Skip to main content
After the September and October 2026 Microsoft 365 updates on Windows, some users are missing their chat history and uploaded skills in the Claude add-in. Anthropic is working with Microsoft to solve this issue. In the meantime, see what you and your IT team can do.
Organizations using Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry, or an LLM gateway can deploy Claude’s Office add-ins without requiring individual Claude accounts. The add-in connects through your organization’s infrastructure, sends prompts to the provider your organization chooses, and receives Claude’s responses from it. See What Anthropic collects for what reaches Anthropic on Amazon Bedrock, Google Cloud Vertex AI, and Azure AI Foundry.

Connection paths

Four connection paths are available. Your IT admin selects one during deployment. End users see the same interface regardless.

Requirements by connection path

All paths need:
  • Claude for Excel, PowerPoint, Word, or Outlook installed from Microsoft AppSource or via admin deployment.
  • Microsoft 365 with Entra ID for admin consent and token issuance.
  • For Outlook: Microsoft Graph admin consent for Mail.ReadWrite, Calendars.Read, User.Read, and offline_access, granted via Anthropic’s app or your own Entra app registration.
Your organization’s IT team manages these resources. Anthropic cannot provide or reset credentials.
Bedrock direct does not support AWS GovCloud (US) regions. An alternative is an LLM gateway that your organization operates. Anthropic has not tested a gateway with Amazon Bedrock in GovCloud.

Network allowlist

The add-in requires access to specific domains. The required domains differ depending on whether your organization uses the Anthropic API directly (1P) or a third-party platform (3P).
In all configurations, prompts and responses travel only to your chosen inference provider. Domains pointing to Anthropic (such as pivot.claude.ai) serve the add-in’s interface, feature configuration, and operational telemetry, not prompt or response content.Dictation is an exception. It is available only when users sign in with a Claude account and inference goes to api.anthropic.com. The add-in sends the user’s Claude access token and streams the user’s audio over a WebSocket connection to pivot.claude.ai, which forwards them to Anthropic’s transcription service and returns the transcribed text.

Anthropic API (1P)

Use this table if your organization signs in with Claude accounts and inference goes to api.anthropic.com. If your organization has IP allowlisting enabled for Claude, route bridge.claudeusercontent.com through the same proxy egress as claude.ai and api.anthropic.com, for example by placing it in the same Zscaler app segment or Netskope steering policy. If you cannot route it that way, add the egress address your proxy uses for that domain to your organization’s Claude IP allowlist, but only when that address is dedicated to your organization: a shared proxy egress range also admits the proxy vendor’s other customers. Anthropic checks connections to bridge.claudeusercontent.com against your organization’s Claude IP allowlist using the address they arrive from. If your proxy sends traffic for that domain out through an address that is not on that allowlist, work across apps stops while the rest of the add-in keeps working.

Third-party platforms (3P)

Use this table if your organization signs in with Microsoft Entra ID and inference goes to your LLM gateway, Bedrock, Vertex AI, or Azure AI Foundry.

Deploy the add-in for your organization

Use the claude-for-msft-365-install plugin to configure and deploy the add-in across your organization. The plugin provisions cloud resources (for Bedrock or Vertex AI direct), generates the add-in manifest, and obtains admin consent in a single guided flow.

Run the setup wizard

Install the plugin from the financial services marketplace, then run the setup wizard from inside Claude. Add the marketplace in your shell:
Install the plugin:
Keep the plugin current before each deployment. List installed plugins with claude plugin list and compare your version against the latest published version. If yours is older, update it:
Then, from inside Claude, run the setup wizard:
The wizard walks you through the path you chose:
  • LLM gateway: collects the gateway URL and token, determines the API format, generates the manifest, handles Azure admin consent.
  • Bedrock direct: creates the IAM OIDC identity provider and role, generates the manifest, handles Azure admin consent.
  • Vertex AI direct: walks through Google OAuth client creation, generates the manifest, handles Azure admin consent.
  • Foundry direct: captures azure_resource_name and azure_api_key, then generates the manifest. For keyless Entra ID sign-in, add the parameters described in Foundry direct without an API key to the generated manifest.
When complete, the add-in is ready for tenant-wide deployment.
Bedrock and Vertex AI paths require Node.js for manifest generation and validation. The wizard checks for it and prompts installation if missing.

Available commands

The plugin exposes the following slash commands once installed. Run /claude-for-msft-365-install:debug whenever a connection or sign-in does not behave as expected. It triages from the symptom, reads the “Copy error details” paste from the connection-failed screen, and explains how each connection path works, so you can resolve most third-party platform questions without escalating.

