Registry

Appearance

Browse & Install ​

The registry is a catalog index. Browse it to find an agent or skill, then copy the command from its card and run it locally.

Find an agent or skill ​

Open Agents or Skills. Each page lists every entry in the corresponding catalog with live filters above the grid:

  • Search is a case-insensitive substring match. On Agents it covers the name, description, and derived tags. On Skills it also covers the vendor and the skill's categories.
  • Tag filter (agents) narrows by any tag derived from the agent's tools and skills.
  • Vendor filter (skills) narrows by the maintainer listed on the skill.
  • Clear filters resets everything; the count above the grid updates live.

Install an agent ​

Every agent card surfaces a copy-paste command. Click the copy icon next to the infer agents add line, then run it:

sh
infer agents add documentation-agent

The CLI resolves any agent published in the catalog by name: it reads the URL from spec.server and the OCI image from spec.deployment, then writes a local config entry pointing at the agent. From there the gateway can route requests to it - see the CLI repo for infer semantics.

Pass an explicit URL when the name is not in the catalog, or to override the resolved defaults:

sh
infer agents add my-agent http://localhost:8080

The OCI Image row (e.g. ghcr.io/inference-gateway/documentation-agent:0.6.3) appears when the agent declares spec.deployment.cloudrun.image or spec.deployment.kubernetes.image, or when the source repo lives in the inference-gateway org (in which case the image is inferred from the agent name and version). Use that tag when wiring the agent into Kubernetes or Docker Compose yourself.

Install a skill ​

Skill cards work the same way:

sh
infer skills install skill-creator --user

A skill is a markdown playbook the agent reads on demand: only its name and description are advertised to the model at startup, and the SKILL.md body is loaded lazily when the model invokes the skill. The SPDX license badge on each card (MIT, Apache-2.0, Proprietary, etc.) lets you vet redistribution terms before installing.

What the card tells you ​

Card fields map directly to the underlying schema:

  • Capabilities pills (Streaming, Push) - spec.capabilities.streaming and spec.capabilities.pushNotifications on the agent's ADL manifest.
  • Protocol binding pill (e.g. JSONRPC, GRPC, HTTP+JSON) - spec.card.supportedInterfaces[0].protocolBinding, the transport the agent prefers for A2A traffic.
  • Tools / Skills counts - spec.tools[] and spec.skills[]. Expand either list for per-entry detail.
  • Model - spec.agent.model when set.
  • Repository button - the upstream GitHub repo. For agents this comes from spec.scm.url, falling back to the _source.url provenance block the catalog aggregator injects.
  • Documentation button - spec.card.documentationUrl when set.

Where the data comes from ​

Both catalogs are fetched at runtime - this site is just a static front-end:

  • Agents: https://cdn.jsdelivr.net/gh/inference-gateway/agents@latest/catalog.json
  • Skills: https://cdn.jsdelivr.net/gh/inference-gateway/skills@latest/catalog.json

The @latest ref resolves to the catalog repo's newest release tag, so a freshly merged entry only shows up here once that repo publishes a release. Release runs in both catalog repos are manual (workflow_dispatch). For local development you can point at a fork or branch via the VITE_AGENTS_CATALOG_URL and VITE_SKILLS_CATALOG_URL environment variables.

Next steps ​