Dubber

dubber-mcp

MCP server for Dubber — call recording upload/retrieval, compliance metadata and tags, group/account/user management, REST-hook (webhook) notifications, and Dub.Point telecom-system (e.g. BroadWorks) integration.

Built on the MCP 2026-07-28 spec via the split v2 SDK (@modelcontextprotocol/server / /node / /client ^2.0.0-beta.5) with dual-era serving: one shared McpServerFactory behind createMcpHandler({ legacy: 'stateless' }) answers both 2025-era initialize-handshake clients (the WYRE gateway today) and modern 2026-07-28 envelope clients, with an identical, deterministic 37-tool surface for every caller. Ships as a GHCR container only (no MCPB bundle). The Dubber client is @wyre-ai/node-dubber.

Tools (37, flat)

• Core: dubber_test_connection.
• Groups: dubber_groups_get, dubber_groups_create_child, dubber_groups_list_unidentified_recordings, dubber_groups_create_unidentified_recording.
• Accounts: dubber_accounts_create, dubber_accounts_get, dubber_accounts_update.
• Recordings: dubber_recordings_list, dubber_recordings_create, dubber_recordings_get, dubber_recordings_get_waveform, dubber_recordings_update_metadata, dubber_recordings_add_tags, dubber_recordings_initiate_multipart_upload, dubber_recordings_get_upload_target, dubber_recordings_complete_upload. Gated: dubber_recordings_delete (⚠ DESTRUCTIVE), dubber_recordings_delete_tags (⚠ HIGH-IMPACT). dubber_recordings_list lists newest first with limit (1–500, default 100), offset, cursor, query, dubPointId/teamId, and a start_time window — see Listing recordings.
• Users: dubber_users_list, dubber_users_create, dubber_users_get, dubber_users_update. Gated: dubber_users_delete (⚠ DESTRUCTIVE).
• Profile: dubber_profile_get.
• Notifications (REST hooks): dubber_notifications_list, dubber_notifications_create, dubber_notifications_get, dubber_notifications_update, dubber_notifications_activate, dubber_notifications_list_unclaimed. Gated: dubber_notifications_delete (⚠ HIGH-IMPACT).
• Dub.Point: dubber_dub_points_list, dubber_dub_points_create, dubber_dub_points_get, dubber_dub_points_find.
• OAuth: gated dubber_oauth_revoke_token (⚠ HIGH-IMPACT).

Credentials

Four values, all required — none of them a human login (see "Vendor quirks" below).

In gateway mode a request missing any of the four required headers is answered 401 (JSON-RPC error -32001) before the MCP handler runs. It never falls through to env credentials — a header-controlled base URL is likewise refused entirely (env-mode only), so a caller can never redirect client secrets to another host. Blank or repeated header values count as missing.

/health and /healthz stay unauthenticated and do not read credentials. The stdio transport does not use CONDUIT_S2S_SECRET.

HTTP authentication

With MCP_TRANSPORT=http and CONDUIT_S2S_SECRET set, every request except /health and /healthz must include X-Gateway-S2S. The check is verifyS2sHeader in src/s2s-verify.ts: the header value is t= ,v1= , where the hex is HMAC-SHA256 of the literal string t= keyed with CONDUIT_S2S_SECRET (64 lowercase hex characters). The timestamp must be within 300 seconds of the server clock. A missing or invalid header is HTTP 401, before vendor credentials are read.

TS=$(date +%s)

SIG=$(printf 't=%s' "$TS" | openssl dgst -sha256 -hmac "$CONDUIT_S2S_SECRET" -hex | awk '{print $NF}') curl -sS http://127.0.0.1:8080/mcp \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -H "X-Gateway-S2S: t=${TS},v1=${SIG}" \ -H "X-Dubber-Client-Id: $DUBBER_CLIENT_ID" \ -H "X-Dubber-Client-Secret: $DUBBER_CLIENT_SECRET" \ -H "X-Dubber-Auth-Id: $DUBBER_AUTH_ID" \ -H "X-Dubber-Auth-Token: $DUBBER_AUTH_TOKEN" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

In AUTH_MODE=gateway, /mcp also requires X-Dubber-Client-Id, X-Dubber-Client-Secret, X-Dubber-Auth-Id, and X-Dubber-Auth-Token (X-Dubber-Region is optional). Those headers are the only credential source for that request. AUTH_MODE=env still requires the signed X-Gateway-S2S header, then uses DUBBER_* from the environment.

