MCP Servers
- Desktop, mobile and web
- All plans
What it does
Your own MCP servers (GitHub, Notion, Sentry, a database bridge, anything you already use from a CLI) are handed to every agent AgentsRoom launches, whatever CLI it runs on, next to the AgentsRoom servers. You declare a server once, in the usual mcpServers JSON block that every CLI documents, and the desktop writes it into each CLI's generated configuration at launch. Before version 1.176.0 (2026-09-16) a user-level server only reached Claude Code and, through the Use my personal CLI configuration switch, Codex; every other provider ignored it.
Three levels declare servers, with the same logic as environment variables: merged by name, the narrowest level wins. Settings (every agent of every CLI, every project), project (every agent of that repository) and agent (that agent alone). An agent's editor lists what it inherits from the two wider levels and lets you switch a server off just for it.
Video (1 min 17 s): MCP servers, shown in the real app. In Settings > Agent behavior > MCP servers, the Microsoft Learn server is added from the catalog in one click (no key, no sign-in), then two Claude Code agents and one Codex agent are started and each one calls it from its own terminal.
Where to find it
- Settings > Agent behavior > MCP servers (first group, Agent tools): not inside one CLI's tab of Settings > AI providers, because a server is a tool you want everywhere.
- Project settings > AI providers > MCP servers. Since 2026-11-16 it also lists the servers the project inherits from Settings and from its client folder, each with its switch, and the next row, Hide my claude.ai connectors, keeps the connectors of your claude.ai account away from the project's Claude Code agents.
- Hide my claude.ai connectors also exists one level up: Settings > Agent behavior (right under MCP servers) for every project, and Folder settings > Environment (under MCP servers) for every project of a client folder.
- Add / Edit agent modal > Advanced options > MCP servers. The same field exists on a scheduled task's agent section.
How to use it
- Open one of the three MCP servers fields. Each shows a list, one row per server, with a Remote or Local badge, a Handed to agents switch and a Remove this server button.
- Add from catalog opens the MCP server catalog: servers published to the official MCP registry by their maintainers, with a search box, a Verified publisher mark and a Load more button. Add on a row builds the declaration for you: a remote server (url and headers) when the entry has one, otherwise the first launchable local package. Where a token is needed, the declaration is filled with a
{{secret:NAME}}reference, to store in the Secret Manager under that name. - From my projects opens MCP servers you already use: every server already declared in AgentsRoom (Settings, folders, projects, agents, toolkits) or in the config files of your CLIs on this computer (
.mcp.json,.cursor/mcp.json,.vscode/mcp.jsonand.gemini/settings.jsonin each project,~/.claude.json,~/.cursor/mcp.json,~/.gemini/settings.json). Each row shows where the server was found; the same server found in several places is one row. Add copies its declaration into the block; a server the block already has shows Added. A token written in plain text (an env value or header whose name contains token, key, secret, password, auth...) is moved on the way into this computer's Secret Manager and replaced by a{{secret:NAME}}reference: the server then works in every project of this computer without typing the token again, and a toolkit, which is saved on your account, never carries the token. If the vault refuses the token, nothing is added ("A token could not be saved in the secrets vault, so the server was not added."). Codex'sconfig.tomlis not read. - Paste JSON shows the raw block, so you can paste the
mcpServersobject you already use in a CLI config, with or without the"mcpServers"wrapper. A local server needscommand(plusargsandenv), a remote one needsurl(plusheadersandtype:httporsse). Errors are listed by server name; the block stays open until it parses, and the catalog refuses to add to a broken block ("The current block is not valid JSON: fix it before adding a server."). - Save. In Settings there is no Save button: a server added from the catalog is saved on the spot and its row appears in the list, its Handed to agents switch on. Every agent launched from then on, on any CLI, gets the servers; an agent already running gets them when it restarts.
A remote server that signs in with OAuth (Buddy.Works, Atlassian, Notion, Linear and the like) needs no token. Its row shows This server offers an OAuth sign-in. with a Use OAuth button when the server answers with an OAuth challenge and the entry carries no header; in the JSON it is "oauth": true on the remote entry. The row then shows an OAuth badge, Connected on this computer or Not connected, and Connect (Reconnect once connected) and Disconnect. Connect opens the server's sign-in page in your browser ("Finish signing in in your browser"); once you approve, every agent of every CLI on that computer reaches the server, its tokens renewed on their own. Agents that were already running pick it up when they restart.
