Turn on the AI assistant

Give users the in-app assistant that answers questions about their data and, after they confirm, makes changes for them. It's off by default. About 10 minutes; it sets an api secret, restarts the api, and turns on two feature flags. How the assistant works: Chat agent.

Before you start

• An Anthropic API key
• Access to the api's environment variables
• Access to the database: the Supabase SQL editor, or psql with the DIRECT_URL

Steps

1. On the api, set the key. On Render, add it under monark-api → Environment; it isn't in the blueprint.

   ANTHROPIC_API_KEY=<your-key>
   

   Optional: CHAT_LLM_MODEL (default claude-sonnet-4-5-20250929), CHAT_LLM_MAX_TOKENS (default 4096), and CHAT_ASSISTANT_NAME (default Chrysa).

2. Restart the api (on Render, saving the variable redeploys it). The key is read on first use, so a missing key shows up only when someone sends a message.

3. Turn on chat.enabled (the chat panel and launcher) and chat.ai-agent (the AI replies) for the whole deployment. There's no admin page for flags, so write the overrides directly:

   UPDATE "FeatureFlagOverride" SET enabled = true, "setById" = 'operator', "setAt" = now()
   WHERE module = 'chat' AND "flagKey" IN ('enabled', 'ai-agent')
     AND "organizationId" IS NULL AND "userId" IS NULL AND "roleId" IS NULL;
   
   INSERT INTO "FeatureFlagOverride" (id, module, "flagKey", enabled, "setById", note)
   SELECT gen_random_uuid()::text, 'chat', k, true, 'operator', 'AI assistant turned on'
   FROM (VALUES ('enabled'), ('ai-agent')) AS f(k)
   WHERE NOT EXISTS (
     SELECT 1 FROM "FeatureFlagOverride" o
     WHERE o.module = 'chat' AND o."flagKey" = f.k
       AND o."organizationId" IS NULL AND o."userId" IS NULL AND o."roleId" IS NULL
   );
   

   Flags are read on every request, so the change is live at once. The api must have started at least once since the chat module was added, so the flag rows exist. To let each organization name its own assistant, turn on chat.org-branding the same way.

4. Give access. Built-in administrators already have it. Other users need a role with the chat use permission; see Administer.

Verify

Sign in as a user with access. Click the sparkle icon in the top bar (or press Cmd+J / Ctrl+J), and ask "Which data models do we have?". The reply streams in, and a step card for looking up your data models completes.

Turn it off

Set chat.ai-agent back to off. The assistant stops replying immediately; conversations stay stored.

UPDATE "FeatureFlagOverride" SET enabled = false, "setAt" = now()
WHERE module = 'chat' AND "flagKey" = 'ai-agent'
  AND "organizationId" IS NULL AND "userId" IS NULL AND "roleId" IS NULL;

If something goes wrong

Next: Choose a webhook secret store