Gazebo
    ServicesAgentsVaultsDocsWritingPricing
    Log inSign up
    Log in
    GazeboWritingWindsurf and Devin Credentials: MCP Setup and Trust Boundaries

    Windsurf and Devin Credentials: MCP Setup and Trust Boundaries

    Separate Devin-hosted OAuth from local Cascade configuration, protect authentication material, and verify credential retrieval at the correct boundary.

    August 2, 2026·5 min

    Quick answer: Windsurf/Cascade configuration and Devin-hosted OAuth are distinct integration paths. Use the guide for your actual client, protect its authentication material, and expect an authorised Gazebo retrieval to deliver the provider key to the executing runtime and its host.

    Identify the product and execution location

    “Windsurf setup” can refer to different client versions and execution surfaces. Current vendor documentation describes Cascade configuration within the Devin ecosystem, while older installations may use different files and settings.

    Gazebo's Devin / Windsurf guide separates the tested hosted Devin OAuth flow from manual-token guidance for older clients. A result from one path is not a test of every local or legacy version.

    Start by establishing:

    • Which app and version you are using.
    • Whether tool calls execute locally or through a hosted agent.
    • Whether the connection uses OAuth or a manually issued agent token.
    • Which other integrations and credentials are available in the same context.

    Hosted Devin: use the maintained OAuth guide

    For the documented hosted path, add a custom HTTP MCP connection to:

    https://app.gazebohq.com/api/mcp
    

    Use OAuth and follow the current guide's personal-authentication and execution-location settings. Authenticate in Gazebo, explicitly select the service keys needed and choose an access deadline. Do not supply a manual agent-token header for this OAuth path.

    The hosted agent receives returned credentials. Personal authentication is not proof that no other organisation member can interact with a session; review the host's session-access controls.

    Gazebo has a recorded Devin-hosted OAuth test covering discovery, identity, named fictional-key retrieval, retrieval history and rejection after revocation. The compatibility guide records the environment and evidence limits.

    Local Cascade: follow the configuration your client opens

    Current Cascade MCP documentation describes remote HTTP/OAuth support and an Open MCP config file action. Use that action for your installed version rather than invent a repository filename and assume it will be loaded.

    The current documentation distinguishes the Cortex-managed configuration from editor discovery files. Older Windsurf paths may therefore not describe the active configuration in a newer installation.

    For a manual-token connection, create an agent at Gazebo Agents. Configure the hosted endpoint and Authorization: Bearer header using that client's supported fields. Protect the agent token as a secret.

    Current vendor examples support ${env:NAME} interpolation in relevant fields. That is documentation support, not a completed Gazebo runtime test of a legacy installation. Ensure the application actually inherits the referenced value and do not assume another client's interpolation syntax works.

    Share a secret-free template or installation instructions. Do not commit the resolved token, leave old keys commented out or describe the token as harmless because it is not a Stripe key.

    Verify identity, retrieval and revocation separately

    Call get_identity before retrieving any credential. Inspect the identity and policies, including approved key names for OAuth.

    Use a fictional test value first. Substitute the service you configured in this illustrative call:

    get_credential({ service: "your-test-service", key_name: "api_key", method: "GET" })
    

    OAuth requires the approved key name. Ask the client not to print the value and inspect the corresponding event with list_audit_events. A tools-discovery panel or successful connection check alone does not prove that tool execution worked.

    Revoke the test grant or agent, then verify a new retrieval fails without reconnecting. This stops future retrieval; it does not revoke a key already returned.

    Understand what a policy actually restricts

    A policy can determine which configured credential the identity may retrieve and under which declared method. It does not turn a broad provider key into a payment-intent-only or read-only credential.

    Use restricted provider keys where available. Apply human approval or other execution controls at the tool that performs sensitive operations. Those are separate controls, not features created merely by connecting Gazebo.

    Returned keys can enter tool output, model context, session storage or host infrastructure depending on the client. Removing a preloaded key can reduce unnecessary availability, but runtime retrieval is still delivery of a sensitive value.

    Keep identity and project separation meaningful

    Use separate retrieval identities where different projects, consumers or trust levels need genuinely different access. A name such as billing-staging is useful for review; it does not prove a particular human or unique session performed a later request.

    A prompt selecting one of several configured MCP entries is not isolation. Disable unrelated tools or use appropriate runtime separation when the distinction matters.

    Shared onboarding documents should contain configuration structure, not shared authentication material. Do not assume named profiles imply a particular Gazebo organisation-admin or invitation workflow.

    Investigate with more than retrieval history

    Gazebo records credential-access attempts, not every provider operation or API response. Correlate relevant host/runtime and provider records when investigating unexpected activity. One retrieval can precede many operations, including from another process holding a copied key.

    If a credential may have been exposed, assess provider-side invalidation as well as connection revocation. See MCP configuration security and rotation versus revocation.


    Last reviewed: 8 November 2026. Independent editorial; not affiliated with or endorsed by Windsurf or Devin. Use the documentation for your installed product and execution surface.

    Frequently asked questions

    How does Windsurf's Cascade agent access API credentials?

    Access depends on the installed client, execution host, permitted tools and available environment or configuration values. Hosted Devin and local or legacy Cascade are different paths. A Gazebo retrieval delivers the actual provider key to the executing runtime and its host.

    What is the risk of giving Windsurf full API key access?

    A broad key can permit harmful actions or disclosure if the runtime misuses it. Provider keys can also be restricted; they are not inherently full-account. Gazebo retrieval policies do not narrow the returned key's provider permissions or prevent later misuse.

    Does Windsurf support MCP for external service connections?

    Current vendor documentation describes Cascade remote MCP and OAuth support. Use the configuration opened by your actual client. Gazebo's tested Devin-hosted OAuth path is separate from local or legacy Cascade configuration; consult the maintained guide and compatibility record.

    How do you revoke Windsurf agent access without rotating API keys?

    Revoking the Gazebo grant or agent stops fresh retrieval through that identity. A key already returned can remain usable within its provider permissions. Restrict, revoke or expire it at the provider if needed, and account for other consumers of that key.

    What is prompt injection in the context of Windsurf?

    Prompt injection is when a malicious string in content Cascade reads — a code comment, a README, an API response — redirects what it does next. If Cascade has full-access credentials when this happens, the attack can touch everything those credentials cover.

    Give your agents the access they need

    Scoped credential retrieval, access logs, and revocable agent access — for your AI tools.

    Get started free

    Agent pages

    WindsurfCursorClaude Code

    Service pages

    StripeGitHubVercel

    Related reading

    MCP Config Security: Protect API Keys and Bearer TokensReplit Agent Credentials: Secrets, MCP and Trust Boundaries1Password and Gazebo for AI Agent Credentials
    ← Back to writing
    Gazebo

    Credential management for AI agents. Control credential retrieval, review access logs, and revoke an agent’s future retrieval without rotating shared API keys.

    Gazebo

    • Account
    • Demo
    • Pricing

    Explore

    • Services
    • Agents
    • Vaults

    Developers

    • Docs
    • SDK
    • CLI
    • API
    • MCP

    Content

    • Writing
    • Topics
    • Blog

    Company

    • About
    • Contact
    • hello@gazebohq.com

    © 2026 Gazebo. All rights reserved.

    PrivacyTermsSecurity