Claude MCP for teams: how to share connectors
The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to outside systems: data sources, tools and workflows. [1] Anthropic created it [4], and in Claude’s apps an MCP server you add is called a connector. [2] For one person, adding a server is one command or one settings page. For a team, three other questions decide whether it works:
- Where the server list lives, so that everyone gets the same servers.
- Whose sign-in each call uses: each person’s own account, or one shared key.
- Who says yes before a change, when a tool can write, delete or send.
Here is how Claude Code and Claude’s apps answer each, what to check before you connect a server, and, at the end, a different approach, connections on a company server, which is what we build.
Where a Claude MCP server can live
| Where you add it | Shared with the team? | Where the connection comes from |
|---|---|---|
Claude Code, local or user scope (~/.claude.json) |
No [2] | Where Claude Code runs |
Claude Code, project scope (.mcp.json at the repository root) |
Yes, through version control [2] | Where Claude Code runs |
| Claude Code, managed configuration from an administrator | Yes, to every managed machine [3] | Where Claude Code runs |
| Claude custom connector (claude.ai, Claude Desktop, Cowork, the mobile apps) | On Team and Enterprise, an owner adds it and each member connects it [4] | Anthropic’s cloud [4] |
Claude Desktop local server (claude_desktop_config.json) |
Per computer | Your local network; not available in Cowork or claude.ai [4] |
Connectors added in claude.ai also show up in Claude Code, but only while Claude Code is signed in
with a claude.ai subscription. They don’t load when an API key, a third-party provider such as Amazon
Bedrock, or a token from claude setup-token is active. [2] A team that runs Claude Code on a company
API key therefore can’t lean on claude.ai connectors and needs one of the Claude Code routes.
Sharing servers in Claude Code
A project file. Add a server with --scope project and Claude Code writes it to .mcp.json at the
repository root; check the file in and everyone on the team gets the same servers. [2] Two things come
with that:
- Approval. In an interactive session Claude Code asks before it uses a project’s servers.
claude -pruns, Agent SDK sessions and some cloud sessions can’t show that prompt, so they load the servers without asking, and so does abypassPermissionssession set to skip that prompt. To keep one out anyway, list it indisabledMcpjsonServersor start with--strict-mcp-config. [2] - Keys stay out of the file.
.mcp.jsoncan reference environment variables, as${API_KEY}or${VAR:-default}, so each person supplies their own key. [2] Claude Code reads its own and your cloud provider’s credentials, such asANTHROPIC_API_KEY, as empty in a remote server’s address and headers, so a project file can’t send them to a server it names. [2]
Precedence. When the same server is defined in more than one place, Claude Code connects once and
uses the first of: local scope, project scope, user scope, plugin servers, claude.ai connectors. A
server your organization provides through managedMcpServers ranks above all of them. [2] So a
teammate’s local entry with the same name replaces the shared one for that teammate.
For an administrator. Managed configuration has three tools. [3]
managed-mcp.jsondeploys a fixed set: users can’t add or use any other server. It is a standalone file at a system path (/Library/Application Support/ClaudeCode/on macOS,/etc/claude-code/on Linux,C:\Program Files\ClaudeCode\on Windows), usually delivered by device management; it can’t be delivered through server-managed settings. [3]managedMcpServers, in managed settings, gives every user a set of remote servers alongside their own (Claude Code v2.1.259 or later). Users can’t remove them but can turn them off for themselves. Entries must behttps://addresses and can’t name a program to run on a user’s machine. [3]allowedMcpServersanddeniedMcpServersfilter what users add. WithallowManagedMcpServersOnly: trueonly the organization’s allowlist counts; without it, a user’s own allowlist can broaden it. [3]
Two cautions from the same page. Any user on the machine can read managed-mcp.json, so keys don’t
go in it: each person’s comes from ${VAR}, OAuth or a header helper. And anyone who can read managed
settings can read a header set in managedMcpServers, so it should be a credential issued for that
whole audience, or left out so each person signs in with OAuth. [3] A managed-mcp.json also hides
the claude.ai connectors Claude Code fetches itself unless you set allowAllClaudeAiMcps, and none of
