Appearance
Move Claude Enterprise connectors to Corveil
A Claude Enterprise connector links Claude to an outside service through a remote MCP server. An Owner adds it for the organization under Organization settings → Connectors, then sets each of its tools to Always allow, Needs approval or Blocked under Customize → Connectors → Tool permissions, a setting that applies to everyone. Each person then signs in to the service with their own account, unless the connector uses a shared credential (Anthropic: Use connectors).
When your organization sends Claude Desktop and Claude Code through Corveil, those clients mostly stop loading the connectors from your claude.ai organization (which ones). This guide re-creates each connector as an MCP server that Corveil governs. Admins grant who may call which tools, and Corveil checks and logs every call.
Corveil has no import. For each connector you:
- Register the server under Connections → Tools, with the credential it needs.
- Grant who may call it, tool by tool, on its Permissions tab.
- Point Claude Desktop and Claude Code at its Corveil URL.
The governed-server model itself (credentials, grants, argument rules, rate limits, audit) is on MCP Gateway.
Before you move
Which clients lose the connectors
| Client | After the move |
|---|---|
| Claude Code with a Corveil virtual key (direct mode) or on Amazon Bedrock | Doesn't load claude.ai connectors. Claude Code loads them only when a claude.ai sign-in is the active credential, not when ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, apiKeyHelper or Bedrock supplies it (Anthropic). |
| Claude Code in passthrough mode on a Claude subscription sign-in | Still loads them, and their calls go through Anthropic, not Corveil. To turn them off, set "disableClaudeAiConnectors": true in any Claude Code settings scope, such as managed settings. It covers the connectors Claude Code fetches itself in the terminal, IDE extensions and Agent SDK, not the ones the Claude Desktop app delivers to its Code sessions (Anthropic). |
| Claude Desktop in gateway mode | Its MCP servers come from the managedMcpServers list in the managed configuration, not from your claude.ai organization's connectors (Anthropic). |
| claude.ai on the web and mobile | Keeps the connectors you leave in Claude Enterprise. Servers you register in Corveil don't appear there. |
Removing a connector from Claude Enterprise
Anyone who still uses claude.ai on the web or mobile loses a connector you remove from Claude Enterprise. Keep each one there until those people have moved, or tell them before you remove it.
Where the calls come from
A Claude Enterprise custom connector reaches its MCP server from Anthropic's cloud. A server registered in Corveil is reached from Corveil. If a server sits behind a firewall that admits only Anthropic's IP ranges, Corveil's requests are refused until the firewall admits Corveil as well. Ask your Corveil contact before you move a server like that.
Step 1: List your connectors
In Claude Enterprise, open Organization settings → Connectors for the list of connectors your organization added. For each connector, write down:
- its name, and its remote MCP server URL if it has one;
- how people sign in: each person with their own account, one shared credential, Enterprise-managed auth, or no sign-in;
- each tool's setting: Always allow, Needs approval or Blocked. An Owner sees these under Customize → Connectors, on the connector's Tool permissions;
- which people or teams use it.
Then put each connector in one of the three cases below.
Step 2: Pick the case
| Connector in Claude Enterprise | Examples | What to do in Corveil | Status |
|---|---|---|---|
| A. A vendor's remote MCP server, where each person signs in with their own account | Atlassian, Notion, Linear, Asana | Register the server in per-user OAuth mode. Each member connects their own account. | In preview: ask your Corveil contact. |
| B. A shared credential, or a custom connector that takes a token | Your own MCP server; a vendor server you call with a service account's token | Register the server with the organization's credential, or install its catalog template. | Available |
| C. A connector Anthropic hosts, with no MCP server URL from the vendor | Microsoft 365, Gmail, Google Calendar (Anthropic) | No equivalent in Corveil today. See the options below. | Not available |
To tell B from C: look in the vendor's own documentation for a remote MCP server URL. A connector whose vendor publishes none is case C.
A connector on Claude Enterprise's Enterprise-managed auth (in beta at Anthropic), which an admin authorizes once for the whole organization, is closest to case B. Before you move one, check how the service tells members apart today. With a Corveil shared credential, every call reaches the service as one account, so read the warning under case B.
A. Each person signs in (in preview)
In preview: ask your Corveil contact before you move a connector of this kind.
This is how it works. In Connections → Tools → Add Server, set Upstream auth mode to Per-user OAuth (Dynamic Client Registration) (auth_mode: per_user_dcr in the API). Corveil registers itself with the server's sign-in service, and each member connects their own account, so the service limits each member to what their own account can reach. Members are asked only for the scopes your organization allows: the service's read-only scopes until you choose others on the server's OAuth scopes tab (see Per-user sign-in). Changing them asks everyone already connected to reconnect. The server's User connections tab lists who has connected, and Revoke removes a member's stored sign-in.
