Agentfolio

Agentfolio safely plans and applies generic external collections. A collection is a version-controlled repository or folder explicitly identified by collection.yaml; this CLI repository is never implicitly a collection.

It is designed to be used by coding agents and their operators. A collection is the source of truth for an agent's desired skills, tool catalog, harness setup, and profiles. Agentfolio validates that configuration, previews its effects, and applies it repeatably; it does not run agents or track work.

flowchart LR
  A[Operator or coding agent] --> B[Version-controlled agent collection]
  B --> C[collection.yaml]
  C --> D[agentfolio plan / apply]
  D --> E[Pinned Skills CLI]
  D --> F[Collection-owned adapters]
  E --> G[Installed skills]
  F --> H[Harness-specific agent setup]

Loading

What belongs in a collection

• Skills: reusable instructions installed through the pinned local Skills CLI.
• Tool catalog: a reference inventory of tools and skills available to agents.
• Harness workflows: ordered, adapter-owned setup actions.
• Adapters: trusted, collection-owned implementations of those actions.
• Profiles: named subsets of skills and harnesses for an environment.

A collection tracks configuration, not activity. Keep tasks, agent runs, telemetry, and project status in the separate systems that own them.

Install and start

Requires Node.js 20+.

Install from npm:

npm install -g @luabagg/agentfolio

The same package is also on GitHub Packages. To install from there, add a .npmrc with the scope mapping and a token that has read:packages:

@luabagg:registry=https://npm.pkg.github.com

The command is agentfolio in every case.

To work from source instead:

npm install
npm link
agentfolio init ./my-collection
cd ./my-collection
agentfolio plan
agentfolio apply --dry-run
agentfolio apply

Collection discovery is, in order: --collection DIRECTORY, AGENTFOLIO_COLLECTION, then the nearest ancestor that physically contains collection.yaml.

Skills support uses Agentfolio's pinned local Skills CLI binary. Direct Skills installation is intentionally not documented as a floating network command.

When to use Agentfolio

Use it when agents or a team need a repeatable, reviewable setup across one or more AI harnesses. Do not use it as an agent runtime, task tracker, or replacement for harness-specific tooling.

Adapter boundary

adapters:
  local:
    executable: node
    args: [./scripts/adapter.mjs]

Agentfolio executes this literal argv without a shell; running it trusts collection code. Adapters receive one protocol JSON request on stdin and return one JSON object on stdout. Agentfolio inherits normal process authentication but never discovers, stores, or prints credentials.

plan reports selected work. apply --dry-run passes a dry-run request to adapters and does not run Skills changes. Collections own all action semantics.

Commands

agentfolio init [dir]
agentfolio list skills|harnesses|tools|profiles
agentfolio plan [--profile name]
agentfolio apply [--profile name] [--dry-run]
agentfolio doctor [--profile name]
agentfolio verify [--profile name]

See architecture and collection schema.