these settings reach the connectors the desktop app delivers to its own local and SSH sessions. [2][3]
Claude Code has no built-in registry that people can browse and install from: Anthropic suggests
sharing the approved list and its claude mcp add commands somewhere people will find them, such as an
internal wiki, or shipping the servers as plugins through a managed plugin marketplace. [3] To see
which servers and tools people actually use, Claude Code can report them through OpenTelemetry with
OTEL_LOG_TOOL_DETAILS=1. [3]
Sharing in Claude Team and Enterprise: custom connectors
On a Team plan only Owners and Primary Owners can add a custom connector; on Enterprise, Owners, Primary Owners and people with a custom role that includes Manage access to Libraries. [4] The owner adds it in Organization settings › Connectors with the server’s URL, chooses how people connect (“Sign in now”, “Sign in when needed” or “No sign in”) and can add request headers with a fixed credential such as an API key. [4] Members then find it under Customize › Connectors and connect it themselves. [4]
Five things to know before you roll one out:
- Each member signs in, unless you hand them a key. Members connecting individually means Claude reaches only the tools and data that member can reach, “unless the connector uses a shared credential such as an API key, which gives everyone who uses it that credential’s access.” For such a key, Anthropic’s advice is to use one limited to what members should access. [4]
- The connection comes from Anthropic’s cloud, in Claude Desktop and Cowork too. The server must be reachable over the public internet from Anthropic’s IP ranges; one on a private corporate network, behind a VPN or a firewall won’t connect unless you allowlist those ranges. [4]
- People approve tool calls. Claude can ask before it uses a connector’s tool: Allow once, or Always allow for that tool from then on. On the connector’s page each person sets tools, one by one or in groups, to Always allow, Needs approval or Blocked. [13] Anthropic’s advice is to choose “Allow always” only for a server and tool you trust to run unsupervised, and in Research, Claude can use connector tools without further approval. [4]
- The organization can hold tools back too. It can set a connector’s tools to
blocked, which keeps them from Claude in Claude Code, the desktop app and claude.ai chat, or toask, and Claude Code then prompts on every call, even inbypassPermissionsmode, in the sessions where it fetches connectors itself (the desktop app’s local and SSH sessions apply their ordinary permission rules instead). [2] - Editing means re-adding. To change a custom connector, you remove it and add it again. [4]
Whose sign-in does each call use?
Once a whole team uses a server, this matters more than where the configuration lives, because the service you connect sees the credential, not the person typing.
| Credential | What the AI can do | Whom the service’s log can name |
|---|---|---|
| Each person’s own sign-in or key | What that person can do [4] | That person [7] |
| One shared key | What the key can do, for everyone who uses it [4] | At some services, the key’s owner for every call [6][7] |
If you do use a shared key, scope it to what everyone using it should be able to reach. [4]
Before you connect a server
- Trust the source. Anthropic reviews connectors against its listing criteria before adding them to the Anthropic Directory, but doesn’t security-audit or manage any MCP server. [3][5] Claude Code’s docs say to verify that you trust each server, and that servers which fetch outside content can expose you to prompt injection risk. [2]
- Tool descriptions reach the AI. They are part of what the AI reads, so, as our own integration notes put it, connecting a server means trusting it that far. [7]
- Know the full list. Reviewing a project’s
.mcp.jsondoesn’t show every server a session can load: servers at other scopes, claude.ai connectors and plugins can add their own. [5] - Know who approves. In Claude Code’s Auto mode, the starting mode for interactive terminal sessions, a classifier model reviews actions instead of you and blocks the ones it judges unsafe; your explicit ask and deny rules still apply. [5]
- Start narrow. One person’s own sign-in before a shared key, permissions limited during sign-in, and tools held at approval or blocked until you know what they send. [4][13]
A different approach: connections on a company server
Another option is to put the connections where the team’s AI already runs, so that who may use a service, and who says yes before a change, is decided in one place. That is what we build, so read this section as ours.
