manifest — команда для Claude Code: что делает и как установить
Команда manifest для Claude Code собирает XML-манифест надстройки с настройками облачного провайдера. Для Outlook она выбирает отдельный шаблон, а перед развёртыванием требует согласия администратора Microsoft Graph.
Apache-2.0 ★ 41 441 обновлён 2026-09-21
/plugin marketplace add anthropics/financial-services
/plugin install claude-for-msft-365-install@claude-for-financial-servicesЧто делает manifest
Команда запускает скрипт сборки манифеста: он получает основной XML и добавляет конфигурацию к URL панели задач. Чувствительные параметры попадают после #, остальные передаются как query-параметры. Для Excel, Word и PowerPoint используется общий шаблон, для Outlook — отдельный.
Запуск выполняется через node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" с аргументами для хоста, имени выходного файла и ключей провайдера. Нужные ключи зависят от облачного пути: например, Bedrock требует aws_role_arn и aws_region, а Vertex — идентификаторы проекта и OAuth-клиента Google. Outlook читает почту и календарь через Microsoft Graph, поэтому до развёртывания нужно получить согласие администратора арендатора.
Какие задачи решает
- собрать XML-манифест с параметрами облачного провайдера
- выбрать отдельный шаблон для Outlook
- добавить настройки Entra SSO в конфигурацию надстройки
- закрепить параметры национального облака в манифесте
Примеры запросов агенту
- Собери манифест Outlook для Bedrock с моими ключами провайдера.
- Создай манифест для Excel и включи вход через Entra SSO.
- Собери манифест Outlook для DoD с
graph_cloud=us-gov-dod.
Как установить manifest
1Через маркетплейс плагинов
/plugin marketplace add anthropics/financial-services/plugin install claude-for-msft-365-install@claude-for-financial-services2Вручную: положить файлы
.claude/commands/manifest.mdВызов — /manifest. Для всех проектов — ~/.claude/commands/
Вручную: положить файлы
~/.codex/prompts/manifest.mdВызов — /prompts:manifest. В Codex этот формат помечен устаревшим, замена — скиллы
Вручную: положить файлы
.github/prompts/manifest.prompt.mdВручную: положить файлы
.gemini/commands/manifest.tomlGemini хранит команды в TOML: текст запроса — в поле prompt
Вопросы о manifest
Как установить manifest в Claude Code?
Выполни /plugin marketplace add anthropics/financial-services , затем /plugin install claude-for-msft-365-install@claude-for-financial-services.
Что делает manifest?
Команда собирает XML-манифест надстройки и добавляет к нему конфигурацию облачного подключения. Для Outlook она использует отдельный шаблон.
Что нужно для работы manifest?
Нужен Node.js для запуска скрипта и ключи, требуемые выбранным облачным провайдером. Для Outlook до развёртывания также требуется согласие администратора Microsoft Graph.
Исходный текст
manifest.md · Apache-2.0
---
description: Generate the add-in manifest XML with your cloud config baked in
---
# Generate add-in manifest
The script fetches the canonical manifest and appends your config to the
taskpane URL. Sensitive settings (`gateway_token`, `azure_api_key`,
`google_client_secret`, `otlp_headers`, `inference_headers`, `mcp_servers`) go
after `#`; other settings are query parameters. Regenerate and redeploy
manifests built with earlier versions. Outlook uses a separate template
because Microsoft's `MailApp` schema is distinct from the `TaskPaneApp` schema
Excel/Word/PowerPoint share, so ask which apps they're deploying and generate
one file per host.
| Host arg | Apps | Template |
|---|---|---|
| `office` | Excel, Word, PowerPoint | `pivot.claude.ai/manifest.xml` |
| `outlook` | Outlook (mail + calendar) | `pivot.claude.ai/manifest-outlook-3p.xml` |
## Keys by cloud
Prompt only for the keys their cloud path needs. Don't ask for all of them.
| Cloud | Keys |
|---|---|
| Vertex | `gcp_project_id` `gcp_region` `google_client_id` `google_client_secret` |
| Bedrock | `aws_role_arn` `aws_region` |
| Foundry | `azure_resource_name` `azure_api_key` — or keyless: `azure_resource_name` + Entra (see [Foundry keyless](#use-the-entra-token-for-foundry-keyless)) |
| Gateway | `gateway_url` `gateway_token` `gateway_auth_header` `gateway_api_format` |
| Gateway (`gateway_api_format=vertex`) | also `gcp_project_id` `gcp_region` |
## Outlook — Microsoft Graph
Outlook reads the user's mailbox and calendar via Microsoft Graph, which
requires a one-time tenant-wide admin consent regardless of which cloud serves
the model. Run [consent](consent.md#outlook--microsoft-graph-consent) before
deploying — otherwise every user hits "Need admin approval" on first open.
