generated: '2026-08-13'
method: searched
status: published
source: https://docs.svix.com/ai/app-portal-mcp
sources:
  - https://docs.svix.com/ai/app-portal-mcp
  - https://docs.svix.com/receiving/using-app-portal/connecting-your-coding-agent
  - https://github.com/svix/ai

deployment:
  mode: remote
  endpoint: https://mcp.us.svix.com/app/{app_id}
  auth: api-key
  verified: probed
  note: |
    Verified live on 2026-08-13 by POSTing an anonymous JSON-RPC tools/list to
    https://mcp.us.svix.com/app/app_test. The server answered HTTP 401 with
    `WWW-Authenticate: Bearer realm="svix-mcp"` and the body
    `missing Svix token: pass it as ` + "`Authorization: Bearer <token>`" + `.
    That is a real MCP endpoint refusing an unauthenticated caller, not a 404 or
    an SPA shell. No stdio/npx package is shipped for this server — it is
    remote-only, so an agent can reach it over HTTPS with no local install.

server:
  name: Consumer App Portal MCP
  transport: http
  url_template: https://mcp.{region}.svix.com/app/{app_id}
  regions: [us, eu, ca, au]
  example_url: https://mcp.us.svix.com/app/app_2ErlDgQ1QzKvSAqxdMQnjHNL
  client_name_convention: "<app-portal-slug>-webhooks"
  enabled_by_default: false
  enablement: |
    Off by default for existing accounts. The Svix customer enables it per
    environment in Dashboard -> Settings -> General -> "Enable App Portal MCP".
    Once on, an MCP tab appears under Settings in the Consumer App Portal.

authentication:
  scheme: bearer
  header: 'Authorization: Bearer <token>'
  token_issuer: Consumer App Portal (Settings -> MCP -> Connect to MCP)
  token_scope: single application
  token_ttl_days: 7
  revocable: true
  revocation: MCP tab -> token menu -> Expire (immediate)
  not_an_api_key: |
    Docs are explicit that an MCP token is NOT usable as a Svix API key for the
    organization. It encodes the one application it was issued for and cannot
    read or write anything outside it.
  permissions:
    read:
      - application
      - endpoints
      - event types
      - integrations
      - messages
      - message attempts
    write:
      - endpoint transformations (update)
      - messages (create)
      - message attempts (create — this is what powers resend/recover)

audience: |
  This is a two-sided product and the distinction matters for agent readiness.
  The MCP server is not aimed at the Svix customer operating their own account;
  it is aimed at THAT customer's END USERS — the developers consuming their
  webhooks through the Consumer App Portal. The Svix customer turns it on; the
  consumer generates the token and points Claude Code / Cursor / VSCode / Codex
  at their own application's webhook traffic. It is an embedded agent surface a
  provider ships to its own downstream integrators.

tools:
  - name: get_application
    description: The application this session is scoped to (name, UID, metadata).
    mutating: false
  - name: list_endpoints
    description: List the application's endpoints (URL, enabled state, filtered event types).
    mutating: false
  - name: get_endpoint
    description: Full configuration of one endpoint.
    mutating: false
  - name: get_endpoint_stats
    description: Success / fail / pending / sending counts over a time window.
    mutating: false
  - name: get_transformation
    description: An endpoint's transformation code, enabled state, and variables.
    mutating: false
  - name: update_transformation
    description: Set an endpoint's transformation code and/or toggle it on or off.
    mutating: true
    consequence: changes live endpoint configuration
  - name: list_messages
    description: List messages sent to the application (filter by event type, channel, time).
    mutating: false
  - name: list_attempts_by_endpoint
    description: Delivery attempts for an endpoint (e.g. only failures), with response bodies.
    mutating: false
  - name: list_attempts_by_message
    description: Every endpoint a single message was attempted against, and how each responded.
    mutating: false
  - name: get_message
    description: A message's event type, channels, and JSON payload.
    mutating: false
  - name: get_attempt
    description: One attempt in full, including the response status code and body.
    mutating: false
  - name: resend_message
    description: Resend one message to an endpoint.
    mutating: true
    consequence: performs a real webhook delivery
  - name: recover_endpoint
    description: Replay all failed messages for an endpoint since a given date.
    mutating: true
    consequence: performs real webhook deliveries in bulk

tool_schemas:
  available: false
  reason: |
    The live tools/list is auth-gated (401 without an application-scoped bearer
    token), so the real inputSchema per tool could not be introspected
    anonymously. The 13 tool names and descriptions above are transcribed
    verbatim from the provider's own docs table at
    https://docs.svix.com/ai/app-portal-mcp — names and semantics are provider-
    published, input schemas require an authenticated introspection pass.
    mcp/svix-tool-crosswalk.yml binds each tool to its backing OpenAPI
    operationId, which is where the real parameter contract lives.

safety:
  provider_guidance: |
    "Most of these are read-only. Three of them are not: resend_message and
    recover_endpoint perform real deliveries, and update_transformation changes
    live endpoint configuration. The server instructs agents to only use them
    when explicitly asked, and most agents will ask for confirmation before
    running a tool, but as with any agent, your users should review those calls
    before approving them." — docs.svix.com/ai/app-portal-mcp
  data_exposure_warning: |
    Provider states plainly that message payloads the agent reads are sent to
    whichever model provider backs the coding agent. Recorded because it is an
    unusually honest disclosure and it is the correct governance input for
    agentic-access/svix-agentic-access.yml.
  mutating_tool_count: 3
  read_only_tool_count: 10

related_agent_surfaces:
  - kind: AgentSkill
    url: https://github.com/svix/ai
    detail: Two provider-published Agent Skills; see skills/_index.yml.
  - kind: agent-runtime-plugin
    name: svix-openclaw
    url: https://github.com/svix/ai/tree/main/svix-openclaw
    detail: |
      Polls a Svix sink and hands each message to the OpenClaw runtime as if it
      were an inbound POST — webhook delivery into an agent with no public URL.
  - kind: agent-runtime-plugin
    name: svix-hermes
    url: https://github.com/svix/ai/tree/main/svix-hermes
    detail: Same inversion for the Hermes Agent gateway.

gaps:
  - id: docs-ai-index-404
    detail: |
      github.com/svix/ai and its plugin.json both point at https://docs.svix.com/ai
      as the homepage, and the repo README links "our docs" at
      https://docs.svix.com/ai/mcp. Both 404 as of 2026-08-13, and neither /ai/
      page appears in docs.svix.com/llms.txt. The MCP documentation is only
      reachable at https://docs.svix.com/ai/app-portal-mcp, which is discoverable
      solely through llms-full.txt. A real agent surface is effectively
      undiscoverable from the docs index.
  - id: no-organization-scoped-mcp
    detail: |
      The only MCP server Svix ships is application-scoped and consumer-facing.
      There is no MCP surface over the 230-operation organization API that a Svix
      customer's own agent could call, so the sending side of the product has no
      agent surface beyond the REST contract and the Agent Skills.