Trina AX runs Claude Code, and optionally Codex, on one company server with the company’s files, the team’s skills and each person’s folder permissions; people use it through a chat app in the browser, on the desktop or on a phone. [10]
- Personal and team connections. A member adds a personal connection from the services listed in Settings › Connectors, or any other public https address by asking the AI in a chat [19]; only they use it, and nobody else sees it, admins and the owner included. An admin adds a team connection and names who may use it: everyone, groups or people. Admins manage team connections but don’t use one unless a rule names them, and who may use a connection is checked on every call. [6][7] An admin can turn members’ own connections off for the whole team, except connections they made from an admin’s template and their own Google Drive, Gmail, Calendar, Docs and Sheets, which every member can connect once the app is set up (by an admin, or by the hosting service; the four Google apps on Linux servers only); guests use no connections, and there is no admin list of allowed addresses yet. [7][14][20][24] (The screen was called External services before version 0.4.0 and MCP in 0.4.0; since 0.5.1 it is Connectors; today it is Settings › Connectors for your own and Admin › Connected apps for the team’s. [6][14][15][17][21])
- An MCP server by its address. An owner or admin can add any MCP server by its address; the server lists its tools, and you sign in with OAuth or paste a key. Discover, on the same page, lists Google Drive and the Google apps, and services that support sign-in with Trina AX, such as Notion, Linear and Atlassian; members can ask an admin to add any of them for the team. [6][14][15] A few services, such as Figma, Square and Intercom, accept only apps they approved, and since version 0.5.6 Trina AX says so instead of offering a key. [23][24] That list was checked without a real sign-in, so a service can still refuse. [7] You can also say “Connect Linear.” in a chat: the AI starts the connection, and you only sign in or paste the key into a sealed card; the card calls an address official only when it is exactly the service’s own, and shows any other address as it is, with a warning. [6][14] The same connections serve Claude Code and Codex. [7]
- Changes need a yes. Reads run at once (after a team room was read in the chat, an HTTP API’s reads
ask too). Everything else is a write and follows the connection’s policy: Ask every time, the default, shows a card with everything that will be sent, and nothing goes until a person says yes; the card offers
Send once, Always allow (for that person only) and Don’t send. Block writing refuses, and an admin can set a team connection to run without asking for the people it names. In version 0.5.2, once a chat had read something from outside (a web page, someone else’s file, a service’s answer), every change asked again; since 0.5.3 (2026-11-13), Always allow and run without asking hold after outside content too, in a chat the person runs from the Trina app, while the team leaves the Always allow option on (the default). If an admin turns that option off, members’ Always allow on team connections stops counting, and then, as for a question from Slack, changes after outside content ask again (Block writing still refuses; scheduled runs keep the rule below). [17][18] In a scheduled run nobody can say yes, so a write goes only to a team connection set to run without asking that the schedule’s card approved, through the person’s own connection to the recipients that card named, or, since version 0.5.4 and while the team’s Always allow switch is on, where the task’s own person already chose Always allow (a task that pinned recipients still keeps to them). [6][14][21][22] What counts as a read is the server’s own
label: a tool the server marks
readOnlyHint: trueruns without a card, so the card is only as good as the server’s labels. Whoever manages a connection can switch single tools off. [7] - A member’s own key when the service logs people. An admin can send a member a request with the service filled in; the member connects it with their own key or declines, and the admin never sees the key. [6][7] Since version 0.4.0 (November 9), an admin can also register a service once as a template that each member connects with their own key or sign-in. [7][12]
- Keys stay away from the AI. Keys and sign-ins are stored encrypted on the server and never shown
again. [6] The AI’s input never places a key, and a key a service echoes back is replaced with
[hidden key]before the AI, the chat or the logs see it, in the forms the server recognises: not if the service hashes or otherwise re-encodes it, and not for values shorter than 8 characters. [7] - A record of each call. Each call is kept as a summary (the tool, or the method, path and status) without its arguments or bodies, and connections being added, changed or removed go to the admin audit log. [7]
- The other direction. Your own Claude (Claude Code, Claude Desktop or claude.ai) or Codex can connect to a Trina AX team as a remote MCP server and work with the team’s files with your own permissions: reading tools, and with the write scope, writing and editing. Reading your conversations is ticked by default on the sign-in page; writing in them is off unless you tick it, and a message the app sends looks like one you typed, with no audit entry. An owner’s or admin’s connection can also manage the team’s groups. It stays off until the operator gives the server a public HTTPS address. [9][16]
What it doesn’t do. A connection is an https address on the public internet, on the standard
port: office-network and plain http addresses can’t be connected, even when Trina AX itself runs on
an office Mac. [7][8] Claude does better here: a custom connector can reach a server on a private
network once you allowlist Anthropic’s IP ranges, and Claude Code’s servers and Claude Desktop’s local
servers use the network of the computer they run on. [2][4] Trina AX speaks Streamable HTTP only and
doesn’t follow redirects; Claude Code also connects over SSE, WebSocket and local commands (stdio). [2][7]
Signing in with OAuth works with a server that supports dynamic client registration or, since version 0.5.6, a client ID metadata document (that route needs the team server to have a public https address), or, since version 0.5.1, with an app you register at the service yourself (its client ID and secret); otherwise use a key, unless the service accepts only apps it approved. [23][24]
Services that need an app of their own, such as GitHub, are left out of the Discover list; Slack is a built-in app that an owner or admin adds in Admin › Connected apps, and since 0.5.3 the AI reads Slack threads, and channels the app was invited to that the person is in, through the Slack app’s own permissions. [7][14][17] Claude can also sign in with its own published identity or your OAuth client.
