.

## Safety
- Hosted DataChef operator pages remain behind app authentication.
- MCP runtime calls require `Authorization: Bearer <mcp-token>`.
- Stored secrets are never returned by discovery docs or source catalog tools.
- Destructive mutation tools require `confirm=true`.
- Smart Form definitions are versioned and action-audited.