Install GroundX Agent Harness
This page covers installing GroundX Agent Harness for each supported client,
connecting the optional hosted MCP server, and verifying the install. The
public plugin source is
GroundX-Studio/groundx-agent-harness
on GitHub.
GroundX Agent Harness is built on the open Agent Skills (SKILL.md) standard.
Claude and Codex install it as a one-command plugin; any other skills-capable
agent (Cursor, Replit, Gemini, Windsurf, Copilot, and more) uses the same
portable skills by adding the repository’s skills/ folder. Installing the
skills is the primary step; the hosted MCP server is a separate, optional add-on.
Before You Start
- A GroundX API key. Sign in or create an account at
https://dashboard.groundx.ai, then create or copy an API key. Wherever you enter it, use the GroundX sign-in page, an environment variable, or an approved local secret store. Never paste an API key into chat or a tool argument. - The public repo. Install commands and marketplace prompts reference
GroundX-Studio/groundx-agent-harness. Some clients want the full URL, others the shorthandGroundX-Studio/groundx-agent-harness; both are called out below. - The hosted MCP endpoint, if you plan to connect it:
https://api.groundx.ai/mcp. It is optional and prod-only today.
The plugin/skills, GROUNDX_API_KEY, and MCP are three separate pieces.
Installing the plugin does not set up API access, and connecting MCP does not
install the plugin. Set up whichever of the three you actually need. As in
What Is GroundX Agent Harness?,
“the plugin” below means just one of these three pieces (the instructions,
references, and workflows), not the whole harness.
Client support:
Claude Desktop
Most Claude app builds expose the Plugins/Cowork surface and install the plugin from a marketplace. Builds that expose only Connectors use the connector-only path below.
Install the plugin (personal marketplace). No GitHub account is needed for this public repo. Go to Customize -> Plugins -> Personal plugins + -> Add marketplace -> Add from a repository, and enter:
If the repository list cannot load, type GroundX-Studio/groundx-agent-harness
directly. Click Sync, install GroundX Agent Harness, then run
/reload-plugins or start a new session.
Organization distribution (Team/Enterprise admins). Claude organization
GitHub sync uses a private or internal marketplace repository; the public repo is
not supported as the direct organization marketplace sync target. Vendor or copy
the public bundle into your organization’s private marketplace repository at
plugins/groundx-agent-harness/. The private marketplace repository should
include:
Create the root .claude-plugin/marketplace.json with a complete marketplace
manifest. Use your organization for the root owner, keep the public bundle
plugin entry’s description, strict, and skills fields, and change only the
plugin source to the repo-relative path. You may replace author with your
approved organization publisher value:
After an admin completes the organization sync, users install GroundX Agent
Harness from Claude Cowork or Code (the organization’s plugin list),
then run /reload-plugins or start a new session.
Connect MCP (optional):
Run /mcp, connect groundx, enter your key on the GroundX sign-in page, and
start a new Claude Code session.
Claude Desktop with Connectors only
Builds that only expose Connectors (no Plugins/Cowork surface) have no plugin skills; the connector only provides MCP API tools.
Go to Customize -> Connectors, click +, then Add custom connector, and enter:
Add the connector first, then connect/authorize on the next screen; these are separate steps. Leave advanced OAuth fields empty unless Claude asks you to review discovered settings. Enter your key on the GroundX sign-in page, then enable the connector in a conversation.
Codex Desktop
Install the plugin first:
- Open Plugins -> Manage (or Manage marketplaces).
- Add a marketplace from
https://github.com/GroundX-Studio/groundx-agent-harness, refmain, leaving sparse paths empty. - Install GroundX Agent Harness.
- Start a new Codex session.
Connecting MCP is a separate, optional step: open Settings -> MCP servers,
add a server, toggle it to Streamable HTTP, and enter URL
https://api.groundx.ai/mcp. The saved MCP server entry in the MCP server list
should show an Authenticate button; click it and enter your key on the
GroundX sign-in page.
Claude CLI
Install the plugin:
Run /reload-plugins or start a new session after installing. If claude plugin
is not found, update Claude Code first.
Connecting MCP is a separate, optional step:
Then run /mcp, connect groundx, enter your key on the GroundX sign-in page,
and start a new session.
Codex CLI
Install the plugin:
Connecting MCP is a separate, optional step:
Verify and start a new Codex session:
VS Code
Open the VS Code integrated terminal and run the install for your agent.
Claude Code:
Codex:
Then reload plugins or start a new session. If claude plugin is not found,
update Claude Code first. Connecting MCP is optional and uses the same commands
as the CLI sections above; enter your key on the GroundX sign-in page.
Everything else
The harness works on any skills-capable agent (Cursor, Replit, Gemini, Windsurf,
Copilot, and more) through the open Agent Skills (SKILL.md) standard.
-
Clone the harness:
-
Add its
skills/folder to your agent’s skills directory. See your agent’s docs for where skills live. -
Reload or restart your agent so it picks up the skills.
If your agent supports remote MCP, you can also add the optional tools: in your
agent’s MCP settings, add a Streamable HTTP server with URL
https://api.groundx.ai/mcp, authenticate, and enter your key on the GroundX
sign-in page.
Verify
Run these checks without pasting secrets into chat:
If you have a local checkout of the repository, you can also run the local helper:
You can pass a specific client or check name as an argument to scope the check,
for example node scripts/doctor.mjs vscode-claude or node scripts/doctor.mjs mcp.
Next Steps
- If a verification check above didn’t behave as expected, see Verify and Troubleshoot.
- Connected MCP and ready to try a real workflow? See Ingest, Search, and Answer With an Agent or Extract Data With an Agent.