[4] Signing in to a service has to be done on a computer, in a browser or the desktop app, not on a
phone, and confidential chats don’t offer connected services. [7][8] A team connection’s sign-in is an
admin’s, and everyone its rule names then uses that account, so the shared-key caution above applies to
it too. [7] There are at most 20 personal connections per person and 50 team connections. [7] The
requests made to a connected service leave the server, people with administrator access to the server
can read the stored data (on a hosted plan that includes Trina’s operators), and Trina AX has no security certification yet. [11]
The software is free and open source (AGPL-3.0) to run on your own server. Trina hosting runs a private Trina AX server for your company, in Seoul, with a 14-day free trial (a card is needed to start) and plans from $50 a month for up to 5 people. AI usage is paid separately to your AI provider. [10]
Fits when the team runs Claude Code or Codex on company accounts or API keys rather than Claude’s apps on a Team or Enterprise plan, people don’t use a terminal, and you want the connections next to the company’s files, with an admin deciding who uses each team connection.
Side by side
Claude Code .mcp.json |
Claude custom connector | Trina AX | |
|---|---|---|---|
| Who adds it | Anyone who can commit to the repository [2] | An owner (Team); an owner or a custom role (Enterprise) [4] | A member (personal: a listed service, or any address by asking in a chat) or an owner or admin (any address; team connections and templates) [6][14][15][19] |
| Who uses it | Everyone with the repository, after approving it [2] | Members who connect it [4] | Its member, or the people a team rule names [6] |
| Whose credential | Each person’s (OAuth or ${VAR}), or a key in a header [2] |
Each member’s sign-in, or a shared header key [4] | Personal or template: the member’s own; team: one account for everyone it names [7] |
| Before a tool runs | Claude Code’s permission mode and rules [5] | Allow once or Always allow; per tool Always allow, Needs approval or Blocked [13] | Reads at once by the server’s label; writes Ask every time (default; Send once, or Always allow for yourself), Block writing, or run without asking on a team connection; after outside content, Always allow still holds in the app while the team’s option is on (the default) [6][7][14][18] |
| A server on the office network | Reachable from where Claude Code runs | Only if you allowlist Anthropic’s IP ranges [4] | No [7][8] |
How to choose
- Developers in one repository →
.mcp.jsonwith each person’s OAuth or${VAR}keys, and managed settings if you need to control what loads. - A team already on Claude Team or Enterprise, working in Claude’s apps → an owner adds custom connectors, members sign in themselves, and a shared key only if it is scoped. For skills, see our post on sharing Claude skills.
- No Claude Team or Enterprise plan, Claude Code or Codex on company accounts or API keys, and people who don’t use a terminal → connections on a company server, like Trina AX.
Questions people ask
How do I share an MCP server with my team in Claude Code?
Add it with claude mcp add --scope project. Claude Code writes it to .mcp.json at the repository
root; commit that file. Keep keys out of it with ${VAR} references, and expect each teammate to
approve the project’s servers the first time in an interactive session. [2]
Can Claude Team admins add MCP servers for everyone?
Owners and Primary Owners can add a custom connector to the organization; on Enterprise, so can people with a custom role that includes Manage access to Libraries. Each member then connects it with their own sign-in, unless the owner set a shared credential. [4]
Do claude.ai connectors work in Claude Code?