A Google Cloud remote server (the BigQuery MCP server at https://bigquery.googleapis.com/mcp, and Google's other remote MCP servers) signs in with your own Google login instead, since 2026-11-11. Run gcloud auth application-default login once in a terminal, then paste the block Google documents for Gemini CLI, as is:
{ "mcpServers": { "bigquery": {
"httpUrl": "https://bigquery.googleapis.com/mcp",
"authProviderType": "google_credentials",
"headers": { "x-goog-user-project": "YOUR_PROJECT_ID" }
} } }
If you pasted only the URL ("url": "https://bigquery.googleapis.com/mcp"), the row says "This Google Cloud server signs in with your Google login, not with OAuth." with a Use Google login button, which writes the setting for you. httpUrl is read as url, and "authProviderType": "google_credentials" is what turns the Google login on (Gemini's oauth scopes and timeout are accepted and ignored). The row shows a Google login badge and "Uses the Google login of this computer: run gcloud auth application-default login once in a terminal. No key is stored in AgentsRoom." Every agent of every CLI then reaches the server through a small bridge shipped with AgentsRoom, which adds a fresh Google token to each request (renewed before it expires) and the billing project x-goog-user-project (yours from the header, otherwise the one gcloud saved with the login). If the computer has no Google login, the agent's tool answers "No Google login found on this computer. Run "gcloud auth application-default login" in a terminal, then restart the agent." From my projects also finds this block in ~/.gemini/settings.json and keeps the Google login when you add it.
The project field (since 2026-11-16) and the agent editor list the servers inherited from the wider levels as rows marked "from Settings", "from client folder" (and, in the agent editor, "from project"), so a client project can keep only some of them.
Switching a server off writes "disabled": true in its entry. On an inherited row (the agent editor shows the servers coming "from Settings" or "from project") the switch writes an override holding only { "disabled": true }, which turns the server off at that level without redeclaring it. A disabled server is skipped at launch.
Any string value (url, header, env value, argument) accepts {{secret:NAME}}; the value is resolved on the desktop at launch, never stored resolved in the declaration. Names starting with agentsroom are reserved and refused.
Settings
mcpServers(global, Settings > Agent behavior): the JSON block handed to every agent of every CLI and every project. Empty by default.mcpServers(project override, Project settings > AI providers): same block for every agent of the project; wins over the global block by name.mcpServers(agent, agent editor > Advanced options): the agent's own block; wins over both. An entry holding only{"disabled": true}switches an inherited server off.disableClaudeAiConnectors(Hide my claude.ai connectors, since 2026-11-16), three levels like the other inherited settings: global (Settings > Agent behavior), client folder (Folder settings > Environment) and project (Project settings > AI providers). Off by default at every level: the connectors of your claude.ai account (Slack, Notion, Gmail, the ones you connected on claude.ai) load in Claude Code agents as before. On: the Claude Code agents concerned start withENABLE_CLAUDEAI_MCP_SERVERS=false, Claude Code's own switch, so those connectors are not reachable at all. The nearest level that decides wins: a project's own choice, then its folder (then the parent folder), then Settings. A folder or a project can also switch it back off to keep the connectors where Settings hides them. The AgentsRoom tools and every server declared in AgentsRoom stay. Saved on your account, so every computer follows. Restart running agents to apply."oauth": trueon a remote entry (any of the three blocks): AgentsRoom runs that server's OAuth sign-in instead of sending headers. The sign-in itself (the Connect button) is not a setting: it belongs to the computer."authProviderType": "google_credentials"on a remote entry (since 2026-11-11, Gemini CLI's spelling): every request carries a token of this computer's Google login (Application Default Credentials). AnAuthorizationheader in the entry is ignored; the other headers are sent.
Agent tools (MCP)
settings_get/settings_set: read or write themcpServersblock at the global or project scope, anddisableClaudeAiConnectorsat the global, folder and project scopes (truehides the claude.ai connectors,falsekeeps them; at folder and project scopenullinherits the level above again).agents_save: acceptsmcpServerson a saved agent.
Providers
The declaration is the same everywhere; the desktop translates it into each CLI's own format at launch.
- Claude Code and GitHub Copilot CLI: a private launch file passed on the command line, so nothing lands in the project's shared
.mcp.json. For Claude Code that file is passed again when a conversation is resumed: a resumed agent keeps the servers declared in AgentsRoom (fixed 2026-09-19; before, a resumed Claude Code agent lost them, completely on an AgentsRoom account profile, invisibly when the same servers also sat in the CLI's own configuration). - Codex, OpenCode, Grok Build, Mistral Vibe, Kimi Code, Antigravity, Amp, Cursor, oh-my-pi, Freebuff, Devin: written into the CLI's project-level config file, which AgentsRoom keeps out of git.
- Codex and Mistral Vibe have no SSE transport: an
sseserver is ignored for them;httpworks. - Aider has no MCP client at all: servers are not delivered.