If their policy forbids consenting to a third-party app, prompt for
`graph_client_id` (their own single-tenant Entra app's client ID with
Mail.ReadWrite, Calendars.Read, People.Read, User.Read, offline_access
delegated permissions and admin consent granted). Otherwise leave it unset and
the add-in uses Anthropic's multi-tenant app.
## Sovereign / national clouds (GCC-High, DoD, 21Vianet)
The add-in auto-detects the tenant's national cloud at sign-in (from the
authority host Office reports) and resolves the matching Graph + Entra
endpoints, so most sovereign tenants need **no cloud config**. The only
required step is bringing your own Entra app via `graph_client_id` —
Anthropic's multi-tenant app exists only in the commercial cloud; see
[entra-app](entra-app.md#gcc-high--dod--21vianet) for the registration steps in
the Azure Government / 21Vianet portals. A GCC-High Outlook manifest needs
nothing beyond the usual keys:
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" outlook manifest-outlook.xml \
<provider keys> entra_sso=1 graph_client_id=<your-app-guid>
```
### When to set `graph_cloud`
The cloud is configured as a single enum value — never a URL. Each value maps
to the fixed Graph + Entra endpoint pair from Microsoft's
[national-cloud deployments](https://learn.microsoft.com/graph/deployments)
inside the add-in.
| Tenant | `graph_cloud` | Notes |
|---|---|---|
| Commercial or GCC | `global` | default; may be omitted |
| GCC High | `us-gov-high` | auto-detected; set explicitly to pin it in the reviewed manifest |
| US Gov DoD | `us-gov-dod` | **always required** — DoD shares an authority host with GCC High, so auto-detect picks GCC High |
| China (21Vianet) | `china` | auto-detected; set explicitly to pin it |
A DoD Outlook manifest:
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" outlook manifest-outlook.xml \
<provider keys> entra_sso=1 graph_client_id=<your-app-guid> graph_cloud=us-gov-dod
```
The build script enforces the same rules the add-in does at load: an
unrecognized value is a hard error, and any non-global `graph_cloud` requires
`graph_client_id` (without one, sign-in fails with an opaque AADSTS700016).
`graph_cloud` also governs the Entra SSO authority for Word/Excel/PowerPoint —
they share the auth path — so include it in the `office` manifest too if you
set it.
**Bedrock / WIF note:** a `.us`-issued idToken has issuer
`https://login.microsoftonline.us/{tenant}/v2.0` — your AWS OIDC identity
provider must be configured with that issuer, not the `.com` one.
## Entra SSO
`entra_sso=1` makes the add-in acquire an Entra ID token at startup. Set it
when your deployment needs the user's Microsoft identity — Bedrock uses it as
the STS web identity, the bootstrap endpoint uses it as Bearer auth, and
per-user attrs ([update-user-attrs](update-user-attrs.md)) ride inside it as
`extn.*` claims.
**Admin consent is a prerequisite.** Without it, every user hits a Microsoft
consent dialog on first open. Run [consent](consent.md) first so
`entra_sso=1` is silent for your users.
If you don't need Entra — static gateway config, Vertex with Google OAuth —
leave it off. Users won't see a Microsoft prompt for a setup that doesn't
involve Microsoft.
**Bring your own Entra app.** By default the token is requested as Anthropic's
multi-tenant app (`c2995f31-…`), so its `aud` claim is that GUID. If your
bootstrap endpoint or token-exchange service requires `aud` to match an app
registered in *your* tenant, set `graph_client_id=<your-app-guid>`. See
[entra-app](entra-app.md) for the registration steps (redirect URIs, API setup,
admin consent). [consent](consent.md) covers Anthropic's default app only.
**Send an access token instead of the ID token.** With `graph_client_id` alone
the add-in still sends an *ID token* to your bootstrap endpoint — `aud` is your
app's GUID, but there's no `scp` claim. If your endpoint is a standard OAuth2
protected resource that validates `aud` + `scp`, or an RFC 8693 token-exchange
service, set `entra_scope=api://<your-app-guid>/<scope>` and the add-in
requests an *access token* for that scope instead. The Bearer it sends carries
`aud` = your API's App ID URI and `scp` = the granted scope. In Entra, on your
app registration: **Expose an API** (Application ID URI `api://<guid>`), add a
scope such as `access_as_user`, and grant the same app delegated permission to
it, then grant admin consent for the tenant. In the app manifest, set
`accessTokenAcceptedVersion: 2` so the issued token uses v2.0 claims
(`iss = login.microsoftonline.com/<tid>/v2.0`, `azp`, `preferred_username`);
leave it unset and you get v1.0 tokens, which your validator may reject.