Yes, while Claude Code is signed in with a claude.ai subscription; they appear in /mcp. Not when an
API key, a third-party provider or a claude setup-token token is active. [2]
Can a Claude connector reach a server on our office network?
Not directly. Custom connectors connect from Anthropic’s cloud, so the server must be reachable from Anthropic’s IP ranges: allowlist them, or use a local server in Claude Desktop’s configuration file, which uses your local network but isn’t available in Cowork or claude.ai. [4]
Try it
- Claude Code:
claude mcp add --transport http <name> --scope project <url>, then/mcpto see what loaded. [2] - Trina AX: Connect an external service, Notion, Linear and other apps and Team AX MCP connector. Source: github.com/trina-app/trina-ax. Hosting plans: ax.trina.app/#pricing.
Claude is a trademark of Anthropic. Trina AX is made by Trina Inc. and is not affiliated with Anthropic.
Sources
- [1] What is the Model Context Protocol (MCP)? · modelcontextprotocol.io/docs/getting-started/intro · read 2026-11-07
- [2] Connect Claude Code to tools via MCP · code.claude.com/docs/en/mcp · read 2026-11-07
- [3] Control MCP server access for your organization · code.claude.com/docs/en/managed-mcp · read 2026-11-07
- [4] Get started with custom connectors using remote MCP · support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp · read 2026-11-07
- [5] Security · code.claude.com/docs/en/security · read 2026-11-07
- [6] MCP connections · ax.trina.app/docs/use/mcp/ · read 2026-11-07
- [7] trina-app/trina-ax: docs/INTEGRATIONS.md · github.com/trina-app/trina-ax/blob/main/docs/INTEGRATIONS.md · read 2026-11-07
- [8] Connect an external service · ax.trina.app/guide/external-services · read 2026-11-07
- [9] Team AX MCP connector · ax.trina.app/docs/agents/mcp-connector/ · read 2026-11-07
- [10] Trina AX for LLMs (llms.txt) · ax.trina.app/llms.txt; Trina AX: Claude Code and Codex for your whole team · ax.trina.app/ · read 2026-11-07
- [11] Security and how it works — Trina AX · ax.trina.app/security · read 2026-11-07
- [12] Release v0.4.0: trina-app/trina-ax · github.com/trina-app/trina-ax/releases/tag/v0.4.0 · read 2026-11-07
- [13] Get started with connectors · claude.com/docs/connectors/getting-started · read 2026-11-07
- [14] Changelog (v0.5.2) · github.com/trina-app/trina-ax/blob/v0.5.2/CHANGELOG.md · read 2026-11-07
- [15] MCP connections (v0.5.2) · github.com/trina-app/trina-ax/blob/v0.5.2/docs/site/use/mcp.md · read 2026-11-07
- [16] Team AX MCP connector (v0.5.2) · github.com/trina-app/trina-ax/blob/v0.5.2/docs/site/agents/mcp-connector.md · read 2026-11-07
- [17] Changelog (v0.5.3) · github.com/trina-app/trina-ax/blob/v0.5.3/CHANGELOG.md · read 2026-11-07
- [18] server/src/integrations/connections/tools.ts (v0.5.3) · github.com/trina-app/trina-ax/blob/v0.5.3/server/src/integrations/connections/tools.ts · read 2026-11-07
- [19] server/src/chat/connect-service.ts (v0.5.3) · github.com/trina-app/trina-ax/blob/v0.5.3/server/src/chat/connect-service.ts · read 2026-11-07
- [20] External-service connections (v0.5.3) · github.com/trina-app/trina-ax/blob/v0.5.3/docs/INTEGRATIONS.md · read 2026-11-07
- [21] Changelog (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/CHANGELOG.md · read 2026-11-07
- [22] server/src/integrations/connections/tools.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/integrations/connections/tools.ts · read 2026-11-07
- [23] Changelog (v0.5.7) · github.com/trina-app/trina-ax/blob/v0.5.7/CHANGELOG.md · read 2026-11-07
- [24] server/src/integrations/connections/oauth.ts (v0.5.7) · github.com/trina-app/trina-ax/blob/v0.5.7/server/src/integrations/connections/oauth.ts · read 2026-11-07