- claude.ai account connectors are a Claude Code feature: no other CLI loads them, so Hide my claude.ai connectors only changes Claude Code launches. A server switched off in the project field is left out for every CLI.
- A
"authProviderType": "google_credentials"server reaches every CLI as the same local command too: a bridge shipped with AgentsRoom, run by the same Node as AgentsRoom's own MCP servers (nonpx, nothing downloaded), on macOS, Windows and Linux. It speaks Streamable HTTP only (the transport of Google's servers). - An
"oauth": trueserver reaches every CLI as the same local command, the open-source mcp-remote bridge (npx -y mcp-remote@<pinned version> <url>, throughcmd.exeon Windows), so its sign-in works the same on every CLI, SSE transport included, whatever that CLI's own OAuth support.
Mobile
No editing screen on the phone (the mobile companion edits no launch settings, environment variables included). The behaviour is identical: an agent started from the phone is launched by the desktop with the same servers, from the three levels. The OAuth Connect happens on the desktop, in its browser; an agent started from the phone then uses that connection.
Limits
- Values are written resolved into the CLI's project config files, which AgentsRoom adds to
.gitignore. Committing one of those files by hand would publish a resolved token. - A
{{secret:NAME}}that does not exist in the vault is written as is, never replaced by an empty string. - Not available on team steps: a team node has no MCP servers field.
- Per project you can switch off the servers declared in AgentsRoom (Settings, client folder), and hide the claude.ai connectors as a whole for a project, a client folder or everything. Not yet per project: one claude.ai connector out of several, the servers of the CLI's own configuration (
claude mcp add,~/.claude.json, the repository's.mcp.json) and the servers of a toolkit attached to the project. - A server name only uses letters, digits, dots, dashes and underscores.
- OAuth servers need Node.js on the computer (
npx): without it, Connect says "Connecting needs Node.js: npx was not found on this computer." - The OAuth connection is per computer: sign in again on another machine. Disconnect forgets it on this computer only.
- Google login: the bridge reads the file
gcloud auth application-default loginwrites (GOOGLE_APPLICATION_CREDENTIALSfirst, thenCLOUDSDK_CONFIG, then%APPDATA%\gcloudon Windows or~/.config/gcloudelsewhere). A gcloud user login or a service account key file works; workload identity federation and impersonated service accounts are refused with that message. The Google account needs the IAM roles Google lists for the server (for BigQuery: MCP Tool User, BigQuery Job User, BigQuery Data Viewer). Queries the agent runs through Google's BigQuery server are billed by Google like any query, without the Max billed per query cap of a BigQuery database connection. Not sent to agents running on an SSH host (the login stays on this computer). Requests go out directly, not through a system proxy. For your safety the Google login only goes tohttps://*.googleapis.comaddresses: an entry that asks for it on any other address is skipped at launch (a shared block or an imported config cannot collect your Google token). - On Windows, an OAuth server whose URL or a header contains
&,|,%,^,<,>,!or"cannot be started throughcmd.exe: it is skipped and Connect says so.
Common questions
- My MCP server works in Claude Code but my Codex or Grok agents do not see it. Declare it in Settings > Agent behavior > MCP servers (or in the project settings): from 1.176.0 that block reaches every CLI. A server declared only in a CLI's personal config depends on each CLI's import rules; Grok can also import Claude Code and Cursor servers.
- Grok stays busy at startup and large MCP servers time out. On Windows, AgentsRoom sends its own tool catalog to Grok in smaller pages, preserving all tools. If large servers imported from Claude Code still cause the hang, add
GROK_CLAUDE_MCPS_ENABLED=falsein Edit agent > Advanced options > Environment variables, then close and reopen that console. For all Grok agents, put the line in Settings > AI providers > Grok tab > Advanced settings > Environment variables. For Cursor imports, useGROK_CURSOR_MCPS_ENABLED=false. This leaves Grok's status hooks, account and the servers declared in AgentsRoom working. Remove the line to restore the CLI's normal import rules. It does not fix an xAI connection failure or a slow server declared directly for Grok. - My Claude Code agents on an AgentsRoom account do not see the servers I added with
claude mcp add. Those live in the CLI's own config, which a managed account does not read by default. Turn on Use my personal CLI configuration (Settings > Multiple accounts): user-scope servers and, since 2026-09-20, project-scope servers added for that folder are then merged into every account. See Multiple Accounts. - My Claude Code agent had my MCP servers, but after a restart or a resume they are gone. Update AgentsRoom: the launch file that carries the servers declared in Settings, the project or the agent was not passed again when the conversation was resumed. It is now, on a resume as on a fresh start. If you only noticed it on an AgentsRoom account profile, that is expected: on the default profile the CLI's own configuration often lists the same servers and hid the loss.