What the wizard provisions

The setup wizard creates resources in your cloud account based on the connection path you choose.

Per-user configuration

If gateway, Bedrock, or Vertex AI connection values vary per user, such as different gateway tokens or AWS roles for different teams, run /claude-for-msft-365-install:update-user-attrs after initial setup to write them as Microsoft Graph extension attributes. Set entra_sso=1 in the manifest so that the add-in reads them at sign-in, where they arrive as claims in the user’s ID token. By default, the add-in signs in through Anthropic’s Microsoft Entra application, which passes only these eight keys to the add-in:
  • LLM gateway: gateway_url and gateway_token
  • Bedrock direct: aws_role_arn and aws_region
  • Vertex AI direct: gcp_project_id, gcp_region, google_client_id, and google_client_secret
Deliver every other key, such as the Foundry direct keys or disabled_features, through the manifest. To vary one per user, return it from a bootstrap endpoint instead. If the add-in uses your own Entra app registration through graph_client_id, you choose which keys it receives, and entra_sso=1 is still required. Configure each one on that application as a directory extension attribute with an optional ID token claim, which reaches the add-in as extn.<key>, such as extn.otlp_endpoint. An attribute value holds at most 256 characters, so deliver a longer value, such as most access_policies values, through the manifest or a bootstrap endpoint. With either application, neither a claim nor a bootstrap response can set entra_sso, graph_client_id, graph_cloud, entra_scope, or gateway_auth_source, so keep them in the manifest. With Anthropic’s application, bootstrap_url stays in the manifest too. At load, the add-in resolves configuration from three sources in order of precedence: a bootstrap endpoint, Microsoft Entra ID extension attributes, then manifest parameters. Per-user values override the manifest defaults, so one deployed manifest can serve teams with different settings.
The add-in resolves configuration from a bootstrap endpoint, then Entra ID extension attributes, then manifest parameters.

Configuration resolution at add-in load: bootstrap, Entra ID attributes, then manifest parameters. With Anthropic's Microsoft Entra application, Entra ID attributes carry only eight connection keys, not every configuration key.

Keep credentials out of the manifest URL query string

Prefer a bootstrap endpoint (bootstrap_url) or Entra ID sign-in with gateway_auth_source=entra, so no credential appears in the manifest or in any URL. If you keep URL configuration, put gateway_token, azure_api_key, google_client_secret, otlp_headers, inference_headers, and mcp_servers after #, never after ?: the browser sends the query string to the server on every load, where request logs record it, while the fragment stays in the browser. The add-in warns in the console when one of these keys arrives in the query string and ignores it there from October 19, 2026. The following example shows the change. Apply it to both SourceLocation and the Taskpane.Url resource, and keep & written as &amp; in the XML.
Before you redeploy the manifest from the Microsoft 365 admin center, raise the number in its <Version> element, for example from 1.0.0.1 to 1.0.0.2. Microsoft requires a higher version number for every manifest change, and without one, Office can keep using the old manifest.

Admin feature controls

The disabled_features configuration key turns off individual add-in features for your users. Set it as a comma-separated manifest parameter to apply one list org-wide, or return it from a bootstrap endpoint as a JSON array to vary it per user. Unknown slugs are ignored, so setting a slug from a newer add-in version on an older deployment is safe. Disabling web_search pairs with the mcp_servers key: attach your own search tool from a server inside your network, and with the built-in search disabled the model uses the tool you provide. This keeps search queries on infrastructure you control. For document-scoped control, such as disabling a feature only on files carrying certain sensitivity labels, use the access_policies key instead; a statement without a resource behaves exactly like disabled_features. The setup wizard’s /claude-for-msft-365-install:manifest and /claude-for-msft-365-install:access-policies commands document both keys in full.

Skill limits for a bootstrap endpoint

A bootstrap endpoint can return a skills array of admin-provisioned skills. The add-in applies these limits to each response. If a skill fails to load, the add-in doesn’t replace it with a later skill. To see which skills failed to load, select the + button at the bottom of the chat pane, and then select Skills. A note at the top of the list names up to three of those skills and gives the number of any others. The note leaves out a skill that still has a copy from an earlier load. If a bootstrap call takes more than 10 seconds when the add-in loads with an existing connection, the add-in has no admin-provisioned skills until a later bootstrap call succeeds. To stay within these limits:
  • Put the most important skills first.
  • Return a different list for each app. The X-Claude-User-Agent header on each bootstrap call starts with claude-word, claude-excel, claude-powerpoint, or claude-outlook.
  • Use url instead of inline content for large skills. Inline content is part of the bootstrap response, so it counts toward the 10 seconds. The add-in downloads skill URLs after the bootstrap call, and only for the skills that it keeps.
If a response holds more than 20 different skill names, the add-in sends the warning log record Bootstrap skill count exceeds cap — truncating to your OpenTelemetry collector at {your_url}/v1/logs. Your collector might not receive this record for the bootstrap call made when the add-in sets up a new connection. On that record, count is the number of different skill names in the response, and cap is the limit, 20.

