SIEM Streaming
Stream Gatana audit events to your own SIEM as HMAC-signed NDJSON batches, over HTTPS or into an S3 or Cloud Storage bucket.
Introduction
A step-by-step version with receiver recipes and the delivery guarantees in one table is the guide How to get audit logs of every MCP tool call into your SIEM.
SIEM streaming delivers your organization's audit events to a security platform you control. Gatana pushes events to a destination you configure, so your own detection rules, dashboards, and retention apply to Gatana activity alongside the rest of your estate.
Choose a destination:
| Kind | What Gatana does | Typical use |
|---|---|---|
| HTTPS endpoint | POSTs each batch to your URL | Splunk HTTP Event Collector, CrowdStrike Falcon LogScale, Elastic, Microsoft Sentinel through a relay, in-house pipelines |
| S3 bucket | Writes each batch as one object into a bucket on Amazon S3 or any S3-compatible store | Splunk SQS-based S3 input, Microsoft Sentinel S3 connector, Datadog, Elastic, Athena, long-term archive |
| Cloud Storage bucket | Writes each batch as one object into a Google Cloud Storage bucket | Google SecOps feeds, BigQuery external tables, Elastic, long-term archive |
Configure it under Settings → SIEM Streaming in the dashboard. One destination per organization. The feature requires a paid plan.
Event sources
Two sources are streamed, and each can be turned off independently.
| Source | Contents |
|---|---|
audit_logs | Server, profile, team, and organization configuration changes; sign-ins; user and team membership changes; secret fetches; and every MCP request, including tool calls |
audit_logs_credentials | Credential creation, updates, and deletion, plus automatic and manual OAuth token refreshes |
How much detail MCP events carry is controlled by the MCP audit log level in your organization settings; see Audit Logs for the levels. At terse, the default, an event records which tool was called, how long it took, and which credential made the call.
Internal events visible only to Gatana staff, such as billing webhook traffic, are never streamed.
Envelope schema
Each event is one JSON object on its own line.
{
"id": "audit_logs:184223",
"version": 1,
"time": "2026-08-04T09:52:04.118Z",
"tenant": "acme",
"source": "audit_logs",
"action": "mcp_server.update",
"actor": { "type": "user", "id": "usr_8Fq2..." },
"entity": { "type": "mcp_server", "id": "srvr_pQ1..." },
"details": {}
}| Field | Description |
|---|---|
id | Stable, unique, and repeatable for the same event. Use it to deduplicate |
version | Envelope version. A new version is introduced only for a breaking change |
time | When the event happened, ISO 8601 UTC |
tenant | Your organization identifier |
source | audit_logs or audit_logs_credentials |
action | Dotted action name, described below |
actor | { "type": "user", "id": "..." }, or { "type": "system", "id": null } for background work and database triggers |
entity | What the event is about. Either field can be null |
details | Event-specific payload. Shape varies by action |
Event catalog
Most actions follow {entity type}.{event name}, so an action appears for a new kind of event without any change on your side. Write rules against the prefix where you can.
| Action | Meaning |
|---|---|
mcp_server.create / .update / .delete | An MCP server was added, changed, or removed |
mcp_server.start / .stop | A hosted server was started or stopped |
mcp_server.credentials.upsert / .delete / .delete_all / .copy | Server credentials were changed |
profile.create / .update / .delete | A profile was changed |
team.create / .update / .delete | A team was changed |
team.member_join / .member_leave | Team membership changed |
tenant_configuration.update | Organization settings changed |
siem_destination.create / .update / .delete | This streaming configuration itself changed |
alert_rule.create / .update / .delete | An alert rule was changed |
alert_rule_channel.create / .delete | A channel was added to or removed from the selected channels of an alert rule |
notification_channel.create / .update / .delete | A notification channel was changed |
connected_client.create / .update / .delete | A user's OAuth client connection appeared, was configured, or was disconnected |
connected_client_profile.create / .delete | A profile was attached to or detached from a connected client |
user.login_success | A user signed in |
user.created / .disabled / .disabled_due_to_seat_limit | A user account changed |
secret_store.secret_fetched | A secret was read from an external secret store |
firewall.deny / firewall.log | Egress from a hosted server was blocked or recorded |
mcp.tools/call | A tool was called. details.outcome says how it ended (ok, denied, not-found, needs-credentials, credentials-rejected, upstream-error, invalid-arguments, failed, unauthenticated), and a denied event carries details.reason; see Audit Logs. Other JSON-RPC methods appear under the same mcp. prefix |
credential.create / .update / .delete | A credential record changed |
credential.auto_oauth_refresh_success / .auto_oauth_refresh_failed | Gatana refreshed an OAuth token in the background |
credential.manual_oauth_refresh_success / .manual_oauth_refresh_failed | A user refreshed an OAuth token |
credential.expired_no_refresh | An access token expired with no refresh token available |
credential.rejected_by_upstream | The upstream MCP server rejected the access token; Gatana tries to refresh and tries again |
gatana.test | A test event sent from the dashboard |
Configuration changes are audited too
Changes to the streaming destination arrive in the stream as siem_destination.*, so a rule can alert when somebody redirects your audit trail. Delivery bookkeeping such as the internal read position is not an event and produces nothing.
