guides

gcontext add: how the agent registry works

Browse a registry of pre-built agents. Install one with a single command. It lands in your instance, uses your connections, and its commands become MCP tools.

What an installable agent is

An agent is a folder. The entry point is index.md, a markdown file with YAML frontmatter that declares id, name, description, parameters, connections, and tags.

The commands/ subfolder holds slash commands. Each command is a markdown file exposed as an MCP tool. Optional subfolders: scripts/ for executable code, data/ for configuration and state, steps/ for multi-step workflows.

Step 1: Browse the registry

The registry lists all available agents. Browse them on the web at gcontext.ai/agents/browse or search from your MCP client. Each listing shows the name, description, required connections, and tags.

Pick an agent that matches what you need. The listing tells you which capabilities it requires before you install it.

Step 2: Install

Run gcontext add with the agent ID from your instance directory. The command pulls the agent folder from the registry and copies it into agents/browser-cookbook/.

The agent is now part of your instance. You can read its files, edit its configuration, and inspect its commands before you run anything.

install an agent
$ cd my_instance
$ gcontext add browser-cookbook

What appears on disk

The agent folder contains the entry point, commands, and data. When you install an agent, the folder lands inside your instance's agents/ directory.

The agent runs with your connections and modules. It does not bring its own runtime or dependencies beyond what its connections declare.

my_instance/agents/browser-cookbook/
index.md
commands/
setup.md
record.md
replay.md
data/
config.json

Step 3: Bind connections

Agents declare abstract capabilities, not specific services. The browser-cookbook agent needs a "browser" connection. It does not specify which browser tool you use.

During /setup, gcontext reads the agent's frontmatter, finds the capabilities it needs, then looks in your connections/ folder. Two users can install the same agent and bind it to different connections. The agent code stays the same.

Step 4: Run a command

The agent's commands become MCP tools. Your client calls them directly. The tool name follows a predictable pattern.

Each command runs in the context of your instance. It has access to the agent's data, your connections, and your modules.

MCP tool names
/mcp__my_instance__browser_cookbook__setup
/mcp__my_instance__browser_cookbook__record

Updating and removing

The agent tool checks for updates from the registry. When a new version is available, it pulls the latest files into your local copy.

To remove an agent, delete the agents/name/folder. The agent's commands disappear from your MCP tools on the next server restart. Your connections and modules stay untouched.

Browse the registry

  • browser-cookbook saves browser tasks as replayable Playwright scripts.
  • post-crafter builds marketing posts on top of persuasion formulas.
  • diagram-craftergenerates diagrams from your instance's context.

See all available agents at /agents/browse.

Questions

Where do agents come from?

The registry at github.com/bleak-ai/agents. Anyone can publish an agent by adding a folder with an index.md that follows the required frontmatter schema. The gcontext add command pulls from this repository.

Can I make my own agent?

Yes. Create a folder with index.md and commands/ inside your instance. The agent works locally right away. To publish it, add the folder to the registry repo and open a pull request.

Do agents share my data?

No. Agents run inside your instance. They use your local connections and state. Nothing leaves your machine unless a connection you configured sends it.

Related

Browse agents
See all available agents in the registry
Share context across tools
Connect multiple AI clients to one state folder
Self-improving agents
The learnings-file pattern
Install an agent in one command

Browse the registry. Pick an agent. Run gcontext add.

Browse agentsView on GitHub