Last updated: October 2026
Your coding agents load Model Context Protocol (MCP) servers from their own config files. Kepler reads those files and lists exactly the servers each agent loads. From Kepler, without editing the agent’s config by hand, you can:
- See the servers each agent loads.
- Turn a server, or one of its tools, off where you do not want it.
- Add or remove a server.
- Sign in to a server that needs it.
You manage MCP servers in 3 places:
- Per repository, from the composer’s Agent settings pill
- Per repository, in Settings → Repositories → repository → MCP servers
- As global defaults, in Settings → Agents → MCP servers
Requirements and limits
- Works for Claude Code, Codex, GitHub Copilot, Cursor, OpenCode, Auggie, and Grok Build. Pi, Antigravity, and custom agents do not expose their MCP servers to Kepler. For these agents, the manager reads This agent does not expose its MCP servers to Kepler.
- The agent must be installed. Settings lists only the agents that answer.
- Kepler shows the servers the agent’s config files define. Kepler does not list servers that are not stored on disk, such as claude.ai connectors.
Where you manage them
| Where | What a change applies to |
|---|---|
| The composer’s Agent settings pill → MCP servers → Manage | This repository, for that agent |
| Settings → Repositories → repository → MCP servers | This repository, for the agent you pick |
| Settings → Agents → MCP servers | Every repository, as the default each one inherits |
The MCP servers dialog in the composer includes Open repo settings, which opens the same list in Settings → Repositories.
Every place shows the same manager. It groups servers by where they are defined:
| Group | Where the server comes from |
|---|---|
| Added by you | Your own config for the agent, used in every repository |
| In this repo, shared with your team | Checked into this repository |
| In this repo, just you | This repository only, not shared |
| From plugins | Installed by a plugin |
| Set by your organization | Managed by your organization |
Each row is a server. The switch on a row turns the server on or off, and clicking the row opens the server’s own page. A row can carry these badges:
- Sign in needed: the server needs authentication before the agent can use it.
- Every repo: a change to the server reaches every repository.
- Locked: Kepler cannot edit this entry in the agent’s config file, such as a server your organization manages.
Per repository
To manage the servers for one repository:
- In a session or in the New task composer, open the Agent settings pill. Its MCP servers row reads {enabled} of {total} enabled.
- Click Manage. The MCP servers dialog opens. Its subtitle reads {agent} in {repo}, the agent and the repository the dialog applies to.
When you open a server’s page, All servers takes you back to the list.
A change here applies to that repository only. Worktrees of one checkout share the setting, so a server you turn off in a repository stays off in every task’s worktree of it.
If a server is off in Settings and you have not turned it on here, the server says Off for every repo in Settings. Turn it on to use it here., with Open Settings to go to the global default. Turning it on in the dialog turns it on for this repository only.
Turn off a single tool
A server’s page lists its Tools under the Let {agent} use this server switch. Kepler connects to the server to ask for its tools, so the list can take a moment to load (Asking the server for its tools…). If the server needs authentication first, the list reads Sign in to this server to see its tools.
To turn off one tool, open the server’s page and turn off the tool’s switch. Turn a tool off to keep {agent} from calling it.
Turning off a single tool works for Claude Code, Codex, and Grok Build. For the other agents, turn the whole server off instead.
Global defaults
Settings → Agents → MCP servers is the same manager with no repository. Its description reads Give your agents extra tools, like a database, a browser or your issue tracker. Use the Agent picker at the top to choose the agent, and the account when that agent has more than one.
Keep the GitKraken MCP server installed in every agent is on by default. Gives every command-line agent Git tools and your issues and pull requests through GitKraken. Turn off to manage it yourself.
The repository and global levels combine like this:
- A repository uses the global value until you change that server in its own dialog.
- A repository’s choice wins over the global one, in either direction: a repository can turn off a server that’s on globally, or turn on a server that’s off globally.
