Configure LLM Providers

Add guild-scoped provider credentials and select Gigi’s chat and routing models.

Configure LLM Providers

Guild LLM behavior is live when a server has a sealed provider credential and active model profiles.

Prerequisites

• Discord gateway enabled
• slash commands synced
• GIGI_LLM_SECRET_KEY_BASE64 configured
• your Discord user has llm.provider.write, llm.provider.test, and llm.provider.select

Generate the sealing key with:

openssl rand -base64 32

Set it as:

GIGI_LLM_SECRET_KEY_BASE64=<base64-32-byte-key>
GIGI_LLM_SECRET_KEY_ID=local-v1

Provider API keys are not environment variables. Enter them through /llm provider add or /llm provider rotate private modals.

Grant access

Create and assign an LLM manager role:

/permissions role create name:Gigi LLM Managers preset:llm-manager
/permissions role assign role:@Gigi LLM Managers user:@person

Add and test a provider

Start credential entry:

/llm provider add provider:gemini label:main

Paste the provider key into the private modal.

Then test it:

/llm provider test label:main

Select models

Set the chat model for guild mention replies:

/llm model set purpose:chat label:main model:auto

Set the routing model for semantic external app dry-run routing:

/llm model set purpose:routing label:main model:auto

model:auto picks Gigi’s static default for the credential provider and purpose. Pass a provider model ID instead when you want a specific model:

/llm model set purpose:chat label:main model:gemini-2.5-flash

Confirm:

/llm model show purpose:chat
/llm model show purpose:routing

Enable or inspect LLM tool routing:

/llm routing show
/llm routing set mode:dry-run
/llm routing set mode:enabled

New guilds default to off. Use dry-run first when testing flexible guild mention routing. Use enabled only when you want validated native memory tools and public external app dispatch to execute after capability checks.

Verify usage

Ask Gigi a guild mention after the chat model is set:

@Gigi explain recursion in one paragraph

Or use /ask:

/ask question:explain recursion in one paragraph context:none visibility:private
/ask question:how often did we mention postgres here context:channel visibility:public

context:channel lets the agent load permitted recent full-mode channel memory into a cited context pack and choose current-channel memory tools only when tool routing and memory permissions allow it.

Then check aggregate token usage:

/llm usage guild

Usage events store actor, guild, provider, model, purpose, billing owner, status, and token counts. They do not store raw prompts, completions, provider responses, or secrets.

Current limits

• Personal BYOK commands are not live.
• Rich DM chat is not live.
• Provider fallback between guild, user, and tenant owners is not automatic.