AI Report Builder (MCP)

The AI Report Builder lets an administrator connect their own AI assistant, claude.ai (web), Claude Desktop, or Claude Code, to the SALT server and build Quarto reports by chatting. The assistant can read the survey data dictionary and aggregate summaries, write the report’s R/Quarto, and render it on the server, iterating until it looks right. The finished report appears in Reports → History like any other.

It works over the Model Context Protocol (MCP). Your AI subscription powers the work, so SALT never sees an API key. The assistant receives the survey data dictionary, aggregate summaries, and the reports you generate. Reports are normally aggregate-only, but anything included in a report (including any individual-level detail) is shared with the AI service you connect, so design your reports accordingly.

Enabling it (administrator / ops)

The feature is on by default. The main thing to set is the server’s public URL so AI clients can connect (and to disable it, set MCP_ENABLED=false):

MCP_PUBLIC_URL=https://your-salt-host.example.org   # the public HTTPS URL of this server
# MCP_ENABLED=false                                 # only if you want to turn it off

The server must be reachable over HTTPS at a public URL so the AI client can connect. Sessions expire automatically: an access token lasts 30 minutes (refreshed silently) and every session is capped at 6 hours, after which you sign in again.

The management dashboard’s Reports tab has an AI Report Builder button that shows your exact connector URL and these connection steps.

Connecting from claude.ai

1. In claude.ai, open Settings → Connectors → Add custom connector.
2. Enter the connector URL: https://your-salt-host.example.org/mcp.
3. Click Connect. A SALT sign-in page opens; log in with your administrator account and click Allow.
4. Start a chat and say “Help me build a SALT report” (see below).

Connecting from ChatGPT

1. In ChatGPT, open Settings → Connectors (this requires a plan that supports custom connectors / developer mode).
2. Add a connector / MCP server and enter the connector URL: https://your-salt-host.example.org/mcp.
3. Authorize by signing in with your SALT administrator account.

Connecting from Claude Code / Claude Desktop

Claude Code:

claude mcp add --transport http salt https://your-salt-host.example.org/mcp

Run a tool or start a chat; Claude Code opens a browser for the same SALT sign-in. Claude Desktop uses its Settings → Connectors dialog with the same /mcp URL.

Building a report

Once connected, in a new chat just say:

Help me build a SALT report

The assistant takes it from there: it reads the survey’s data dictionary and summaries, writes the report, renders it, and refines it as you describe what you want (for example, “add HIV prevalence by age band 18 to 25, 25 to 40, 40+”). Finished reports appear under Reports → History.

What the assistant can do

Capability	Notes
List surveys, read the data dictionary	One row per export variable (survey, rapid tests, labs)
See aggregate summaries	table()/summary() per variable, never raw rows
Read example report templates	The same templates as the Reports editor
Create, update, and render reports	Rendered HTML/PDF/DOCX appear in Reports → History

The data tools never hand back raw response rows, so the assistant works from aggregates; it does see the report it builds (its rendered markdown preview), so whatever a report displays is visible to it. Reports it produces are ordinary reports: you can edit, schedule, and download them from the Reports tab.

Security notes

• Connecting requires an administrator login; MCP access grants the same report capabilities an administrator already has.
• Tokens are stored hashed and can be revoked; every session ends within 6 hours.
• Rendering executes report code on the server, exactly as the existing Reports engine does.