- Kepler stores a choice only while it differs from the level beneath it, so switching a server back to match the global default hands control back to the default.
You cannot change plugin servers or Cursor servers in Settings, although plugin servers are listed there. Cursor keeps its on and off state per project, so switch a Cursor server from a repository’s dialog instead.
Add and remove servers
To add a server, click Add server: beside the section’s heading in Settings, or at the foot of the list in the MCP servers dialog. The Add MCP server ({agent}) page opens with these fields:
| Field | What it takes |
|---|---|
| Name | Letters, numbers, ., _, or -, up to 64 characters, unique where it is saved |
| Applies to | Every repo, This repo only, or This repo, shared with your team. The choices depend on the agent and on whether you are in a repository, and the field shows only when there is more than 1 choice, as in a repository’s MCP servers section |
| Runs as | Command for a local server, or URL or URL (SSE) for a remote one |
| Command and Arguments | For a command server, for example npx -y @some/server |
| URL | For a remote server, an http:// or https:// address |
| Environment variables / Headers | Key and value pairs. Values stay hidden until you choose Show value |
Kepler writes the new server into the agent’s own config file, in the format that agent uses, so the agent’s CLI sees it too.
To remove a server, open the server’s page and click Remove server. Kepler deletes the server from the agent’s config. You cannot undo this.
You cannot remove these servers from Kepler:
- Plugin servers.
- Managed servers.
- GitHub Copilot servers defined in a repository. You can remove only the Copilot servers in your own config.
Copy a server to other agents
To use a server you already set up with another agent or account, copy it instead of adding it again. Kepler copies the server’s full definition, including its environment variables and headers.
- Open the server’s page.
- Click Also add to….
- Select each agent, or each account of an agent, that should get the server.
- Click Add to N destinations.
Kepler adds the server to each destination in turn. If one copy fails, the others still go ahead, and the failed destination stays selected with its error.
You cannot copy servers from plugins, so their page does not show Also add to….
Sign in to a server
For a server that needs authentication, its page has a Sign-in section. It offers Sign in, or Re-authenticate once you are signed in, and Clear authentication to sign the server out. Kepler signs in by running the agent’s own CLI, under the same account the session uses, so the credentials end up where the agent looks for them. Sign-in is available for Claude Code, Codex, OpenCode, and Cursor servers.
In a local window, Open sign-in page opens your browser, and Kepler finishes when you are done (Finish signing in in your browser). In a window connected to a remote environment, sign in on that page, then paste the URL you land on into Redirect URL.
When changes apply
| Where the change lands | When it takes effect |
|---|---|
| A Claude Code Chat session | Straight away. Kepler reconnects the session with the new servers and keeps the conversation and your model and mode picks |
| A terminal session, or any other agent’s session | On restart. The Agent settings pill and the dialog show {count} MCP changes apply when the agent restarts. with Restart agent |
| The New task composer | The session you start comes up with that state |
Where Kepler saves an on or off choice depends on the agent:
- Claude Code, Codex, OpenCode, and GitHub Copilot: Kepler does not edit the agent’s config. Kepler keeps the choice itself and passes it to the agent when the agent launches, so the same agent run from your own terminal is not affected. The one exception is turning on a Claude Code server that Claude’s own config lists as disabled: Kepler can do this only by editing that list.
- Cursor and Grok Build: Kepler writes the change into the agent’s own files, which their CLIs also read.
- Auggie: Kepler lists the servers without a switch, because Auggie has no way to turn a server off. To stop using a server, edit Auggie’s config.
Adding and removing a server always edits the config file.
Each agent account keeps its own MCP setup. When Kepler launches a Claude Code account other than the default one, it first copies in the default account’s servers.
Related
- Agent Settings: Settings → Agents, including the MCP servers section.
- Agent Integrations: the agents Kepler runs and how to set them up.
- Agent Sessions: the composer’s Agent settings pill, and restarting a session.
