Aller au contenu principal

Agents Catalog

When you create an Agent, you can start from a template — a pre-built specialist (a PM, a Coder, a Researcher, …) with an identity, a role prompt, default Skills, and a starter knowledge base. These templates live in the separate ever-works/agents GitHub repo and are surfaced in the Create-Agent wizard as chips and a View All catalog.

This page covers the catalog repo structure, the starter agents, and how the platform reads it. For the Agent runtime and REST surface, see Agents (Your AI Employees) and the Agents API.

Key sources:

  • apps/api/src/agents/agent-template-catalog.service.ts — the catalog reader
  • apps/api/src/agents/agent-templates.controller.ts — GET /api/agent-templates
  • ever-works/agents — the template repo (manifest.json + templates/<slug>/…)

Catalog repo structure​

The ever-works/agents repo has a top-level manifest.json that indexes the templates, plus one folder per template under templates/<slug>/. A template folder contains (verified from templates/starter-coder/):

templates/starter-coder/
  SOUL.md              # the agent's identity / persona
  .works/agent.yml     # the agent manifest (config)
  skills.yml           # default Skills to attach (required / recommended)
  kb/                  # starter knowledge base (playbooks, checklists, examples, templates)
  prompts/             # system prompt + per-task prompts
  README.md
  icon.svg

Each manifest.json templates[] row carries a slug, a display name/title, a summary, a scope (e.g. TENANT, WORK, MISSION), an avatarIcon, and tags.

The skills.yml file is the attach manifest documented in Skills Catalog → Attaching Skills to Agents: its required / recommended slug lists resolve against the ever-works/skills catalog.

The 12 starter agents​

The catalog ships 12 starter agents spanning common team roles:

SlugRoleScope
starter-pmPMTenant
starter-coderCoderWork
starter-researcherResearcherTenant
starter-copywriterCopywriterTenant
starter-marketerMarketerTenant
starter-salesSalesTenant
starter-supportSupportTenant
starter-curatorCuratorWork
starter-growthGrowthTenant
starter-designerDesignerWork
starter-devopsDevOpsWork
starter-founderFounderMission

(The scope column reflects each template's scope in manifest.json and controls where the resulting Agent operates — across the whole Tenant, a single Work, or a Mission.)

How the platform reads the catalog​

AgentTemplateCatalogService.list(entity) (backed by GET /api/agent-templates?entity=agent) fetches manifest.json from ever-works/agents, maps each row to the stable AstTemplateEntry shape the web app consumes, and caches it for 1 hour in the shared cache store. Details verified in agent-template-catalog.service.ts:

  • Only agent is repo-backed. Requests for skill / task entity types return []; the web layer keeps its own fallback for those.
  • Resilient by design. Every failure path — no token, repo unreachable, malformed manifest — returns an empty array, so the web app falls back to its built-in list and the chips never break.
  • Security hardening. Because the manifest is external input, the reader rejects slugs that don't match a strict allowlist, strips HTML from every string field, and caps field lengths — so a compromised repo can't inject XSS into the catalog UI.

Authentication​

The ever-works/agents repo is private. The reader resolves a GitHub token in priority order:

  1. The platform GitHub App's installation on the ever-works org — the same App that already lets the platform create repos there. No extra secret is needed in the standard hosted deployment.
  2. EVER_WORKS_AGENTS_TOKEN / GITHUB_TOKEN env override — for self-hosted installs that don't run the GitHub App, or local dev.

When neither resolves, the service logs once and returns [].

Pin the ref in production

The catalog ref is EVER_WORKS_AGENTS_REF (default main). The service warns when it's a mutable branch and recommends pinning to a commit SHA (40 hex chars) or a version tag (vX.Y.Z) in production, to prevent supply-chain substitution after the cache expires.