Publish a new or changed skill

The add-in loads the skills array when the pane loads and at each scheduled bootstrap call before the response expires. It also calls the bootstrap endpoint when a user does one of these:
  • Selects New chat.
  • Starts the first message of a chat, by typing it or by selecting a suggested prompt.
  • Opens a list of skills in any chat: types / at the start of a word in the chat input, or selects the + button at the bottom of the chat pane and then selects Skills.
At New chat and at a first message, that extra call happens at most once every 30 seconds in each open pane, counted from the last time the pane read the skills. When a user opens a list of skills, the add-in always calls, unless an earlier call still waits for its answer or the skills are still loading. So a pane has one of these calls open at a time, and it ends a call that has no answer after 10 seconds. Nothing else limits how many of these calls a pane makes in a minute. A / typed while Claude is replying calls when the reply ends. The add-in retries a call up to three times if the endpoint can’t be reached or answers with a 5xx, 408, 425, or 429 status. Skill file downloads aren’t retried. From the response, the add-in uses only the skills array. It doesn’t apply a token or any other setting, so keep tokens from earlier calls valid until they expire. After that call, the add-in loads the array again if an entry might have changed. An added, removed, or reordered entry also counts as a change, and so does an entry that failed to load the last time. When the add-in loads the array again, it downloads the url of each skill within the limit of 20, unless the entry has inline content. It does the same when the pane loads and at each scheduled bootstrap call. To avoid downloads on that extra call when nothing changed, give each url entry a sha256: the SHA-256 of the file, as 64 hexadecimal characters. Update the sha256 whenever the file changes: the add-in refuses a file that doesn’t match it. The load runs in the background and a chat never waits for it:
  • A message sent before the load ends goes without the new skill. A message sent after the load ends has it.
  • If the call fails or takes more than 10 seconds, the add-in keeps the skills that it has.
  • A skill that fails to load keeps the copy the pane already had, until a later load of that skill succeeds. The copy is held only in memory: if the pane reloads while the file still fails, the skill is missing until it loads. The add-in sends the warning log record Bootstrap skill failed to load to your OpenTelemetry collector.
  • A list of skills that is open when the load ends shows the new skill.
To remove every skill, return an empty skills array. If a response has no skills key, the add-in keeps the skills that it has. Serve the bootstrap response with Cache-Control: no-store, so that no cache between the add-in and the endpoint answers with an older copy. A response with Last-Modified and no Cache-Control can be stored too.

Deploy to Outlook

Outlook requires a separate manifest file from Excel, PowerPoint, and Word. Microsoft uses a different add-in schema for mail applications, so the two cannot be combined into one file. When you tell the setup wizard you are deploying to Outlook, it generates a second file named manifest-outlook.xml alongside manifest.xml. Upload each file as its own custom app in the steps below. Claude for Outlook reads mail and calendar data through Microsoft Graph, which requires a one-time tenant-wide grant from a Global Administrator or a Cloud Application Administrator, regardless of which platform serves the model. Complete the Microsoft Graph admin consent step before deployment so users are not prompted individually. The Graph token stays in the user’s Outlook client and is never sent to your gateway or to Anthropic. If your organization’s policy does not permit consenting to a third-party multi-tenant application, register your own single-tenant Entra application with the same delegated Graph permissions and provide its client ID to the setup wizard as graph_client_id. See Use your own Entra app instead.
Claude for Outlook on third-party platforms supports Claude Opus 4.7 and later and Claude Sonnet 5 and later. Earlier model generations are not available on the Outlook surface.