- My Claude Code agent starts with "MCP server ... connection timed out after 30000ms" for a remote server that answers fast, and
/mcpreconnect fixes it. Update AgentsRoom: a Claude Code agent that carries a remote server you declared (Settings, folder, project, agent or toolkit) now starts with the classic MCP handshake (MCP_PROTOCOL_NEGOTIATION=legacy) and a 60 s connect deadline (MCP_TIMEOUT=60000). Those are defaults: a value you put in an Environment variables block (agent, project, folder or Settings > AI providers) always wins, so you can set a longerMCP_TIMEOUT, orMCP_PROTOCOL_NEGOTIATION=autoto get the CLI's own negotiation back. - In a client project, my agent's tool list shows the claude.ai connectors of my other clients (Slack, Notion, WordPress). Can I hide them? Yes, since 2026-11-16: Project settings > AI providers > Hide my claude.ai connectors, then restart the project's agents. Its Claude Code agents then start without any claude.ai connector; the AgentsRoom tools and the servers declared in AgentsRoom stay. The default keeps them visible, as before. For all the projects of a client at once, use the same switch in Folder settings > Environment; for every project, in Settings > Agent behavior (a folder or a project can still turn it back off).
- Can I turn off a server from Settings for one project only? Yes: open Project settings > AI providers > MCP servers, the row marked "from Settings" (or "from client folder") has a switch. Off, every agent of the project, whatever its CLI, starts without it; the other projects keep it.
- How do I give the same MCP server to my Claude Code and my Codex agents? Declare it once in Settings > Agent behavior > MCP servers (Add from catalog, search the server, Add), then start the agents: each CLI receives it in its own format at launch, and an agent already running gets it after a restart. Video: https://youtu.be/fOCHLIfk1MQ
- Is there a video that shows it? Yes, a 1 min 17 s demo in the real app: one server added from the catalog, then two Claude Code agents and one Codex agent call it: https://youtu.be/fOCHLIfk1MQ
- How do I remove a server for one agent only? In the agent's editor, Advanced options > MCP servers, switch off the inherited row: the agent stores a
{"disabled": true}override, the other agents keep the server. - My MCP server asks for an OAuth sign-in (Buddy.Works, Atlassian, Notion, Linear). Do I need a token? No: click Use OAuth on its row (or add
"oauth": trueto its entry), then Connect, and sign in in your browser. Every agent on that computer uses the connection, whatever its CLI. - Will the token of my MCP server expire? A personal access token lasts what you chose when you created it at the service (Buddy.Works offers "No expiration"); AgentsRoom does not warn before it expires, and agents just get an authorization error that day. With Connect (OAuth) there is no token to renew: the sign-in is renewed on its own, and you connect again only when the service ends it (a revoked access, or months without use: six for Buddy.Works).
- Can I use Google's BigQuery MCP server (or another Google Cloud remote MCP server) with my Google login, not as a database connection? Yes, since 2026-11-11, on every CLI: run
gcloud auth application-default loginonce, then paste Google's Gemini CLI block (httpUrl+"authProviderType": "google_credentials") into an MCP servers field (see How to use it). Do not use Use OAuth for it: Google's servers do not accept the automatic client registration that button relies on ("Dynamic client registration not supported"), and a token pasted in anAuthorizationheader expires after one hour. The BigQuery database connection (Database Connections) stays the way to browse tables yourself with a billing cap. - I connected, but an agent still does not see the server. That agent was already running: restart it.
- Do I have to paste my API token in the JSON? No. Store it in the Secret Manager and write
{{secret:NAME}}in the header or env value; the catalog does that for you. - Can I reuse an MCP server I already set up in another project, or in Claude Code, without pasting its JSON again? Yes: From my projects in any MCP servers field (Settings, project, folder, agent or toolkit) lists them, and Add copies one in, its token moved to the Secret Manager.
- Why is my server name refused? Names starting with
agentsroomare reserved for the servers AgentsRoom manages itself. - Where does a project-level server show up in the agent modal? As an inherited row marked "from project", next to the rows "from Settings".
Related
- AgentsRoom MCP: the servers AgentsRoom injects itself, always present next to yours.
- Secret Manager: the vault behind
{{secret:NAME}}. - Multi-Provider Support: the CLIs an agent can run on and their per-provider options.
- Scope Levels: the global, project and agent levels and how they merge.
- Customize Agents: the other fields of the agent editor.
- Restore Session: resuming a conversation, which now keeps these servers.
- Database Connections: the BigQuery database connection, the other way to reach BigQuery with the same Google login.