Supported agents
An agent can use your hub’s instructions without sending traces to PIG. Plugin distribution, native trace collection, and repository analysis have different support boundaries.
Agent support
Section titled “Agent support”| Agent | Generated instruction distribution | Native trace collection |
|---|---|---|
| Claude Code | Claude plugin and marketplace | Yes, through the opt-in pig runtime and generated hooks |
| Codex | Codex plugin and marketplace | Yes, through the opt-in pig runtime and generated hooks |
| Cursor | Cursor plugin and marketplace | Desktop Agent sessions, through the opt-in pig runtime and generated hooks, with analyzer 0.3.3 or later |
| Gemini CLI | Gemini extension | Not currently supported |
| Claude Desktop | Not a separate compiler target | Its native traces can be discovered by the Claude collector on the same host |
The targets entries in hub.yaml are claude, codex, cursor, and gemini. Include only the targets you distribute, or keep all four from pig init.
Trace collection defaults to off. When enabled, the toolchain adds runtime files and collection hooks only to the literal pig plugin for Claude, Codex, and Cursor. Claude Code’s hook also attempts local Claude Desktop collection; this does not create a separate collection schedule inside Claude Desktop. The default enrollment policy includes Cursor once your analyzer reports Cursor support, which analyzer releases 0.3.3 and later do. See host enrollment and its Cursor collection limits.
Claude Code and Codex agents in cloud sandboxes, such as Claude Tag, can also send native traces. They enroll each execution with a cloud integration grant and need no pig plugin from your hub. See Capture traces from cloud agents.
Asset support
Section titled “Asset support”The hub stores assets independently from plugins. A plugin’s includes references choose which assets it distributes. Asset metadata determines how each target receives them.
| Asset | Default behavior | Authoring guidance |
|---|---|---|
Skills in assets/skills/<id>/SKILL.md | Agent Skills on all four targets | Prefer this format for shared, task-oriented guidance; include supporting files in the skill directory |
MCP configurations in assets/mcps/ | Native MCP configuration on all four targets | Verify the server’s transport, command, credentials, and host-specific settings on every target you use |
Rules in assets/rules/ | Unsupported until declared | Add explicit target support; do not assume a repository rule applies identically across agents |
Agent definitions in assets/agents/ | Unsupported until declared | Declare native support or a supported conversion for each intended target |
Commands in assets/commands/ | Unsupported until declared | Review the host’s invocation and file format |
Hooks in assets/hooks/ | Unsupported until declared | Review event names, commands, and the host’s execution environment |
Skills use agent-skill mode by default, including in Cursor. They do not become always-on Cursor rules. An explicit Cursor rule uses .mdc output when its support declaration selects that behavior.
Place metadata for a directory asset in its asset.yaml. For a file asset, use a sibling named <id>.asset.yaml. For example, a Cursor-only rule at assets/rules/docs-style.md can use:
support: cursor: mode: nativeOther targets retain the rule’s default unsupported status. To record a deliberate exclusion for a particular target, use mode: unsupported with a reason:
support: cursor: mode: native codex: mode: unsupported reason: This rule depends on Cursor file matching.Support declarations describe the intended output; they do not make arbitrary host formats portable. Use pig verify to check the selected conversions, then install and exercise the asset on each target. A supported Markdown agent definition can be native on Claude and converted to an Agent Skill on Codex. That conversion does not preserve a separate subagent’s execution or model-selection behavior.
Git provider support
Section titled “Git provider support”| Capability | GitHub | GitLab |
|---|---|---|
| Author and review hub source | Yes | Yes |
| Validate and publish generated releases | Reusable GitHub Actions | GitLab CI template |
| Distribute generated plugins | Supported, subject to the agent’s access and import requirements | Supported by the toolchain; the agent must support the chosen import route |
| Bind repositories for trace analysis and finding issues | Yes | Not currently supported |
Cursor’s documented team marketplace import flow uses GitHub. A generated GitLab marketplace and working Git credentials do not prove that a particular Cursor importer accepts it. Pilot that path before rolling it out. See installation by agent.
The current analyzer reads the GitHub repositories an administrator selects in PIG Settings. It uses their repository IDs to associate traces with repositories and manage findings. Publishing an Instruction Hub on GitLab does not by itself add GitLab repository analysis. Choose the hub’s Git provider separately from the repositories your analyzer can currently process.
Verify your combination
Section titled “Verify your combination”Pilot each agent and plugin combination before you roll it out. For Acme, start with one writer and one developer using the docs and dev plugins. Run these checks on each agent during the pilot, and again after an asset’s target support changes:
| Check | Confirm that |
|---|---|
| Plugin installation | The agent lists the installed plugin. |
| Skill loading | The agent loads the intended skill and its supporting files. |
| Task behavior | The agent follows the skill on a representative task. |
| Collection and analysis | When trace collection is enabled, the host is enrolled, a trace is stored, and an analysis run completes. The deployment guide covers these operator checks. |
For the native session files the collector reads on each host, see Trace object and sources.