Skip to content

MCP servers

Two things share this name in VibeGuild, and it's worth keeping them apart:

  • the built-in server every agent pane connects to automatically, which is what makes agents aware of the app they're running in, and
  • third-party servers you install once and hand out to the agents that should have them.

The built-in server

Every agent pane connects to VibeGuild's own MCP server (vibeguild) with zero setup. The agent's own permission system still gates every call — the app never bypasses it — so expect one prompt per tool the first time.

What it gives an agent:

  • Where it is — its workspace, pane number, nickname, branch, and working directory, so an agent knows it is "Nix, pane 3" (get_workspace_info).
  • Who its neighbors are — every pane in its workspace with the number you see, the nickname, the agent type, the activity state, the cwd and the branch (list_panes). "Is pane 3 busy?" just works.
  • The whole repo's git state — across every worktree, which an agent's own git can't see from inside one checkout (git_overview).
  • Your editor — the open file, the tabs, and the point of it all, your current selection, so "look at this" works without pasting (get_editor_context). read_editor_buffer reads an unsaved tab's real content instead of the stale file on disk, and open_in_editor / open_in_browser let an agent show you what it made.
  • The browserread_browser_console and read_browser_page let an agent open its dev server preview, look at what rendered, and fix what it sees.
  • Panesopen_pane opens one ("a codex pane called Reviewer on a new branch fix-auth", or just "3 claude panes") without stealing your focus, optionally on its own branch and, in a multi-repo workspace, in a named repo.

Agents also work together. open_pane can hand a new agent its first task, and send_prompt messages an existing agent by nickname or pane number ("ask Nix to review my diff"). Messages appear in the target's terminal, visibly attributed — [message from Vex, pane 2] … — and are delivered only when the target is idle: a busy agent refuses rather than gets interrupted, or parks the message for its next idle moment. Relay loops are rate-limited, and plain terminals are never messaged.

A delegating agent can check on its helpers with read_pane (a read-only look, never a keystroke) and tidy up with close_pane — though a running pane you opened is yours alone, and agents are told to ask instead. set_task_description keeps pane titles honest: "Rex: frontend work" while it runs, cleared when done. The name itself stays yours.

Agents can only open panes in the workspace that's on screen. The memory and skills tools ride the same connection — see Memory & skills.

Third-party servers

The v m panel is where Playwright, Context7, GitHub, Supabase, Exa, Firecrawl and anything else get installed. All of them are static-key servers, so there is no OAuth dance; the custom form takes the mcpServers JSON snippet every server README ships, pasted whole, and fills itself in for you to check.

The app is the single source of MCP truth for its panes. Servers configured in a personal ~/.claude.json or a repo's .mcp.json do not load inside VibeGuild, which is what makes the panel's list trustworthy: what it shows is what your agents actually have. It also means eight agents never spawn eight copies of the same npx server, and your API keys stay in the app rather than in every agent's environment.

Installing validates immediately — "connected · N tools", or the real error with a link to go get a key — and enables the server nowhere. From there:

  • Each row's switch turns the server on for this project.
  • The row expands to hold an everywhere toggle and a reset back to it.
  • Each agent pane's menu can force a server on or off for that agent alone — give only the reviewer pane the GitHub tools — or fall back to the project's choice.

Toggling a server off cuts running agents off instantly. Toggling one on needs an agent restart, so the panel names which agents are affected and restarts this workspace's in one click (busy agents behind a confirm, sessions resuming where they left off). Agents in other open workspaces get a "switch there to apply" line instead — restarting stays a visible, local act.

If a server dies in the background, the status bar raises an amber flag rather than letting your agent quietly lose its tools.

Agents that read MCP config from their own files

A few CLIs can't be handed MCP config at launch. Grok Build and Kimi Code instead take a one-time entry in their own config file, offered with a copy button from the pane's menu → MCP setup…. Neither entry contains a URL or a key, and both stay inert outside VibeGuild panes. Third-party servers you enable ride along that same connection, their tools namespaced by server. A pane that needs the entry but has never connected grows a slim dismissible banner instead of quietly lacking tools. The details are in Agents & profiles.

One known wrinkle: Codex cannot be told to ignore its own config, so a server you configured both there and here may show its tools twice.