Vibe Coding
Back to Multica

Multica / agent

Skills

Attach a "professional knowledge package" to the agent - compatible with the Anthropic Agent Skills open standard.

View official sourceDocs index
Skills key concepts infographic
Skills key concepts infographic

Overview

Skill is a professional knowledge package for agents - a SKILL.md plus optional support files (scripts, configurations, reference templates, etc.) that tells the agent "what to think and do when encountering certain types of tasks." Multica adopts the Anthropic Agent Skills open standard, and all Skills that comply with the specification (packages on Anthropic official repository, ClawHub, and skills.sh) can be directly imported.

Workspace Skills and Native Skills

Multica supports two sources of Skills:

  • Workspace Skill - Available in Multica Cloud. After being attached to the agent, it will automatically synchronize to your local daemon process when the task is executed. This is the standard way for teams to share skills.
  • Local skill - directly stored in a directory on your local machine. Daemon Process scans this machine according to your request. After discovery, you can manually select it into the workspace. The daemon will scan two root directories according to priority: first scan the runtime's own skill directory (each AI programming tool has an agreed default path, such as Claude Code's ~/.claude/skills/), and then scan the cross-tool common directory ~/.agents/skills/ - this is the location shared by Codex, Gemini CLI and other ecosystems. When a skill with the same name exists in both places, the provider-specific directory takes precedence, so the common root directory will only additionally bring out new skills, and will never change the parsing results of existing skills.

In most cases, use Workspace Skill: import it once and it can be used by all team members' agents. Native Skills are suitable for testing locally first or scenarios involving sensitive local content.

Repository level skills

Repository-level skills are expected behavior. Some AI programming tools will discover project-level skills submitted in repositories, such as .claude/skills/, .cursor/skills/, .opencode/skills/, or .agents/skills/. When a Multica task checks out this repository, these files will remain in the working directory and the underlying tools can load them according to their own native discovery rules. You do not need to import these repo skills into Multica first in order to use them in this repository; Multica does not automatically import them into the workspace Skill Registry.

For tools that support project-level discovery, workspace skills are also synced to the same provider native location. If a workspace Skill conflicts with an existing repository Skill directory, Multica will write a copy of the workspace under a sibling name like review-helper-multica instead of overwriting the repository file. The tool may then see both Skills at the same time, with the workspace copy using the adjusted name.

Import Skill

Workspace Skills come from four sources:

  • New - Write SKILL.md and related files directly in the UI
  • From GitHub - Paste a repository URL (for example https://github.com/owner/repo/tree/main/skills/my-skill), Multica will automatically pull SKILL.md and all files in the directory
  • From ClawHub - Search and import from ClawHub public market, support version selection
  • From this machine - The daemon scans the skill directory of your local machine and imports the ones you want to use into the workspace.

There are size caps for individual files and for entire Skill packages (approximately 1 MB per file when importing from GitHub). The precise rules will be prompted in the import interface - an error will be reported if exceeded.

Hooked to the agent

After the Skill is imported, it needs to be mounted to a specific agent to take effect. One agent can be linked to multiple Skills, and one Skill can be linked to multiple Agents.

After hanging it, the agent will automatically get the hanging Skill when it starts working next time - different AI programming tools have their own Skill discovery paths (Claude Code is .claude/skills/, Cursor is .cursor/skills/, Antigravity is .agents/skills/, etc.), and Multica will automatically be placed in the right location. **However, 3 tools (Gemini/Hermes/OpenClaw) currently use the common fallback path .agent_context/skills/ - whether these tools can actually read skills from here depends on whether the tool itself supports **. For the difference between full path comparison and native discovery vs fallback, see [AI Programming Tool Control → Where should the skill file be placed] (/providers#skill-where should the file be placed).

After modifying the content of the Skill, only the newly created tasks will get the new version - the tasks currently being run will continue to use the old version of the Skill.

Security of third-party skills

Skills imported from GitHub or ClawHub may contain scripts and executable content. Multica itself does not perform signature verification, code review, or sandbox isolation - the content in the Skill is handed over to the corresponding AI programming tool intact, and how the tool uses these files (whether to execute them as scripts) is determined by the tool itself.

**Before importing a third-party Skill, review SKILL.md and all files that come with it. ** The "ClawHavoc" incident occurred in February 2026 - someone planted malicious instructions in popular Skill packages, and the API keys of the victim users were stolen. ClawHub later added VirusTotal scanning, but automatic scanning is not a substitute for your own review. Only import from sources you trust. For projects involving sensitive data, consider using only native skills that you write yourself.

The difference between Skill and MCP

Both are mechanisms to "enhance capabilities" for intelligent agents, with different directions:

  • Skill = structured knowledge package (static content + instructions). The agent reads the Skill to learn "what to think and do when encountering type X problems".
  • MCP (Model Context Protocol) = Tool Channel. The agent connects to external services (database, file system, third-party API) through MCP and calls them.

Both can be used at the same time. Currently, Multica's MCP support is implemented by tool: Claude Code, Codex, Cursor, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw will consume mcp_config; other tools will receive this field but will not actually use it. The topic of MCP will be expanded in subsequent versions.

At this point you already know what an agent is, how to create it, and how to add skills. Next question: **Where exactly does it run? Why does my agent sometimes get stuck? ** The next chapter talks about execution architecture—how daemons, runtimes, and tasks collaborate.

Next step