Eve Design Agent Template

Eve Design Agent Template is a boilerplate for deploying a Slack design assistant that answers questions strictly from an approved design corpus committed to your repository. It is built on the Eve agent framework (version 1.0.0-rc.0, a release candidate), runs on Vercel, and is published by Vercel Labs under the MIT license.

What is the Eve Design Agent Template?

It is a Slack-connected AI agent template whose sole job is to answer design questions using a curated knowledge corpus rather than general web or Slack knowledge. It takes inputs in the form of Slack direct messages and channel mentions (including attached images and documents as temporary context) and produces replies grounded in the bundled knowledge/guidelines/ rules, flagging equal-priority conflicts and, in channels, mentioning the designated design owner. It deploys to Vercel and connects to Slack as its only connector.

Key Features

• Slack DM and channel support — answers DMs directly; in channels it responds only when mentioned, and each DM or top-level mention starts a conversation that continues in its reply thread (channel replies require re-mentioning the agent).
• Corpus-bounded answering — the agent reads and searches only the design corpus bundled at build time from knowledge/sources/ (immutable source snapshots) and knowledge/guidelines/ (concise rules), indexed by knowledge/manifest.json for identity, ownership, provenance, precedence, access, and approval.
• Temporary attachment context — message text, images, and documents are used as temporary context; runtime conversations and attachments never modify the corpus.
• Conflict flagging — surfaces equal-priority conflicts and, in channels, notifies the design owner.
• Labeled recommendations — labels general recommendations when that option is enabled.
• Locked-down runtime — shell, file writes, web access, delegation, todo management, and sandbox network access are disabled; Slack is the only connector.
• Model override — set DESIGN_AGENT_MODEL to replace the default model, anthropic/claude-sonnet-4.6.
• Toolchain — requires Node.js 24 and pnpm 10; ships a pnpm check validation script plus verify:knowledge, test, type-check, and build commands.

How does it work?

Deploy the template with the Vercel button and authorize Slack, clone the generated repository, install dependencies, then open it in a coding agent (Codex, Claude Code, Conductor, or similar) and follow BOOTSTRAP.md to run a question-by-question interview that drafts the design corpus. Until the corpus is explicitly approved, the agent replies that setup is incomplete; after approval, commit and push to main to deploy. A CLI path (pnpm install then pnpm run setup) links the project, creates or reuses a Slack connector, attaches the production trigger, deploys, and checks the Eve health and Slack routes. REFRESH.md covers later reviewed knowledge updates.

Who is it for?

• Design system owners: publish guidelines once and let teammates query them in Slack instead of hunting through docs.
• Engineering and product teams: get answers about approved design rules where conversations already happen.
• Teams without written guidelines: use the bootstrap interview to produce a first reviewed corpus with the design owner.

What are its limitations?

It cannot search Slack or the web, change its corpus, edit artifacts, generate websites, or deploy projects. The knowledge base is fixed at build time, so updates require a commit and redeploy.