Tiger Skills MCP Server

Emulate Claude Skills with any LLM via a Model Context Protocol (MCP) server.

What are Skills?

Skills are modular components that enhance the capabilities of an MCP-compatible agent by providing specific functionalities, workflows, and domain expertise. They transform a general-purpose agent into a specialized agent equipped with procedural knowledge that no model can fully possess.

The goal is to be fully compatible with Anthropic's skill format. See their Agent Skills Spec and related documentation for more details.

skill-name/
├── SKILL.md (required)
│   ├── YAML frontmatter metadata (required)
│   │   ├── name: (required)
│   │   └── description: (required)
│   └── Markdown instructions (required)
└── Bundled Resources (optional)
    ├── scripts/          - Executable code (Python/Bash/etc.)
    ├── references/       - Documentation intended to be loaded into context as needed
    └── assets/           - Files used in output (templates, icons, fonts, etc.)

Configuration

Environment Variables

• SKILLS_FILE: Path to the YAML file configuring the set of skills. Default: ./skills.yaml
• SKILLS_TTL: Time (in milliseconds) to cache loaded skills. Default: 5 minutes

Skills Configuration File

The set of skills is configured via a YAML file. Both local directories and GitHub repositories are supported. Config can point to individual skills or collections of skills.

local-directory-collection:
  # A collection of local skills stored in the `./skills` directory.
  # Each skill should be in its own subdirectory with a `SKILL.md` file.
  type: local_collection
  path: ./skills
local-individual-skill:
  # An individual local skill stored in the `./skills/skill-name` directory.
  type: local
  path: ./path-to/individual/skill-name
anthropic-github-collection:
  # A GitHub repo containing a collection of skills.
  # Each skill should be in its own subdirectory with a `SKILL.md` file.
  type: github_collection
  repo: anthropics/skills
  # path: ./ # not needed for this example since skills are at the root of the repo
  # Optionally specify skills/paths to ignore in this collection
  ignored_paths:
    - .claude-plugin
    - document-skills
  disabled_skills:
    - canvas-design
  # Setting enabled_skills will _only_ load the specified skills from the collection
  # enabled_skills:
  #   - frontend-design
  #   - webapp-testing
single-github-skill-example:
  # A GitHub repo containing an individual skill.
  type: github
  repo: anthropics/claude-cookbooks
  path: ./skills/custom_skills/creating-financial-models

Skill names must be unique across all configured skills. Any duplicates will be ignored with a warning.

Connection string parameters

Individual clients can control the set of skills that are enabled, as well as the protocol(s) used, via parameters in the connection string.

• enabled_skills: Comma-separated list of skill keys to enable. If not provided, all configured skills are enabled.
• disabled_skills: Comma-separated list of skill keys to disable. If not provided, no skills are disabled.
• tools=0: Disable all tools (for resource-only integration).
• resources=0: Disable all resources (for tool-only integration).

Example

http://tiger-skills-mcp-server/mcp?disabled_skills=foo,bar&resources=0

Subagent Task Execution

This MCP server (optionally) provides a subagent tool that can be used to break up complex tasks into smaller subtasks, each handled by its own agent instance. This is useful for tasks that require multiple steps, especially when those steps may require consuming large amounts of data in LLM context.

The subagent will automatically have access to the same set of skills as configured for the view skill tool, as well as the ability to (recursively) invoke further subagents. In addition, a mcp.yaml configuration file can be provided to specify additional tooling to be made available to the subagent.

If you do not wish to use subagents, you can set SUBAGENT_DISABLED=true in the environment to disable the tool.

MCP Configuration for Subagents

Create a mcp.yaml file to specify additional tools for the subagent to have accessible. This file is read from the root directory by default, or you can specify a different path via the MCP_PATH environment variable.

Only the streamable HTTP transport is supported at this time.

pg_aiguide:
  type: http
  url: https://mcp.tigerdata.com/docs

Development

Cloning and running the server locally.

git clone git@github.com:timescale/tiger-skills-mcp-server.git

Building

Run ./bun i to install dependencies and build the project. Use ./bun watch to rebuild on changes.

You will need a GitHub token with the correct scopes. Here is a direct link to create such a new token.

Create a .env file based on the .env.sample file.

cp .env.sample .env

Then update the GITHUB_TOKEN value in .env.

Testing

The MCP Inspector is a very handy to exercise the MCP server from a web-based UI.

./bun inspector

Test via HTTP

./bun watch http

Test via stdio

./bun watch stdio

Testing in Claude Desktop

Create/edit the file ~/Library/Application Support/Claude/claude_desktop_config.json to add an entry like the following, making sure to use the absolute path to your local tiger-skills-mcp-server project, and use a valid GitHub token.

{
  "mcpServers": {
    "tiger-skills": {
      "command": "node",
      "args": [
        "/absolute/path/to/tiger-skills-mcp-server/dist/index.js",
        "stdio"
      ],
      "env": {
        "GITHUB_TOKEN": "ghp_whatever"
      }
    }
  }
}