Deploy to Microsoft 365

After the wizard generates your manifest files:
1

Upload the manifest

Open the Microsoft 365 Admin Center and go to Settings, Integrated apps, Upload custom apps. Select “Office Add-in” as the app type, then upload the manifest.xml file. If you are deploying Outlook, repeat this step with manifest-outlook.xml as a second custom app.
2

Choose who gets the add-in

If all users share the same configuration, select “Entire organization”. If you wrote per-user attributes, assign to “Specific users/groups” matching exactly who was configured. Others would open the add-in with no configuration.
3

Finish deployment

Accept permissions and finish deployment.
Propagation to users takes up to 24 hours, usually faster. The add-in appears under Tools, Add-ins on Mac or Home, Add-ins on Windows in Excel, PowerPoint, and Word once deployed. In Outlook it appears in the message ribbon when an email is open. Custom manifest deployment is where most issues surface: the add-in does not appear, users see old configuration after an update, or sign-in fails. Run /claude-for-msft-365-install:debug to diagnose these, or to sideload and validate a manifest locally before a tenant-wide upload.
Start with a pilot group to confirm functionality, then widen assignment. You can change assignment later without redeploying.

Manage the add-in from the Enterprise Admin Console

If you manage Claude Desktop from the Enterprise Admin Console, you can manage Claude for Microsoft 365 there too. For the requirements and the steps for admins and members, see Manage Claude for Microsoft 365 from the Enterprise Admin Console.

Connection instructions for end users

LLM gateway

1

Open the add-in

Open Excel, PowerPoint, Word, or Outlook and launch the Claude add-in.
2

Select your connection mode

On the sign-in screen, select “Cloud provider or gateway”. Then choose your connection: Gateway, Vertex, Bedrock, or Azure. Contact your IT team for connection details if you’re unsure which one to select.
3

Enter your credentials

For Gateway, enter the gateway URL (HTTPS base URL of your LLM proxy, for example https://llm-gateway.example.com) and the API token your IT team provided. By default the add-in sends the token in the x-api-key header with every request. If your admin set gateway_auth_header: authorization in the manifest, the add-in sends Authorization: Bearer <token> instead.
4

Connect

The add-in checks the connection by sending a test request to the gateway. On success, you see the main add-in experience.
Your credentials are stored locally in your browser’s localStorage within the add-in’s sandboxed iframe and are not synced to Anthropic’s servers. Because the Office add-in runs in a sandboxed iframe within Microsoft applications, it cannot use your OS keychain the way Claude Code does. Only enter gateway-issued tokens, not raw cloud-provider credentials. In this path, every request travels from the add-in to your gateway, which forwards it to the provider you configured.
Requests flow from the add-in to your LLM gateway, which forwards them to the configured model provider.

LLM gateway request flow: the add-in calls your gateway, which routes to your chosen provider.

Bedrock, Vertex AI, or Foundry direct

1

Open the add-in

Open Excel, PowerPoint, Word, or Outlook and launch the Claude add-in.
2

Authenticate

For Bedrock, sign in with your Microsoft work account. The add-in uses your Entra ID token to assume the AWS role your admin configured, so no separate AWS credentials are needed. For Vertex AI, sign in with the Google account your admin authorized via the Google OAuth client created during setup. For Foundry, the add-in connects automatically if your admin pre-filled the Azure resource name and API key. If your admin enabled keyless sign-in, the add-in uses your Microsoft work account and no key is involved. Otherwise, enter the values your IT team provided and select Connect.
3

Start working

The add-in reads the configuration your admin provisioned and connects to Bedrock, Vertex AI, or Foundry directly.
If you see an error at sign-in, confirm with your IT team that your account is in the group assigned to the add-in. In a direct connection, the add-in authenticates with your identity provider and calls the model provider without an intermediary gateway. The flow differs by provider. Bedrock direct uses your Microsoft Entra ID token to assume an AWS role, then calls Amazon Bedrock.
The add-in uses a Microsoft Entra ID token to assume an AWS role and call Amazon Bedrock directly.

Bedrock direct flow: the add-in exchanges an Entra ID token for an AWS role, then calls Bedrock.

Vertex AI direct authenticates through Google OAuth, then calls Vertex AI.
The add-in authenticates through Google OAuth and calls Google Cloud Vertex AI directly.

Vertex AI direct flow: the add-in authenticates with Google OAuth, then calls Vertex AI.

