Tiger Skills

The Tiger Skills MCP Server lets you run modular Skill packs that empower an MCP-compatible agent with domain-specific workflows, tools, and knowledge. It enables you to load skills from local files or GitHub repos, control which skills are active, and optionally spawn subagents to break complex tasks into manageable subtasks, all through a streamlined, protocol-based interface.

How to use

You connect your MCP client to the server using the provided HTTP endpoint or run the server locally for development. The server exposes skills as configurable modules that the client can enable or disable. Use the HTTP endpoint to send MCP requests and receive streaming responses that drive your agent’s behavior. If your task is complex or data-heavy, you can enable subagents to handle subtasks in parallel or sequentially, each inheriting the same skill set and optional extra tooling.

Key considerations when using the server include selecting the right set of skills for your task, managing large prompts by leveraging skills’ bundled references, and optionally providing a mcp.yaml to customize tooling available to subagents. You can also tune how long skill configurations are cached to balance startup time with up-to-date behavior.

How to install

Prerequisites: you need a JavaScript runtime and a tool to fetch and build dependencies. This guide uses Bun for a fast, modern development flow.

# 1) Ensure you have Bun installed
curl https://bun.sh/install | bash

# 2) Clone the project
git clone git@github.com:timescale/tiger-skills-mcp-server.git
cd tiger-skills-mcp-server

# 3) Install dependencies and build
bun i
# If a build step exists, run it (example standard pattern)
bun run build

# 4) Start the server in HTTP mode for testing (default port 3001)
bun run start http

# 5) Optionally start the inspector for a web-based UI
bun inspector

# 6) Test via HTTP endpoint (in the client, not as a script here)
curl http://localhost:3001/mcp

Additional sections

Configuration and environment variables: you can configure which skills are loaded and how long they stay cached. The server supports these environment variables to tailor behavior:

• 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

Skill configuration basics: Skills are organized in a directory structure with a required SKILL.md and optional bundled resources for scripts, references, and assets. The SKILL.md must include YAML frontmatter with a name and description, which guide when the agent should use the skill. Bundled resources can include scripts for deterministic tasks, references for in-context knowledge, and assets used in outputs.

Subagents and tooling: you can enable a subagent to handle complex multi-step workflows. A mcp.yaml file can be provided to specify additional tools for the subagent. The file is read from the project root by default, or you can set MCP_PATH to point to a different path. Only the streamable HTTP transport is supported for subagents at this time.

Testing and development: you can test the MCP server using the built-in inspector UI and validate the HTTP transport via the provided endpoints. For local development, you can run commands to watch and test both HTTP and STDIO transports.