Author, validate, and test standard Copilot Studio agents through YAML from your AI coding assistant.
Skills for Copilot Studio helps you work on standard Microsoft Copilot Studio agents from Claude Code or GitHub Copilot CLI. Describe a change to a specialist agent for authoring, cloud synchronization, testing, or design advice; the specialist invokes the relevant skills internally. Reference skills supply project structure and YAML guidance without requiring you to invoke them separately. The bundled session-start hooks are configured to supply instructions for routing requests to specialists and prepare local dependencies.
This plugin targets standard agents; GitHub Copilot harness agents have a separate Copilot Studio plugin. It is an experimental research project, not an officially supported Microsoft product. Review and validate generated YAML before pushing changes to your environment.
Microsoft describes VS Code Copilot Chat support as “in progress” because its session-start context handling differs from the CLI. Follow issue #116 for updates.
Ask the Author agent to create a topic, collect information with questions or Adaptive Cards, and adjust its triggers. For a connector action, create the connection in the Copilot Studio portal before editing its YAML.
Use the Manage agent to clone an existing agent, pull cloud changes, inspect differences, and push local edits. These operations need the Copilot Studio VS Code Extension and access to your Power Platform environment. Pushing creates a draft; publish separately in Copilot Studio when you want the changes live.
Have the Test agent run an existing evaluation test set against your agent's draft and analyze failures. This needs an App Registration with MakerOperations and Copilots.Invoke permissions. Supply a connection ID when the evaluation needs authenticated knowledge sources or connector actions; without it, those resources are not used.
Give the Test agent a results CSV exported from Copilot Studio's Evaluate tab. It can identify failures and propose changes to the relevant YAML without API access or a published agent.
Ask the Advisor agent to review your agent for design problems or troubleshoot a specific behavior. Its pattern references cover topics such as knowledge search, Adaptive Cards, and child agents; schema lookup helps investigate YAML definitions.
The specialist agents invoke skills internally. Shared project context and reference skills support their work; deprecated chat skills remain listed for compatibility.
Showing all 31
Select a cloud environment and agent, then download its YAML files.
Synchronize cloud and local agent content through the Copilot Studio Extension.
Check YAML, Power Fx, schema, and cross-file references using extension diagnostics.
Guide connector setup in the portal before editing the action's YAML.
Insert an interactive Adaptive Card into a topic.
Add topic nodes for knowledge search or AI-generated answers.
Create a variable shared across topics within a conversation.
Add website or SharePoint knowledge sources for the agent to search.
Add or edit questions, messages, conditions, and other ordinary topic nodes.
Add child agents or YAML that calls a connected agent.
Change existing connector or MCP actions, including inputs and connection mode.
Change agent instructions, display name, conversation starters, and AI settings.
Change a topic's trigger phrases or model description.
Create a topic YAML file for a new conversation flow.
Analyze exported evaluation CSVs and propose YAML fixes without API access.
Test no-auth or manual-auth agents with a token endpoint or DirectLine secret.
Test agents using integrated authentication; requires an App Registration Client ID.
Deprecated; replaced by authentication detection and separate DirectLine or SDK chat skills.
Create scenarios for testing the plugin itself, rather than Copilot Studio evaluations.
Generate test-case CSVs for import into Copilot Studio's Evaluate tab.
Query Dataverse to identify the agent's authentication mode and chat connection details.
Deprecated DirectLine chat skill retained for backward compatibility.
Run evaluation test sets against draft agents and investigate failures.
Run published-agent test suites with the Copilot Studio Kit and Dataverse permissions.
Authenticate evaluation and SDK chat workflows using an App Registration.
Supply design patterns and YAML examples for authoring, reviews, and troubleshooting.
Supply shared project structure and skill guidance to the specialist agents.
Supply YAML reference tables preloaded by the Author and Advisor agents.
List available YAML kind values, optionally filtered by keyword.
Inspect topics, trigger types, trigger phrases, and action counts.
Look up YAML definitions, properties, and schema structure.
No matching results.
Review agent designs, investigate problems, and suggest implementation patterns.
After installing the plugin in Claude Code or GitHub Copilot CLI, select @copilot-studio:copilot-studio-advisor and describe the design question or behavior you want investigated.
Create and edit topics, actions, knowledge sources, triggers, and variables.
Open your AI tool in the cloned agent's directory, select @copilot-studio:copilot-studio-author, and describe the topic or configuration change you want.
Clone agents and synchronize local content with the cloud.