Foundry direct without an API key

Instead of a shared resource key, each user can authenticate to your Foundry resource with their own Microsoft Entra ID token. The add-in acquires the token through Nested App Authentication inside Office, sends it to <resource>.services.ai.azure.com as Authorization: Bearer, renews it silently before it expires, and re-authenticates once if Azure rejects a token early. No key is stored on the device and no key is embedded in the manifest. This uses the same Entra app registration that Claude Desktop’s in-app Foundry sign-in uses (inferenceFoundryClientId), with the add-in’s redirect URI added. Set up:
  1. In your Entra app registration, add the Azure Cognitive Services delegated permission user_impersonation and grant admin consent. Register the add-in’s redirect URI as described in Use your own Entra app instead.
  2. Grant the users or groups who will sign in the Cognitive Services User role on the Foundry resource.
  3. Put these parameters in the manifest URL (no azure_api_key):
When gateway_auth_source=entra is set, the add-in ignores any azure_api_key it receives: the administrator chose keyless sign-in. Each user sees a one-time Microsoft sign-in prompt if silent sign-in is not available; afterwards the add-in connects automatically.

Route Foundry direct through your own gateway

If your organization reaches Azure AI Foundry through a gateway or proxy it operates, such as Azure API Management, set azure_base_url to that gateway’s base URL, path included. The add-in then sends every Foundry request to <azure_base_url>/v1/messages instead of https://<resource>.services.ai.azure.com/anthropic/v1/messages, with the same credential and headers it would send to the resource: the API key as x-api-key, or each user’s Microsoft Entra ID token as Authorization: Bearer with keyless sign-in. azure_resource_name is still required and names the resource behind the gateway. Use the same value you would set as Claude Code’s ANTHROPIC_FOUNDRY_BASE_URL. The URL must use HTTPS. Because the add-in calls the gateway from the browser, the gateway must meet the CORS requirements, and your network must allow its domain as listed in the network allowlist. On this path the add-in also sends the anthropic-beta header, and with keyless sign-in the authorization header, so the gateway’s CORS preflight must allow both and the gateway must forward anthropic-beta to Foundry. Users who connected before you set azure_base_url keep their direct connection until they log out and sign in again, as described in Change or update your gateway connection. Until the setup wizard accepts this key, add azure_base_url to the generated manifest URL by hand, or deliver it per user through a bootstrap endpoint. The following manifest parameters configure this path.

Route Vertex AI direct through your own gateway

If your organization reaches Vertex AI through a gateway or proxy it operates, set gcp_base_url to that gateway’s base URL, API version included. Each user still signs in with Google, and the add-in keeps the Vertex AI request format, but it sends every request to <gcp_base_url>/projects/<project>/locations/<region>/publishers/anthropic/models/<model>:rawPredict, or :streamRawPredict for streaming responses, instead of https://<region>-aiplatform.googleapis.com/v1/..., with the user’s Google access token as Authorization: Bearer. Google sign-in and token refresh still go to accounts.google.com and oauth2.googleapis.com. gcp_project_id, google_client_id, and google_client_secret are still required, and gcp_region still names the region in the request path. Use the same value you would set as Claude Code’s ANTHROPIC_VERTEX_BASE_URL. The URL must use HTTPS. Because the add-in calls the gateway from the browser, the gateway must meet the CORS requirements, and your network must allow its domain as listed in the network allowlist. On this path the add-in sends the same headers it sends to Vertex AI: authorization, content-type, accept, and user-agent. The token the gateway receives is the user’s Google OAuth token with the cloud-platform scope, so the gateway should forward it to Vertex AI and not store it. Users who connected before you set gcp_base_url keep their direct connection until they log out and sign in again, as described in Change or update your gateway connection. Until the setup wizard accepts this key, add gcp_base_url to the generated manifest URL by hand, or deliver it per user through a bootstrap endpoint.

Change or update your gateway connection

If your gateway API token expires or your IT team provides a new URL, select Logout from the account menu and sign in again with your new credentials. This applies to LLM gateway, Bedrock, Vertex AI, and Foundry direct connections alike. Logging out deletes saved Claude for Outlook conversations, as described in Data storage and retention.

Gateway requirements for IT teams

The Office add-ins support the same three API formats as Claude Code. Set gateway_api_format in your add-in manifest to specify which format your gateway uses.

CORS requirements

The add-in’s taskpane loads from https://pivot.claude.ai. Every request to your gateway is cross-origin, and the browser silently discards responses lacking CORS headers. Your gateway must return Access-Control-Allow-Origin: https://pivot.claude.ai (or *) on every response: GET, POST, OPTIONS, and all error responses. Setting it only on the OPTIONS preflight is insufficient. For the preflight, return Access-Control-Allow-Headers listing the request headers the add-in sends, such as x-api-key, authorization, content-type, anthropic-version. The * wildcard does not cover the Authorization header per the Fetch specification, so list it explicitly if you set gateway_auth_header: authorization.

