Search, skills & MCP

Sooner or later the agent needs something that is not in your repository: the current state of a library, a workflow you repeat, or a tool you already use. Ogcode has three ways to reach outside the codebase.

Web search

Web search is built in and on by default. The agent can search, fetch specific pages, and synthesise the results alongside your local code and documents — useful for checking an API, a version, or an error message without leaving the session.

To turn it off:

export OGCODE_SEARCH_ENABLED=false

Point it at your own search if you have one (for example a self-hosted SearxNG), or supply a provider key:

Variable	Purpose
OGCODE_SEARXNG_URL	Use a SearxNG instance for search
TAVILY_API_KEY	Use Tavily for search and page reading

Two knobs tune deeper research — how many pages it reads and how much of each — but the defaults are sensible and most people never touch them.

Skills

A skill is a folder with a SKILL.md file that teaches the agent a repeatable procedure — a release checklist, a migration, how your team writes docs. When a task matches, the agent loads the skill’s instructions and follows them, instead of you explaining the process every time.

Skills are discovered from standard locations automatically: an .agents/skills/ folder in the project, and ~/.ogcode/skills/ (plus a couple of other common homes) on your machine. Drop a SKILL.md in one of those and it is available.

A skill can also declare the environment variables it needs (an API token, an endpoint), and you can decide per skill whether it is allowed, denied, or gated behind an approval — so a skill you do not fully trust never runs without you saying so.

MCP servers

The Model Context Protocol (MCP) is how you connect external tools — an issue tracker, a design tool, a documentation service. Once connected, an MCP server’s tools become available to the agent like any built-in one.

Servers are declared in ogcode.json at your project root (or in your global config) under an mcp key:

{
  "mcp": {
    "ctx7": { "url": "https://context7.example/mcp" }
  }
}

• A server with a url is a remote endpoint. A command instead of a url runs a local program as the server.
• A remote server that needs a login gets an OAuth flow automatically — the first time it asks for authorisation, a browser opens and the token is remembered.
• A server can be turned off without deleting it by setting "disabled": true; nothing about it reaches the agent while it is off.

MCP tools are named after their server, so you can always tell where a tool came from in the transcript.

Where to put the config

Project settings go in ogcode.json at the repo root; machine-wide settings go in your global config. A project file overrides the global one for anything it sets — so a shared ogcode.json can describe the project’s servers while your personal tokens stay on your machine.