Vee

Vee is V1 at Michigan's internal Slack agent, built with the Eve framework and hosted on Vercel.

Vee responds to Slack mentions and direct messages, keeps durable conversation state within Slack threads, can use Eve's built-in tools, and connects to the V1 team's shared Notion workspace, public Slack history, Granola meeting notes, and the Vee and website GitHub repositories through MCP. Its identity and behavior are defined in agent/instructions.md; model selection lives in agent/agent.ts; and the Slack channel is configured in agent/channels/slack.ts.

Vee uses openai/gpt-5.6-luna through AI Gateway, with routing restricted to Azure. AI Gateway automatically tries the V1 Michigan team's enabled Azure BYOK credential first and may fall back to Vercel's Azure credentials if BYOK fails.

Vee is retrieval-first for factual answers. It must verify claims with at least one relevant source: Notion at minimum for V1 facts, Granola for meetings, Slack for team discussions and recent operational context, GitHub for repository questions, and web search for external facts. It names its sources and states when retrieval cannot verify an answer instead of guessing.

GitHub access is restricted to V1Michigan/vee (Vee's own code) and V1Michigan/website-v2 (V1 at Michigan's website), including filing issues and creating changes on vee/* branches for draft pull requests. Vee may suggest filing self-feedback in V1Michigan/vee; it files website improvement issues only when explicitly requested. Vee is instructed never to push to main, merge its own pull requests, alter repository settings, manage secrets, run workflows, or publish releases. GitHub branch protection is the enforcement boundary for main.

Development

pnpm install
pnpm dev

Verification

pnpm typecheck
pnpm build

Deploy with Eve after linking the project and provisioning the Slack connector:

vercel link
vercel env pull
pnpm exec eve deploy

The first person to use Notion from Slack must authorize Vee with the shared V1 Notion team account. That grant is then reused by Vee across Slack users.

GitHub uses the same shared-grant model. Authorize the github/vee connector once with access limited to the V1Michigan/vee and V1Michigan/website-v2 repositories.

Slack search and Granola use per-user authorization. Slack search requests only search:read.public, so Vee cannot search private channels or direct messages. Granola follows each user's active workspace and note permissions; its MCP server does not support a shared service account.

AI Gateway usage reporting

Vee's check_ai_gateway_usage tool uses Vercel's Custom Reporting API to show team-wide AI Gateway activity for all time, the past 30 days, the past 7 days, and the past 24 hours. Each period includes request volume, token usage, the amount charged through Vercel, and market-price usage across both system and BYOK credentials.

On Vercel, the tool uses the automatically provided VERCEL_OIDC_TOKEN. For local development, add a team AI Gateway key:

AI_GATEWAY_API_KEY=

Custom Reporting currently requires a Pro or Enterprise Vercel team. Reports can lag by several minutes. The tool caches results for five minutes and makes two reporting queries when the cache is cold. Set AI_GATEWAY_USAGE_START_DATE=YYYY-MM-DD only if the default all-time starting date of 2020-01-01 needs to be changed.