Required endpoints

The endpoints your gateway must expose depend on which API format it speaks. gateway_api_format: anthropic (default): gateway_api_format: bedrock: Native Bedrock InvokeModel pass-through. gateway_url must point at the pass-through prefix, for example https://litellm.example.com/bedrock. gateway_api_format: vertex: Native Vertex pass-through. gateway_url must include the API-version segment, for example https://litellm.example.com/vertex_ai/v1. Also requires gcp_project_id and gcp_region so the add-in can build the path.

Required header

For anthropic format, the gateway must forward the anthropic-version request header to the upstream provider. For bedrock and vertex formats, the SDK places anthropic_version in the request body instead. The gateway must preserve it there. Failure to forward the header or preserve the body field may result in reduced functionality or prevent the add-in from working.

Authorization header

The add-in can send your gateway’s authorization token in either the x-api-key header or the Authorization header. The default is x-api-key. To switch to Authorization: Bearer, set gateway_auth_header: authorization in the manifest.

Model discovery

For gateways using gateway_api_format: anthropic, the add-in attempts to discover available Claude models via GET /v1/models on login. If your gateway doesn’t expose a model list at that path, the add-in falls back to prompting the user for a model ID manually. For gateway_api_format: bedrock and gateway_api_format: vertex, the add-in uses a built-in model list and probes the gateway to verify each model is reachable, rather than calling GET /v1/models.

Differences from Claude Code gateway setup

If your team already runs Claude Code through a gateway, the table below summarizes how the Office add-in setup differs. When gateway configuration comes from a bootstrap endpoint, the add-in keeps the token current without user action. It calls the bootstrap endpoint again about five minutes before the expiry declared in the bootstrap response and applies the returned token to the running session. If that call fails, the add-in tries again after about a minute, then at growing intervals of up to 90 minutes, until a call succeeds. Meanwhile, it also calls the endpoint before each message that a user sends, at most once every 30 seconds. If the gateway rejects a request as unauthorized before that expiry, the add-in calls the bootstrap endpoint once and retries the request if the token changed. Gateway tokens entered manually in the add-in UI do not refresh automatically: sign out and sign in again with the new token when it rotates.

Example gateway configuration with LiteLLM

LiteLLM PyPI versions 1.82.7 and 1.82.8 contained credential-stealing malware. Do not install those versions. If already installed, remove the package, rotate all credentials on affected systems, and follow remediation steps in BerriAI/litellm#24518.
LiteLLM is a third-party proxy service. Anthropic does not endorse, maintain, or audit LiteLLM’s security or functionality. This section is informational and may become outdated. Use at your own discretion. The example configurations below route Office add-in requests through LiteLLM to Anthropic, Bedrock, Vertex AI, or Azure.

Route to Anthropic directly

Use this config.yaml to point the gateway at the Anthropic API.

Route to Amazon Bedrock

Use this config.yaml to route requests through Amazon Bedrock.

Route to Google Cloud Vertex AI

Use this config.yaml to route requests through Vertex AI.

Route to Azure

Use this config.yaml to route requests through Azure AI Foundry.
For detailed setup instructions, see LiteLLM’s Anthropic format documentation.

What Anthropic collects

Even when inference goes through your own infrastructure, the add-in communicates with pivot.claude.ai to load its interface and with claude.ai/api/ to evaluate feature flags. These connections transmit operational telemetry such as which features are used, performance timings, and error rates, so Anthropic can maintain and improve the add-in experience. They do not transmit your prompts or Claude’s responses. For members of an organization that manages the add-in from the Enterprise Admin Console, they also carry account details, as What leaves the device describes. When the add-in uses Amazon Bedrock or Google Cloud Vertex AI, directly or through a gateway, Anthropic collects information in accordance with that platform’s terms, consistent with Anthropic’s arrangements with customers. Anthropic does not have access to a customer’s AWS or Google instance, including prompts or outputs it contains. Anthropic does not train generative models with such content or use it for other purposes. Anthropic can access metadata such as tool use and token counts, and uses such metadata for analytic and product-improvement purposes. When the add-in uses Azure AI Foundry, directly or through a gateway, Anthropic operates the Claude models as an independent processor for Microsoft, and Anthropic’s data use terms apply. For deployments hosted on Azure, Anthropic processes prompts and completions on Azure infrastructure and stores retained content there. Retained content includes content flagged by Anthropic’s safety systems for Anthropic to review. Operational data that contains no prompts or completions is stored on Anthropic’s systems outside Azure. Examples include safety classifier scores, identifiers, and usage and billing data. For deployments hosted on Anthropic, prompts and responses are processed on Anthropic’s infrastructure. See Claude in Microsoft Foundry for details. For details on what your organization’s gateway or cloud provider logs, contact your IT team. To route a full audit trail, including prompts, tool inputs, tool outputs, and document references, to your own infrastructure, see Configure a custom OpenTelemetry collector. That page covers the otlp_endpoint, otlp_headers, and otlp_attr_max_chars configuration keys, the CORS requirements for the collector endpoint, and the full span reference.