Running

export NODE_AUTH_TOKEN=$(gh auth token)   # GitHub Packages auth for @wyre-ai/*

npm install npm run build node dist/index.js # stdio (default)

HTTP refuses to start until CONDUIT_S2S_SECRET is set

CONDUIT_S2S_SECRET="$CONDUIT_S2S_SECRET" MCP_TRANSPORT=http node dist/index.js

Local HTTP only, without a service-to-service secret (always binds 127.0.0.1)

MCP_TRANSPORT=http MCP_ALLOW_INSECURE_DEV=1 node dist/index.js npm run smoke # proves both protocol eras serve the same tools

Docker (linux/amd64 per fleet law). The image listens on 0.0.0.0 inside the container (MCP_HTTP_HOST) and defaults to AUTH_MODE=gateway. Publish the port on loopback and pass a real S2S secret. In gateway mode, Dubber credentials arrive on each request as headers; they are not read from the environment.

docker build --platform linux/amd64 --build-arg GITHUB_TOKEN=$(gh auth token) -t dubber-mcp .

docker run \ -e CONDUIT_S2S_SECRET="$CONDUIT_S2S_SECRET" \ -p 127.0.0.1:8080:8080 \ dubber-mcp

Compose refuses to start when CONDUIT_S2S_SECRET is unset. The gateway service sets AUTH_MODE=gateway, does not inject DUBBER_CLIENT_ID, DUBBER_CLIENT_SECRET, DUBBER_AUTH_ID, or DUBBER_AUTH_TOKEN, and publishes 127.0.0.1:8080:8080:

CONDUIT_S2S_SECRET="$CONDUIT_S2S_SECRET" docker compose up

Single-tenant HTTP (AUTH_MODE=env) still requires the S2S secret, then uses the DUBBER_* variables. It is separate from the gateway Compose service above.

docker run \

-e CONDUIT_S2S_SECRET="$CONDUIT_S2S_SECRET" \ -e AUTH_MODE=env \ -e DUBBER_CLIENT_ID=... -e DUBBER_CLIENT_SECRET=... \ -e DUBBER_AUTH_ID=... -e DUBBER_AUTH_TOKEN=... \ -p 127.0.0.1:8080:8080 \ dubber-mcp

Elicitation and destructive-action consent

Every gated tool asks before acting. Elicitation rides the SDK v2 MRTR seam: handlers return input_required results that 2026-07-28 clients fulfil and retry, and that the SDK's legacy shim fulfils server-side for 2025-era stateful connections (stdio). All reads and the confirmation happen before the single mutating Dubber call.

Callers that cannot be prompted (including stateless legacy HTTP requests, which is how the WYRE Conduit gateway connects) fail closed: gated tools refuse to run unless the call passes "confirm_destructive_action": true. That argument is consulted only when no prompt is possible and never skips a confirmation an interactive user would have seen.

Listing recordings and time windows

Dubber's recordings endpoint has only count (max 100), page (max 5 pages), a before_id cursor and query — no limit/offset and no date filter, and it silently ignores unknown params. dubber_recordings_list therefore pages Dubber itself (count=100, chaining before_id) and applies the rest server-side:

• limit (1–500, default 100): max matching recordings returned. offset (≥0): matching recordings (newest first) to skip.
• startTimeFrom (inclusive) / startTimeTo (exclusive for date-times; a date-only value includes the whole day): ISO 8601, no offset = UTC. They filter on each recording's start_time after fetching; recordings with an unparseable start_time are excluded and counted in excludedUnparseableStartTime. Weekly pull in Melbourne daylight time (AEDT, +11:00), Mon–Fri: startTimeFrom=2026-10-19T00:00:00+11:00, startTimeTo=2026-10-24T00:00:00+11:00, limit=500.
• dubPointId / teamId (optional, trimmed, non-empty; sent as Dubber's dub_point / team on every page request and echoed in filters): one dub point or one team per call — Dubber forbids combining them, so passing both is an error. Dubber has no direct user filter on recordings: for one technician's week, find their dub point id(s) with dubber_dub_points_list (each dub point's user is a URL ending in /users/ ), then call once per dub point. A cursor must be reused with the SAME dubPointId/teamId.
• Each call makes at most 25 Dubber requests (2,500 recordings scanned). If that cap or limit stops the scan early, hasMore is true and nextCursor is set: call again with cursor= , offset= (0 unless the scan cap hit while still skipping) and the same filters to continue with no gaps or duplicates. scanCapReached says whether the cap was the reason. With startTimeFrom set, the scan stops once a whole page of recordings starts before it.
• Result: { recordings, returned, hasMore, nextCursor, remainingOffset, scanned, requests, scanCapReached, filters, note }.

The handler sends Dubber's params through a single cast over @wyre-ai/node-dubber 1.0.0 (whose typed params are {limit, offset}); drop the cast when the SDK release that types count/before_id/query is adopted.

Vendor quirks encoded here

• Password-grant auth with no human password. username/password on the OAuth token request are the Dubber Auth ID / Dubber Auth Token pair from the portal's API tab, not a person's login. See docs/DESIGN.md §3.
• Region-scoped hosts, no single production default. sandbox is free/self-serve; production is https://api.dubber.net/ /v1 for a region confirmed with Dubber — shipping against the wrong one silently talks to the wrong tenant.
• No downloadable API spec. Every endpoint and the auth flow were verified against the live /io-docs console and the official Getting Started Guide; full response-body