`/.default` (requests all consented scopes) also works.
**Multiple scopes.** `entra_scope` accepts a comma- or whitespace-separated
list — `entra_scope=api://<guid>/use_llm,api://<guid>/admin`. All scopes must
target the **same resource**: one access token has one `aud`, so MSAL cannot
mint a token spanning two APIs (`api://torii/x,api://other/y` will fail or
silently honor only one). The Bearer's `scp` claim is the space-joined list.
If you need every consented scope, prefer `/.default` over enumerating them.
`entra_scope` requires `graph_client_id` — the build script enforces *that
pairing* but not the scope string itself: any non-blank value is accepted and
Entra validates the syntax at sign-in (a malformed scope surfaces as an
`AADSTS` error, not a build failure). Both keys are manifest-only: the add-in
needs them to initialize NAA *before* it can read extension attrs or call your
bootstrap endpoint, so neither can arrive through those layers. Leave
`entra_scope` unset and the ID token is sent.
## Use the Entra token as your gateway credential
If your gateway already validates Entra JWTs (`aud` + `scp` against your own
API resource), you don't need a separate `gateway_token` or a bootstrap hop —
set `gateway_auth_source=entra` and the add-in sends the Entra access token it
acquired above directly as `Authorization: Bearer` on every gateway call, and
silently re-acquires it before expiry. The end-user experience is zero-input
SSO: open the add-in, start chatting.
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" office manifest.xml \
gateway_url=https://llm-gateway.your-org.example \
entra_sso=1 \
graph_client_id=<client-app-guid> \
entra_scope=api://<resource-app-guid>/access_as_user \
gateway_auth_source=entra
```
`gateway_auth_source=entra` requires `entra_scope` (and therefore
`graph_client_id` and `entra_sso=1`); the build script enforces this. It
implies `gateway_auth_header=authorization`, so you can omit that key. Don't
also set `gateway_token` — it's ignored, and the script warns.
## Use the Entra token for Foundry (keyless)
Foundry direct normally takes the resource's static `azure_api_key`. To drop the
key and have each user authenticate with their own Entra identity, set
`gateway_auth_source=entra` alongside `azure_resource_name` and leave
`azure_api_key` out. The scope targets Azure's Cognitive Services resource, not
an API you expose:
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" office manifest.xml \
azure_resource_name=<your-foundry-resource> \
entra_sso=1 \
graph_client_id=<your-app-guid> \
entra_scope=https://cognitiveservices.azure.com/.default \
gateway_auth_source=entra
```
The add-in acquires an access token audienced to `cognitiveservices.azure.com`
and sends it as `Authorization: Bearer` on every Foundry call, re-acquiring it
before expiry. Azure validates the token, then checks that the user holds the
**Cognitive Services User** role on the resource — no key is stored anywhere.
The Entra app needs the Azure Cognitive Services `user_impersonation` delegated
permission (not a self-exposed API); see [entra-app](entra-app.md). If both
`azure_api_key` and `gateway_auth_source=entra` are present, the key is ignored.
**Why `cognitiveservices.azure.com`, not `ai.azure.com`.** The add-in signs each
user in through Nested App Authentication against your Entra app, and that
per-user *delegated* flow pairs with the app's Azure Cognitive Services
`user_impersonation` permission, whose audience is
`https://cognitiveservices.azure.com`. This is the same scope Claude Desktop's
Foundry sign-in uses, so an Entra app already registered for Desktop works
unchanged. Microsoft's SDK and managed-identity docs use `https://ai.azure.com/.default`
instead; a Foundry resource accepts both audiences, but that scope belongs to the
credential-based SDK flow (`DefaultAzureCredential`, `az login`, service
principal), which doesn't use an app-registration delegated permission. For the
add-in's per-user sign-in, use `cognitiveservices.azure.com/.default`.
**Entra setup:** see [entra-app](entra-app.md) — the *Gateway / bootstrap auth*
row of the permissions table, plus the
[backend validation](entra-app.md#what-your-backend-validates) section for the
`iss`/`aud`/`scp`/JWKS values your gateway should check. GCC High / DoD
deployments are covered in the
[same doc](entra-app.md#gcc-high--dod--21vianet).
## Bootstrap endpoint
`bootstrap_url` points to an HTTPS endpoint you host. At startup the add-in
fetches per-user JSON from it — provider keys, `mcp_servers`, `skills` — and
the response overrides manifest values for that user. The add-in fetches the URL
exactly as written.