Why sign-in redirects through pivot.claude.ai

During Google sign-in for Vertex AI, MCP connector authorization, or Microsoft admin consent, your identity provider redirects the browser to a page on pivot.claude.ai. Security reviewers sometimes ask whether this means access tokens for your cloud provider, connector, or mailbox pass through Anthropic’s servers. They do not. This section explains what the redirect carries in each flow and why the page cannot obtain a token. Claude account sign-in involves no redirect.

OAuth authorization-code redirects

Google sign-in for Vertex AI uses the OAuth 2.0 authorization-code grant and redirects to pivot.claude.ai/auth/callback. MCP connector authorization uses the same grant with a dedicated pivot.claude.ai/auth/gateway-callback redirect. In each case the URL contains two query parameters:
  • code: a one-time authorization code, not an access token
  • state: a random value the add-in generated before sign-in started
The callback page is a static page served from pivot.claude.ai. It reads those two parameters from the URL, shows a Copy button, and instructs you to paste the value back into the add-in inside Office. The page has no server-side logic that stores, forwards, or exchanges the code. An authorization code on its own cannot be redeemed for an access token. The token endpoint requires an additional secret that only the add-in running on your machine holds:
  • MCP connector authorization: a Proof Key for Code Exchange (PKCE) verifier. The add-in generates a random verifier locally, sends only its SHA-256 hash to the identity provider when sign-in starts, and keeps the verifier in browser session storage. The token endpoint rejects any exchange that does not present the original verifier.
  • Google sign-in for Vertex AI: the client_secret belonging to the Google OAuth client your organization created during setup. This value is provisioned into the add-in’s configuration on each user’s machine and is sent only to oauth2.googleapis.com during token exchange. Anthropic does not have this value.
The add-in also verifies that the state value pasted back matches the one it generated and stored locally before sign-in. A mismatch is rejected. This prevents an attacker from tricking a user into completing a sign-in the attacker initiated. After the add-in exchanges the code, the resulting access and refresh tokens are held in browser storage inside the Office add-in sandbox. The add-in presents the access token only to the inference endpoint or MCP server it was issued for, and presents the refresh token only to the OAuth token endpoint. Each of these endpoints appears in the Network allowlist. These tokens never reach pivot.claude.ai.

Why the redirect cannot target localhost

Office add-ins run inside a sandboxed browser frame hosted by Microsoft 365. There is no local web server to receive a loopback redirect, and the browser tab that handles sign-in is isolated from the add-in frame’s storage. The redirect must therefore target a registered HTTPS URL, and the callback page at pivot.claude.ai bridges the two contexts by displaying the code for you to paste back into the add-in.

Claude account sign-in uses no redirect

Signing in with a Claude account uses the OAuth 2.0 device authorization grant, RFC 8628, instead of a redirect. The add-in requests a one-time sign-in code from claude.ai, shows it in the task pane, and sends you to claude.ai in your browser. It waits while you approve the sign-in there, then obtains the access and refresh tokens directly from the claude.ai token endpoint. During sign-in, no authorization code or token passes through pivot.claude.ai, and there is nothing to paste back into the add-in. The Microsoft Graph consent link in Grant Microsoft Graph consent uses Microsoft’s admin-consent endpoint. By Microsoft’s specification, the redirect back to pivot.claude.ai/auth/callback carries only the consent outcome: an admin_consent boolean and the tenant ID. It never carries an access token or an authorization code. The callback page displays a confirmation message and nothing else. The Microsoft Graph access token itself is obtained separately through Nested App Authentication, where the Office host brokers the token directly into the add-in on the user’s machine. The Microsoft Authentication Library (MSAL) caches it in the browser’s local storage, and the add-in calls graph.microsoft.com directly. The Graph token never reaches pivot.claude.ai or any other Anthropic endpoint.

Identify Anthropic’s Microsoft Entra application

The admin consent link and Nested App Authentication both use a single multi-tenant application that Anthropic publishes in Microsoft Entra ID. When you review the consent prompt or the resulting enterprise application in your tenant, confirm it matches these values. The add-in uses the following redirect URIs with this application. Each one exists for a specific Microsoft sign-in path, and none of them receives a Microsoft access token in the URL.

