Latest in AI
Sarvam Signed Webhooks: Setup Checks and Recording-Link Risks
Move Sarvam Voice Agents from inline webhook URLs to signed configs, verify raw-body signatures and keep public recording links separate from authentication.
Sarvam's November 16 Voice Agents update adds two features with similar names but different purposes. Signed webhook deliveries help your server authenticate call-result notifications. Public signed recording URLs let someone play a recording without a Sarvam login. Turning on one does not secure the other.
The practical starting point is to migrate one test campaign to a stored webhook config while leaving public recording access off unless there is a specific sharing need. This is a documentation-based implementation checklist, not a report of a tested production integration.
Decide which data flow you are changing
The release note says a workspace owner can enable Public signed URLs under Settings, Workspace, Data policy. The setting defaults to off. Once enabled, recording links work for anyone holding them and expire after 14 days.
Treat such a link as access to the recording, not as an invitation restricted to a named colleague. Before putting one into a CRM, decide who can read that CRM field, whether it appears in exports and whether support logs capture URLs. Expiration limits future access through that link; it cannot undo a copy somebody already downloaded.
For a recording that should remain inside an authenticated workflow, do not enable public access simply because a webhook integration needs authentication. These are separate decisions with separate recipients and failure modes.
Create a stored config and save its secret
Sarvam's create-config reference documents a workspace-scoped POST endpoint ending in /webhooks. Supply a name and an HTTPS receiver URL, with up to ten optional custom headers. The response returns the signing secret only once. Save it directly in your server-side secret store rather than printing the complete response into shared build logs. A workspace supports five configs; creating another returns HTTP 429.
Before sending the request, name the intended owner and environment. A label such as “QA call results” is easier to manage than “webhook2.” Keep staging and production destinations distinguishable, and decide how the receiving application will obtain the matching secret during deployment.
The config overview explains the migration: reference webhook_id instead of an inline url; exactly one may be set. If the signing secret is lost, create a replacement config, move references and then delete the old one. Do not build an incident plan around retrieving the original secret later. Allow capacity for that replacement within the five-config limit. Deleting a config still referenced by a non-terminal campaign or deployment returns HTTP 409, so removal is not necessarily immediate.
Verify before processing a call result
The campaign payload documentation defines Sarvam-Signature as an HMAC-SHA256 value over the timestamp, a period and the raw request body. Sarvam-Timestamp supplies the timestamp. Compare the expected signature with a constant-time comparison and reject timestamps older than five minutes, following the documented replay check.
Preserve the original request bytes before JSON middleware changes them. Parsing a body and serializing it again can change whitespace or key ordering, so equivalent-looking JSON is not a safe replacement for the signed bytes. An inline URL still produces unsigned deliveries; the receiver must not silently accept missing headers after the migration.
For inbound deployments, use the separate deployment payload reference. It describes notifications after inbound calls and the same stored-config signing mechanism. Keep the deployment and campaign payload schemas distinct rather than assuming every identifier and field is interchangeable.
Prove the failure cases before expanding
I would use a synthetic call result for an initial test sequence. Confirm that a valid signature reaches the processing queue, then change one byte in the body and confirm rejection. Try a missing signature, an expired timestamp and an incorrectly formatted header. These are proposed acceptance tests, not claims that Sarvam has run them for your application.
Also test delivery of the same accepted event twice. Signature validity proves authenticity, not that your business operation should happen again. Use the appropriate event identifiers and your own deduplication policy before updating customer records or sending notifications.
Keep diagnostic logs useful without copying transcripts, signing secrets or recording URLs into them. Log the failure category and a safe correlation identifier so a developer can distinguish an authentication error from a downstream processing failure.
Change shared destinations deliberately
A stored config can serve several campaigns or deployments. Sarvam says changing its URL redirects every referencing caller immediately. Inventory those references before making a production edit, prepare the new receiver and record a rollback destination.
Start with a small migration that has a clear owner. Expand only after authentic requests are processed, invalid requests are rejected and repeated events do not create duplicate business actions. Public recording links should remain a separate approval and access-review task throughout.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help teams connect voice agents to business systems with explicit authentication and data-sharing boundaries.