See [bootstrap](bootstrap.md) for the request/response contract, JWT
validation, and handler scaffolding.
## MCP servers
`mcp_servers` is a JSON array of customer-hosted MCP servers the add-in
connects to directly. Each entry is `{url, label, headers?, discover?}` —
`headers` present means static auth; absent triggers OAuth discovery. `url` and
`headers` values interpolate other config keys via `{{gateway_url}}`-style
templates.
Setting it here applies one list org-wide; per-user lists belong in
[bootstrap](bootstrap.md#mcp_servers), which also has the full schema. The
value is JSON inside a shell arg — single-quote it:
```bash
mcp_servers='[{"url":"{{gateway_url}}/deepwiki/mcp","label":"DeepWiki","headers":{"Authorization":"Bearer {{gateway_token}}"}}]'
```
## Telemetry
`otlp_endpoint` routes the add-in's OpenTelemetry traces to a collector you
operate. Set it to the collector's base HTTPS URL — the add-in appends
`/v1/traces` and posts OTLP/HTTP. gRPC isn't supported (the add-in runs in a
browser WebView). Leave it unset and no custom collector is configured.
`otlp_headers` supplies authentication headers for that collector, in the same
`key1=value1,key2=value2` format as the standard
`OTEL_EXPORTER_OTLP_HEADERS` variable. URL-encode the value in the manifest.
`otlp_resource_attributes` adds attributes to the OpenTelemetry Resource on
every span, in the same `key1=value1,key2=value2` format as the standard
`OTEL_RESOURCE_ATTRIBUTES` variable. Use this when your collector requires
specific resource attributes for routing or attribution (e.g.
`team.name=platform,deployment.environment=prod`). The add-in already sets
`service.name`, `service.version`, and `git.sha`; values you provide here are
merged on top.
Setting these here applies one collector org-wide; per-user routing belongs in
[bootstrap](bootstrap.md#telemetry) — Anthropic's Entra app doesn't carry the
`otlp_*` keys as extension attrs (see [update-user-attrs](update-user-attrs.md)).
## Inference headers
`inference_headers` is a JSON object of extra HTTP headers the add-in attaches
to every request it sends to your gateway (`gateway_url`). Use it for
accounting or cost-allocation tags your gateway expects — e.g., an internal
application ID — so you don't need a header-injecting proxy in front of it.
Applies only when using a gateway; direct cloud connections ignore it.
```bash
inference_headers='{"x-application-id":"app123"}'
```
The add-in treats the values as opaque. `Authorization`, `x-api-key`,
`Content-Type`, `Host`, `Content-Length`, `User-Agent`, `Cookie`, and any
`anthropic-*` / `x-amz-*` / `x-goog-*` header are reserved and silently dropped
— they carry the add-in's own auth and protocol negotiation.
Setting it here applies one header set org-wide; per-user values belong in
[bootstrap](bootstrap.md#inference_headers).
## Auto-connect
Default: when all fields for a provider are set, users skip the connection form
and land straight in chat. Ask: should they instead see the form first
(prefilled, one click)? Yes → `auto_connect=0`.
## Allow Claude.ai sign-in
When any enterprise config key is present, users land on the enterprise
connection screen and the **Back** button to Claude.ai sign-in is hidden
(`allow_1p=0`, the default). Set `allow_1p=1` to keep the **Back** button.
## Disabled features
`disabled_features` is a comma-separated list of feature slugs the admin wants
locked for users. Slugs use `<domain>.<action>` form. Currently enforced:
| Slug | Effect |
|---|---|
| `skills.authoring` | Blocks creating, editing, and uploading skills (create/update tools, `/skillify`, `.skill` upload + drag-drop, skill editing UI). Running admin-provisioned skills is unaffected. |
| `thumbs` | Blocks response feedback (thumbs up / down and the follow-up prompt). |
| `addin.access` | Kill switch — the add-in refuses to run. Almost always wants a document `resource`; use [access-policies](access-policies.md). |
| `file.upload` | Blocks attaching files to the conversation. Same — usually scoped by `resource`. |
| `web_search` | Removes the native web_search and web_fetch server tools (queries otherwise egress to Anthropic's third-party search provider) and the user-facing web-search toggle. Code execution is unaffected. Pair with [`mcp_servers`](#mcp-servers) to substitute your own in-network search tool — with native search disabled, the model uses the tool you attach. |
| `permission_mode.accept_all_edits` | Keeps users in the "Ask before edits" permission mode: "Accept all edits" and "Skip all permissions" can't be turned on, and users already in either are moved back. "Always allow" choices for connectors (MCP servers) are unaffected. |
| `permission_mode.dangerously_always_allow` | Removes the "Dangerously always allow" button from approval cards, so users can't switch to "Skip all permissions". Users already in that mode are moved to "Accept all edits", or to "Ask before edits" if `permission_mode.accept_all_edits` is also listed. |
```bash
disabled_features='skills.authoring'
```
Unknown slugs are ignored (forward-compatible). Setting it here applies one
policy org-wide; per-user policy belongs in [bootstrap](bootstrap.md#disabled_features)
(JSON array) — Anthropic's Entra app doesn't carry `disabled_features` as an
extension attr (see [update-user-attrs](update-user-attrs.md)).