Select @copilot-studio:copilot-studio-manage and ask to clone, pull, or push an agent. Install the Copilot Studio Extension first. Cloning guides you through environment and agent selection with browser sign-in; pushing may also open a browser for sign-in.
Test agent responses, run evaluations, and analyze results.
Select @copilot-studio:copilot-studio-test and describe the test you want, or provide an exported evaluation CSV for analysis. Live chat, draft evaluations, and Copilot Studio Kit suites have different authentication requirements; use the testing setup guide for published-agent testing.
At session start, supplies instructions for routing Copilot Studio requests to specialist agents.
Installing and enabling the plugin includes this SessionStart hook; it has no separate installation step. It supplies instructions that direct authoring requests to Author, design and troubleshooting requests to Advisor, cloud synchronization to Manage, and testing to Test.
For authoring, these instructions require an existing local agent.mcs.yml. If one is absent, they direct the assistant to clone an existing agent with Manage before asking Author to edit it.
In an interactive Claude Code session, open the hooks menu with the following command. Look for the plugin's two SessionStart handlers. This confirms registration, not execution.
/hooksTo check execution, launch a fresh Claude Code session from your terminal with debug logging enabled:
claude --debugIn Claude Code 2.1.296 or later, inspect the debug output for a Hooks: SessionStart ... finished with status entry for each handler and check its status. These entries show that the handlers ran; they do not establish that the assistant followed the routing instructions.
Run the following command in your terminal and find this plugin in the installed-plugin list. The list confirms installation; it does not confirm that either startup hook ran.
copilot plugin listNeither client provides a separate toggle for just one of these hooks. Disabling the whole plugin also makes its other components unavailable. In Claude Code, open the plugin manager's Installed tab, select the Copilot Studio plugin, and disable it. Follow any activation message, then start a fresh session.
For GitHub Copilot CLI, use the installed-plugin list above to find the plugin's name. Replace NAME below with that exact name, run the command in your terminal, then start a fresh session.
copilot plugin disable NAMEBoth clients also provide a disableAllHooks setting. It turns off hooks more broadly: every hook source in Claude Code, or all hooks for the repository in Copilot CLI. Use the Claude Code hooks reference or Copilot CLI hooks reference if you need that broader control.
At session start, prepares npm dependencies and records plugin paths when a plugin data directory is available.
Installing and enabling the plugin includes this SessionStart hook. It uses CLAUDE_PLUGIN_DATA or COPILOT_PLUGIN_DATA as its data directory and exits without setup if neither is available.
When the dependency manifest in that directory is missing or differs from the bundled manifest, the hook copies the bundled manifest and installs npm dependencies. It then writes the plugin root and data paths to ~/.copilot-studio-cli/plugin-paths.json. In Claude Code, it also appends those paths to the session environment file when one is supplied.
After a fresh session starts, check whether ~/.copilot-studio-cli/plugin-paths.json was written or refreshed with the expected plugin root and data directory. This indicates that setup reached the path-writing step. An existing file alone may be from an earlier session, and it does not establish that the request-routing hook ran.
For Claude Code registration and execution checks, the Copilot CLI installation check, and steps to stop both startup hooks, use the instructions under Copilot Studio request routing. Neither client provides a separate toggle for Local dependency setup; disable the whole plugin and start a fresh session to stop its startup hooks.
Install the whole plugin in your AI tool. Node.js 18 or later is required. Its session-start setup can download npm dependencies and write local plugin-path records. Cloning, pushing, and pulling agent content also need VS Code with the Copilot Studio Extension; those operations require access to a Power Platform environment with Copilot Studio.
Add the plugin marketplace from an interactive Claude Code session.
/plugin marketplace add microsoft/skills-for-copilot-studioOpen the Copilot Studio plugin's installation panel in the same session. Choose its installation scope and follow the panel's activation message before checking availability.
/plugin install copilot-studio@skills-for-copilot-studioType @ in the input and check that autocomplete lists copilot-studio:copilot-studio-manage, copilot-studio:copilot-studio-author, copilot-studio:copilot-studio-test, and copilot-studio:copilot-studio-advisor.
Add the plugin marketplace from an interactive GitHub Copilot CLI session.
/plugin marketplace add microsoft/skills-for-copilot-studioInstall the Copilot Studio plugin in the same session.
/plugin install copilot-studio@skills-for-copilot-studioType @ in the input and check that autocomplete lists copilot-studio:copilot-studio-manage, copilot-studio:copilot-studio-author, copilot-studio:copilot-studio-test, and copilot-studio:copilot-studio-advisor.