Skip to main content
Quint has two webhook destinations. They sit in different sections of the Integrations page and carry different things. Creating or changing either needs the Admin or Owner role.

The two webhooks

Custom header values on the generic webhook are treated as secrets. Quint does not show them back to you.

Verify a signed alert delivery

Each alert webhook delivery carries three headers you can check: To verify in plain words:
  1. Read the request body as raw bytes. Do not parse and re-serialise it first; any change to the bytes changes the signature.
  2. Compute HMAC-SHA256 over those bytes with the key Quint showed you when the destination was created or rotated.
  3. Encode the result as lowercase hex and prefix it with sha256=.
  4. Compare it with X-Quint-Signature using a constant-time comparison. Reject the delivery if they differ.
  5. If you also want to reject replays, reject a delivery whose X-Quint-Timestamp is older than you are willing to accept.
During a rotation, verify against X-Quint-Signature with the old key or against X-Quint-Signature-Next with the new key. Either one is enough. Each header carries exactly one value; Quint never joins two signatures into one header.
The key is shown once. If you lose it, click Rotate key and store the new one. The old key keeps working for 72 hours so your receiver does not drop deliveries while you switch; after that it is dropped on its own.

Alert delivery body

An alert delivery is one JSON object with these fields. Fields marked optional are omitted when empty. Approval request deliveries to an alert webhook use a different body. Field reference: ask your Quint contact.

Tool supply-chain findings

An alert for a tool supply-chain finding has one of these subtypes: mcp_tool_redefined (MCP tool redefined), mcp_tool_description_injection (Tool description injection), agent_builtin_tool_drift (Built-in tool changed) or, with an upcoming agent release, mcp_tool_cross_reference (Tool cross-reference). The alert carries tool_name and a one-line message; it carries no labels. The message for a cross-reference names both tools and both servers, unless content suppression is on. With that release, a description flagged the first time Quint saw the tool raises an alert whose message says the tool was first advertised with a description flagged for review. A cross-reference is low severity.

Exported event fields

Every batch exporter, including the generic webhook, sends events in this shape. The generic webhook sends a JSON array of these objects. Splunk receives each one wrapped in the HTTP Event Collector envelope (event, time, index, sourcetype, source). Fields marked optional are omitted when empty. For a tool supply-chain finding, labels holds divergence_type (which finding it is), finding_family = supply_chain and server_id (redacted when content suppression is on). With an upcoming agent release labels also holds:

When Quint changes an event to store it

Quint records an event when one of its named fields, such as tool_name, mcp_server or decision, holds a character the database cannot store, and when labels, arguments or result hold a null character or text that only spells out an escape sequence. The character or that text is removed, or in a few cases replaced by the standard replacement character (U+FFFD). If labels, arguments or result is no longer valid JSON after that, the value is kept whole as text. A few events are still set aside rather than recorded, so they never reach an export. These include an event whose source address is not an IP address, whose arguments, result or labels hold malformed text the database rejects (bytes that are not UTF-8 inside a JSON string, or half of a character pair sent as an escape), whose arguments or result hold a number too large for the database, whose arguments, result or labels repeat a name in one object where a copy holds such a number, whose run_id or workload_id label is too long for the database to index, or whose latency is too large for its field. Five values have a length limit on every event: action_type, tool_name and mcp_server at 1,024 bytes, and decision and the confidence band (stored, but not one of the exported fields above) at 128 bytes. A longer value is cut at that length and ends with …[truncated], so it can be slightly longer than the limit. An event changed in any of these ways carries _cleaned_columns in labels: the names of the changed fields, separated by commas. This includes ordinary tool output with a null character in it, binary output for example, which Quint also stored without that character before but did not mark. Two cases are not marked:
  • A tool result that arrives after its tool call was recorded and has to be changed: the call’s record holds the changed result without the label.
  • A labels, arguments or result value that was not valid JSON as sent: it is kept whole as text, as before.
If the labels your agents sent on a changed event were not a JSON object, Quint moves them under _labels_as_sent to make room for the mark. A label your agents send under the name _cleaned_columns is not stored as sent. A threat detection whose agent id or session id holds a character the database cannot store is recorded without that id and still links to its event.

Program-start rows on macOS

With an upcoming agent release, when one tool call starts many programs that carry a finding Quint has already reported, Quint records summary rows for the starts past a limit. A summary row carries two counts in labels, exec_budget_folded_program and exec_budget_folded_call: how many program-start reports have been folded so far, for the row’s program and for the whole call. Both counts are kept when content suppression is on. Two observers on the Mac can each report the same command, and one can name it after the agent while the other names it after the program that launched it. With that release, a program-start row covers the commands one launching program ran under one parent process in the same half second:
  • The executable label holds the launching program’s full path, for example /bin/zsh or /usr/bin/make, and the row holds the most recent command it ran. With content suppression on, executable holds the redaction marker instead.
  • tool_name is the agent’s name, for example claude-code, when either observer named a command on the row after the agent. Otherwise it is the launching program’s name.
  • The raw_event_count label counts the reports the row stands for, not the commands: a command both observers saw counts as 2.
With that release, a search, saved view or SIEM rule that looks for the launching program in the tool name (tool_name = zsh) should match the end of the executable path instead (/zsh). In a workspace with content suppression on, the launching program is not in executable, and a row named after the agent does not show it.

Admin audit record fields

The second stream every enabled batch exporter receives is the admin audit log: who changed what in the workspace. Token mints, role changes, enforcement changes and sign-ins are recorded here. The generic webhook sends these as a JSON array with the header X-Quint-Stream: admin_audit.

Rotate key and Replace URL

  • Rotate key (alert webhook) creates a new signing key and shows it once. Deliveries carry both signatures for 72 hours, as described above. Store the new key and switch your receiver within that window; the old key then lapses on its own.
  • Replace URL (alert webhook) changes the receiver base URL. Click Send test afterwards and confirm the new receiver verifies the signature.
  • Edit (generic webhook) changes the endpoint URL or custom headers.

Next

Last modified on November 15, 2026