## available_models
`available_models` **overrides** the model picker — when set, users see exactly
the models listed, in the listed order, and nothing else. It replaces the
retired `disabled_models` / `additional_models` pair: because it is an
override, it must include every backend model users should keep, not just
additions or removals.
Two formats:
```bash
# comma-separated ids
available_models='claude-opus-4-8,claude-sonnet-5'
# JSON, when picker labels matter — an array of ids and/or {id, label} objects
available_models='[{"id":"claude-opus-4-8","label":"Opus 4.8"},"claude-sonnet-5"]'
```
Unset (or empty) means the picker is unchanged. Setting it here applies
org-wide; per-user policy belongs in [bootstrap](bootstrap.md#available_models).
## access_policies
`access_policies` is the IAM-shaped successor to `disabled_features` — a JSON
array of allow/deny statements that gates add-in features:
- **No `resource`** — the statement applies everywhere, exactly like a
`disabled_features` entry (off for everyone).
- **With a `resource`** — the rule is scoped to matching documents, identified
by Purview sensitivity label or, for the open document, by file path (block
the add-in entirely on top-secret documents or on a restricted network share,
or refuse attaching restricted Office/PDF files).
It covers everything `disabled_features` does, plus conditions, allowlists, and
per-statement attribution. Building one usually means fetching label GUIDs
from Purview, so use the guided command:
[access-policies](access-policies.md). Then pass the result as one more key:
```bash
access_policies='[{"effect":"deny","action":"addin.access","resource":{"type":"open_file","identifiers":[{"type":"mip_label_guid","equals":"<guid>"}]}}]'
```
This belongs in the manifest or bootstrap — it is not one of the eight keys
Anthropic's Entra app carries as extension attrs, and more than one statement
soon outgrows an attr's 256-char cap anyway (see
[update-user-attrs](update-user-attrs.md)).
## Version
M365 Admin Center caches by `<Id>` + `<Version>` — re-upload with the same
version is silently ignored. After the script writes `manifest.xml`, ask whether
this replaces an existing deployment; if yes, edit `<Version>` to bump the
fourth segment past their last deployed value. First deploy can leave the
template's version as-is.
## Run
```bash
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" office manifest.xml \
gcp_project_id=<value> \
gcp_region=<value> \
auto_connect=0 \
...
# and if they're also deploying Outlook:
node "${CLAUDE_PLUGIN_ROOT}/scripts/build-manifest.mjs" outlook manifest-outlook.xml \
<same provider keys as above> \
graph_client_id=<value> # only if NOT using Anthropic's app via the consent URL
```
The script validates key names (unknown keys fail hard) and shape-hints values
(warns but doesn't block — their infra may look different).
## Validate
```bash
npx --yes office-addin-manifest validate manifest.xml
```
If validation passes but M365 Admin Center still rejects or ignores the upload,
match the symptom below. Edit `manifest.xml` directly, then re-validate.
| Symptom | Fix |
|---|---|
| "An add-in with this ID already exists" | Replace the text inside `<Id>` with a fresh UUID. The template carries the marketplace install's ID. |
| Re-upload accepted but nothing changes | M365 caches by ID + version. Edit `<Version>` to a higher fourth segment (e.g. `1.0.0.9` → `1.0.0.10`) and re-validate. |
| Only want Excel (not PowerPoint) | Remove `<Host>` elements for `Presentation`. **Two parallel lists:** the top-level `<Hosts>` uses `Name="Presentation"`, the one under `<VersionOverrides>` uses `xsi:type="Presentation"` — both must go or the manifest is inconsistent. The `xsi:type` block is multi-line, delete the whole `<Host xsi:type="Presentation">...</Host>`. |
| Only want Excel/PPT, not Outlook | Nothing to remove — Outlook is a separate file. Just don't generate it. |
© Anthropic · Apache-2.0 · anthropics/financial-services