Add Server already lists this mode, but connecting an account from Claude Desktop or Claude Code is part of the preview, and a per-user server isn't listed on the unified endpoint yet. Until then you can:
- leave the connector in Claude Enterprise, where people keep using it on claude.ai and its calls don't go through Corveil; or
- register it as case B with one shared account, after reading the warning there.
B. A shared credential or a custom connector
- Open Connections → Tools. Choose Add Server, or Browse Catalog for a template. Templates cover servers such as GitHub, Linear, Slack, Notion and Sentry, and each takes one organization token.
- For Add Server, enter a Slug (it becomes part of the server's URL), a Name, and the connector's remote MCP URL as Upstream URL. Leave Upstream auth mode on Shared (admin-configured credentials), and keep Gateway in Surfaces. Claude Desktop and Claude Code reach the server only through the Gateway surface.
- On the server's Credentials tab, choose Add Credential and store the token as a Bearer Token, an API Key or a Header. Corveil adds it to each request it sends to the server. Callers never see it.
- Grant access on the Permissions tab (Step 3).
- On the Test tab, list the server's tools to confirm Corveil can reach it.
A shared credential gives everyone you grant its full access
With per-person sign-in, the service limits each person to what their own account can reach. With a shared credential, every person you grant acts as that one account and can reach everything it can. Corveil logs which person or key made each call, but the service's own logs show only the shared account.
Use a service account with the narrowest access the work needs, and deny the tools people shouldn't call (Step 3).
C. A connector Anthropic hosts
Corveil has no equivalent today. These are the options, with what each one gives up:
- Keep it in Claude Enterprise and use it from claude.ai. Corveil doesn't see or log those calls.
- Use a built-in connector in Claude Desktop. In gateway mode, a
managedMcpServersentry with"server": "microsoft365"or"server": "github"turns on a connector built into the app (Anthropic). It runs inside the app and calls Microsoft or GitHub directly, so Corveil doesn't see or log those calls either. The built-in GitHub connector also asks before write tools such asmerge_pull_request, whatever itstoolPolicysays. - Use a server the vendor publishes, if one exists, and register it as case A or B.
- Search the content instead of acting on it. Corveil sensors for sources such as Gmail, Google Calendar, Outlook and SharePoint / OneDrive copy that content into your organization's knowledge graph, and Claude searches it through Corveil MCP. A sensor reads with one organization credential, not each person's own account, so that access is a separate decision. It doesn't send mail, create events or edit files.
- Tell your Corveil contact which of these connectors you need.
Step 3: Map tool permissions
Corveil denies by default. Until a server has an Allow rule that covers a person, the server is hidden from them and every call they make is refused. Each rule on the Permissions tab names a Tool (one tool, or all tools), an Action (Allow or Deny), and a Grant Type: Organization, Team, User or API Key. A Deny rule beats an Allow rule.
| Claude Enterprise | Corveil |
|---|---|
| Always allow | An Allow rule for the tool, or for all tools, with Grant Type Organization or a narrower team, user or key. |
| Blocked | A Deny rule for the tool. A tool with no Allow rule that covers the caller is refused too. |
| Needs approval | No counterpart. Corveil's rules are allow or deny only. |
For a Needs approval tool, choose Allow or Deny in Corveil. If a person still has to approve each call, set that on the client. The device enforces it, not Corveil:
Claude Desktop: in the server's
managedMcpServersentry, set the tool toaskintoolPolicy, as in the example below. Every call then needs approval (Anthropic).Claude Code: add an
askrule in managed settings. The rule names the tool asmcp__<server>__<tool>, where<server>is the name the server was added under in Claude Code (Anthropic):json{ "permissions": { "ask": ["mcp__corveil-github__merge_pull_request"] } }
Claude Enterprise applies one setting per tool to the whole organization. Corveil can grant a tool to one team and deny it to another, limit a tool's arguments on the Arguments tab, and cap calls on the Rate Limits tab. See MCP Gateway.
Step 4: Connect Claude Desktop and Claude Code
Corveil serves registered servers at two kinds of URL:
| URL | What it carries | Status |
|---|---|---|
https://corveil.io/mcp/<slug>/mcp | One server | Available |
https://corveil.io/mcp | Corveil's own tools, plus every server you list there, each tool named <slug>_<tool> | In preview: ask your Corveil contact. |
Use your own host in place of corveil.io if you run Corveil yourself. Calls to a server go through its permissions, argument rules and rate limits at either URL, and Corveil logs each one. Each person authenticates with a Corveil virtual API key. They create it under People & Access → API Keys, or an admin issues it to them from People & Access → Users → API keys (in preview: ask your Corveil contact). Give each person their own key, so the User and Team grants that cover them apply and each call is tied to them. A call without a grant is refused. For Claude Desktop, see where each person's key comes from.
