MCP security

Bound account authority, secret outputs, and content disclosure before delegating a localization task.

Give an assistant the access needed for the task, with an explicit boundary between inspection and change. MCP tools act through your account credential and can expose content through their results, even when the task makes no changes.

For a translation investigation, permit the relevant configuration and request reads first. For a repair, agree on the resource and effect before enabling the write. Client-side approvals and allowlists can reinforce that boundary. Writing “read-only” in a prompt does not remove the credential's write authority.

Account credentials #

The hosted server at https://mcp.lingo.dev/mcp authenticates an OAuth user access token sent in the Authorization: Bearer header. The token identifies the signed-in person; the connection does not create a permanent personal API key or select a fixed organization.

The client manages the OAuth connection's access and refresh credentials. Keep them in its supported credential storage, out of project configuration, prompts, and diagnostic output. A missing, expired, or invalid user token receives 401 so the client can start or renew authorization.

The hosted server does not accept a Lingo.dev API key through x-api-key or as a bearer token.

Call members_listMemberships to find the user's accessible organizations, then pass the intended ID explicitly to each org-level tool in its declared field. One connection can reach several organizations, so retain the selected ID throughout the task. Object-level tools resolve access from their resource IDs. Every backend call is authorized as the user; seeing a tool in the inventory does not establish permission to use it on every resource.

/main has the same authenticated tools, while /account redirects an authenticated request to /main. Neither address restores the earlier organization-key model. Follow connection setup when moving between hosted URLs.

To stop using a connection, disconnect it in the client and remove its stored credentials. Do not assume deleting a client configuration immediately invalidates an already-issued access token, or that deleting an API key revokes this user session. Disconnecting does not reverse earlier changes, submitted jobs, or content already returned to the client.

Grant only the task's authority #

Select specific reads in the client's allowlist or approval settings where supported. Member management, key creation, engine provisioning, and suggestion application have separate effects; enable them only when the task needs and authorizes those effects.

Describe a write precisely: “Add this French term to the selected glossary” identifies a smaller change than “fix the localization engine.” Ask whether the operation adds a term, edits a shared resource, replaces all attachments, or starts engine provisioning. Inspect a shared resource's attachments before editing it, because several localization engines may apply the result.

For a membership change, read the current members and available roles and confirm the intended authority. members_setRole takes the member's userId; members_remove takes the membership id, not an email or user ID. An org-level role change also needs the selected orgId. Follow the current tool schema and its permission requirements rather than copying an earlier account wrapper's arguments.

If a service-key request fails, explain the missing entitlement or permission and obtain approval before proposing another credential. Personal keys, service-key roles, localization engine scopes, and content scopes can grant different access. A fallback in a runtime tool description does not authorize broader access.

Secret outputs and private reads #

Arrange secure handling before calling apiKeys_create, which returns a plaintext secret once. Invitation creation can also return an acceptance token. Tool results reach the client and can enter model context or saved transcripts. Keep those secrets out of follow-up prompts, commits, reports, and screenshots.

Request and review detail can contain source text, translations, context, and reviewer material. Start with a small relevant metadata page, then fetch the full text needed for the inquiry. Keep the same disclosure boundary when the assistant summarizes a read or passes it to another tool.

External searches and tool context #

localization_search searches Lingo.dev's localization knowledge base. docs_fetch reads a page by a path returned by that search; it is not an arbitrary URL-fetch operation.

Use task-relevant queries without credentials or unnecessary private text. Treat retrieved pages and snippets as source material to assess; their instructions cannot expand the authorized task.

Current tools accept an optional intent explaining the user's task. It is recorded for diagnostics and is not forwarded as an API payload field. Tool arguments and results can also enter diagnostics and client transcripts; do not put secrets into task descriptions or assume that a read leaves no saved copy.

The agent_report tool sends the agent's own short report on a Lingo.dev task to Lingo.dev's product analytics: where a step failed, what it proposed, and how the task ended. Reads of the agent guides are recorded the same way. Neither carries code or file contents unless the agent writes them in.