Verify this in your own environment

You can confirm every claim above with a network capture on a test machine:
  • The redirect to pivot.claude.ai/auth/callback carries code and state, or admin_consent and tenant, in the query string. For an MCP connector the redirect to pivot.claude.ai/auth/gateway-callback carries code and state. No access_token parameter appears.
  • The POST that exchanges the code goes to oauth2.googleapis.com for Vertex AI or to the connector’s authorization server for an MCP connector, originates from the add-in frame, and includes the client_secret or code_verifier that never appeared in any request to pivot.claude.ai.
  • A Claude account sign-in produces requests from the add-in frame to claude.ai/v1/oauth/device_authorization and claude.ai/v1/oauth/token, and no request to pivot.claude.ai/auth/callback.
  • Microsoft Graph calls go directly to graph.microsoft.com with a bearer token that was issued by login.microsoftonline.com and never transited an Anthropic domain.

Differences from signing in with a Claude account

When you sign in with a Claude account, the add-ins connect directly to Anthropic, unless you sign in to an organization that manages them from the Enterprise Admin Console. When you connect through a third-party platform, the add-ins send inference requests to your organization’s infrastructure instead, and your IT team controls how that traffic is routed and logged. Some features work differently through third-party platforms, and others aren’t available. The third-party platform column also applies to members of an organization that manages the add-in from the Enterprise Admin Console. On a gateway using gateway_api_format: anthropic, the add-in tells where requests go from the id in the gateway’s response, so the gateway must pass the provider’s response ID through unchanged. If your team needs a feature that isn’t available on your path, talk to your Claude admin about which sign-in path fits your organization.

Troubleshooting

”Connection refused” or network error

The gateway URL or cloud endpoint is unreachable from the user’s network. Verify the URL is correct, the service is running, and there are no firewall or VPN restrictions blocking the connection. Check the Network allowlist to confirm all required domains are allowed.

401 Unauthorized or “Invalid token”

The auth token is invalid or expired. For gateway connections, confirm the token with your IT team. For direct-cloud connections, verify the user’s Entra ID account is in the assigned group and that the OIDC trust or OAuth client is configured correctly. For Foundry with an API key, regenerate the key in Azure Portal, Keys and Endpoint. For keyless Foundry sign-in, confirm the Entra app has the Azure Cognitive Services user_impersonation permission with admin consent and that entra_scope is https://cognitiveservices.azure.com/.default.

403 Forbidden or “Access denied”

The token is valid but lacks the right permissions. For Bedrock, verify the IAM role has bedrock:InvokeModel permissions. For Vertex, verify your Google account has the Vertex AI User role on the project. For gateways, check the token’s scope with your IT admin. For Foundry, check the resource’s networking rules, or confirm the key belongs to the right resource. For keyless Foundry sign-in, confirm the user holds the Cognitive Services User role on the resource.

404 Not found

The add-in could not reach the expected API path. For gateways, verify the URL is the base URL such as https://litellm.example.com:4000. Don’t include /v1/messages in the URL field.

500 or other server errors

The gateway or cloud provider encountered an internal error. Check your gateway logs, such as docker logs litellm for LiteLLM, for upstream provider errors. Try the request again, and contact your IT admin if the issue persists.

”No models available”

The add-in could not find Claude models. For gateways using gateway_api_format: anthropic, your gateway may not expose a model list at GET /v1/models; your IT team can configure the gateway to serve a model list or give you a specific model ID to enter manually. For gateways using gateway_api_format: bedrock or vertex, none of the built-in models responded to the add-in’s probe; confirm with your IT team that the gateway routes to a region or project with Claude models enabled. For Bedrock or Vertex direct, confirm that at least one Claude model (Claude Sonnet 4.5 or later) is enabled in your account and region. For Foundry, confirm at least one Claude model is deployed in the resource Model catalog.

Streaming responses fail or hang

Verify that your gateway supports Server-Sent Events (SSE) pass-through. Some proxy configurations strip or buffer SSE connections, which prevents streaming responses from reaching the add-in.

A feature I expected is not available

Dictation and working across apps aren’t available through third-party platforms, and connectors are limited to the MCP servers your admin configures. If you need these, ask your admin about signing in with a Claude account instead. Web search and code execution depend on where your gateway or cloud provider sends requests; see Differences from signing in with a Claude account.