One server per URL
Each server's Connect tab shows its URL and a Claude Code command. In Claude Code:
bash
claude mcp add --transport http corveil-github https://corveil.io/mcp/github/mcp \
--header "Authorization: Bearer sk-citadel-your-key-here"Add --scope project to write the server to the repository's .mcp.json. To put servers on every machine, use Claude Code's managed MCP configuration.
The unified endpoint (in preview)
In preview: ask your Corveil contact. One connection carries Corveil's own tools and every server you list there, so people add one entry instead of one per server.
When it's on for your organization, Edit Server shows List on the unified endpoint, and the server's details show a Unified endpoint checklist: Opted in, Active and Gateway surface. The server's tools are listed only when all three pass. In Claude Code:
bash
claude mcp add --transport http corveil https://corveil.io/mcp \
--header "X-Corveil-Key: sk-citadel-your-key-here"The server's tools then appear as, for example, github_create_issue, next to Corveil's ontology_* tools. A Claude Code permission rule for that tool is mcp__corveil__github_create_issue.
Claude Desktop in gateway mode
Add a managedMcpServers entry for each URL to the managed configuration that pins gateway mode. Anthropic's reference says credentials don't go in an entry's headers. A headersHelper script prints them instead (Anthropic):
json
{
"managedMcpServers": [
{
"name": "GitHub",
"transport": "http",
"url": "https://corveil.io/mcp/github/mcp",
"headersHelper": "/usr/local/bin/corveil-mcp-headers",
"toolPolicy": { "merge_pull_request": "ask" }
}
]
}The helper prints a JSON object of headers, such as {"Authorization": "Bearer sk-citadel-…"}, or X-Corveil-Key for the unified endpoint. Claude Desktop runs it with CLAUDE_CODE_MCP_SERVER_NAME set to the entry's name, so one script can serve several entries. Build the entry in the in-app configuration window and export it, as Claude Desktop (MDM) describes, rather than encoding it by hand.
Where each person's key comes from
One managed configuration and one helper path serve the whole fleet, but the helper must print each person's own key. A single key in the helper makes every call look like one caller, so User and Team grants stop matching and Corveil can't tell people apart. Where the helper finds the key depends on how the device's Desktop profile sends model traffic to Corveil:
| Desktop profile | Where the MCP helper gets the person's key |
|---|---|
| Virtual-key helper | The key the inference helper already cached in the user keychain or a 0600 file. The MCP helper reads that cache and reprints it. |
| Passthrough helper | The Corveil key on the second line of the same secrets file. |
| Single sign-on | The device holds no key: the sign-in covers model traffic, not MCP servers. Mint one key per person for this, store it in the user keychain or a 0600 file, and have the helper reprint it. A person can create the key under People & Access → API Keys, or an admin can issue it from People & Access → Users → API keys (in preview: ask your Corveil contact) and choose Show the key to me instead so IT can place it on the device. Fleet tooling that runs the exchange instead sends a hostname of its own for each device, such as claude-desktop-mcp-<machine-hostname>. |
Never mint a key from the MCP helper
The helper only reprints a key that already exists, and exits non-zero when it can't find one. It never runs corveil login or its own /auth/cli/exchange. A repeat exchange with the same hostname deactivates the key Desktop is already using, and corveil login sends the machine hostname, so it replaces the key the person's CLI uses. See the helper's contract on Claude Desktop (MDM).
A helper that reprints a cached key:
sh
#!/bin/sh
# Reprints the person's cached Corveil key. It never mints one.
# Point "cache" at wherever your inference helper or your fleet
# tooling stores the key (mode 0600, owned by the user).
set -eu
cache="${HOME}/.config/corveil/desktop-key"
key=$(awk 'NR==1 { print; exit }' "$cache")
[ -n "$key" ] || exit 1
printf '{"Authorization":"Bearer %s"}\n' "$key"Checklist for each connector
- [ ] The server is under Connections → Tools, Active, with Gateway in Surfaces.
- [ ] Case B: the credential is on the Credentials tab, from a service account with the narrowest access the work needs.
- [ ] Permissions has Allow rules for the people and teams that used the connector, and Deny rules for its Blocked tools.
- [ ] Each Needs approval tool is either allowed or denied in Corveil, with an
askrule on the client where a person must still approve each call. - [ ] The Test tab lists the server's tools.
- [ ] Claude Desktop and Claude Code point at the server's Corveil URL.
- [ ] Each person calls with their own key. On Claude Desktop the MCP helper reprints a key that already exists and never mints one.
- [ ] The connector stays in Claude Enterprise until no one still needs it on claude.ai on the web or mobile.