Skip to main content
Agent Plugins is an open package format that can ship Agent Skills and MCP configuration together. OpenWork publishes a standards-compliant package for OpenWork Connect at integrations/agent-plugins/openwork-connect.

​
What the OpenWork package installs

The package contains:
  • plugin.json, targeting the published Agent Plugins 1.0.0 schema for broad client compatibility;
  • mcp.json, pointing at the OpenWork Connect Streamable HTTP endpoint;
  • an openwork-connect skill that teaches the agent to search for an exact organizational capability before executing it.
Install or copy the complete package directory using your client’s Agent Plugins workflow. The standard does not define a registry or one universal installation command, so the last installation step depends on the client. The package does not contain credentials or an Authorization header. Your MCP client discovers OpenWork OAuth from the endpoint, opens browser sign-in, and scopes the resulting authorization grant to the selected organization. The Agent Plugin selects the streamable-http transport; it does not pin a wire-protocol revision. OpenWork Connect negotiates stateless MCP 2026-07-28 with current clients while retaining its MCP 2025-11-25 compatibility path. Existing clients can therefore continue using the same endpoint as they move to the current MCP protocol.

​
What this enables for an organization

An administrator can manage skills, shared MCP connections, and connected services once in OpenWork. Members can then install one portable OpenWork package in a compatible agent and receive only the capabilities granted to them through their organization, teams, and individual access rules. This separates distribution from access control:
  1. the Agent Plugin tells the client how to reach OpenWork;
  2. OpenWork determines what the signed-in member can discover;
  3. search_capabilities returns relevant authorized capabilities;
  4. execute_capability runs the selected capability or returns an actionable connection or permission state.
The result is one reusable organization capability layer across compatible agents instead of a separate MCP and skill configuration for every client, teammate, and machine.

​
Import Agent Plugins into OpenWork Library

OpenWork’s GitHub plugin importer recognizes the published Agent Plugins 1.0.0 schema and the schema-identical 1.1.0 working draft at a repository root or subdirectory. It validates plugin.json, requires mcp.json to target the same schema version, discovers only immediate skills/<name>/SKILL.md entrypoints, and reads the package-root mcp.json. Existing 1.0.0 packages remain valid when 1.1.0 packages are imported. The Den import path currently supports:
  • standards-compliant Agent Skills whose instructions are contained in SKILL.md;
  • remote streamable-http MCP servers;
  • legacy remote sse entries supported by the existing connection runtime;
  • independent failure reporting, so one invalid skill or server does not hide valid siblings;
  • Agent Plugin format and schema-version provenance on the imported Plugin.
The Den import path reports but does not install:
  • local stdio servers, because they require a desktop execution and trust policy;
  • remote entries with static headers, because OpenWork does not yet persist or forward arbitrary package headers;
  • additional scripts or assets referenced by a skill. The SKILL.md entrypoint is imported and the omitted assets are reported.
These limits are explicit in the import preview so an administrator cannot mistake a partial package for a fully configured organizational Plugin.

​
Authentication

Agent Plugins does not define portable OAuth fields. Do not add an OAuth object, bearer token, or secret header to the package. OpenWork Connect uses MCP protected-resource discovery, browser authorization, PKCE, and the client’s credential store as described in the OpenWork Connect reference.