Delivery
Each batch holds up to 500 events or about 1 MB, with one event per line. Export runs once a minute. A short delay lets pending database writes finish before export.
HTTPS endpoint. Batches are POSTed as application/x-ndjson. Respond 2xx as soon as you have accepted the batch. Requests time out after 10 seconds, and redirects are not followed.
S3 and Cloud Storage buckets. Each batch becomes one object. Its key uses the event IDs and the date and hour of the first event:
<prefix><source>/dt=2026-10-24/09/<tenant>_<source>_<first id>_<last id>.ndjson
gatana/audit_logs/dt=2026-10-24/09/acme_audit_logs_184001_184500.ndjson
gatana/audit_logs_credentials/dt=2026-10-24/09/acme_audit_logs_credentials_0192a4b0-…_0192a4b0-….ndjsonThe default prefix is gatana/. The dt= segment is a Hive-style partition that Athena and BigQuery external tables read as a column. Compressed objects end in .ndjson.gz. Use this suffix to select gzip decompression; Content-Encoding is not set. Every object carries custom metadata, which S3 presents with an x-amz-meta- prefix:
| Metadata key | Value |
|---|---|
gatana-signature | The signature described below, over the uncompressed body |
gatana-source | audit_logs or audit_logs_credentials |
gatana-first-id / gatana-last-id | Row ids of the first and last event in the object |
gatana-event-count | Number of events in the object |
Delivery is at-least-once. A batch that is accepted but whose acknowledgement is lost is sent again, so the same event can arrive twice. Deduplicate on id. For buckets, retries with the same events reuse the object key. A retry with more events gets a new key, leaving both objects in the bucket. Events from each source arrive in order.
If your endpoint fails, Gatana retries with a widening delay of 1, 5, 15, and then 60 minutes:
- after 1 hour of failures the destination is marked failing in the dashboard
- after 24 hours your administrators receive one email
- after 7 days delivery is disabled and a second email is sent
Nothing is lost while a destination is failing or paused. The read position is kept, so reactivating resumes where delivery stopped, as far back as audit retention allows.
Verifying signatures
Every batch carries a signature over the exact NDJSON body. An HTTPS receiver finds it in a request header; a bucket object carries it in its metadata under gatana-signature:
X-Gatana-Signature: t=1785825392,v1=6c3b...f19at is a Unix timestamp in seconds and v1 is HMAC-SHA256 of "{t}.{raw body}" keyed with your signing secret. The secret is shown once when you configure the destination and can be revealed again from the dashboard.
Verify against the raw body, before any JSON parsing. For a gzipped object, decompress first; the signature covers the plain NDJSON bytes. Reject a batch whose timestamp is outside a tolerance you choose; five minutes is a reasonable default for a live endpoint. For stored objects, allow for the delay between signing and reading the object.
import { createHmac, timingSafeEqual } from 'node:crypto';
function verify(rawBody, header, secret, toleranceSeconds = 300) {
const parts = Object.fromEntries(header.split(',').map(p => p.split('=')));
const age = Math.abs(Math.floor(Date.now() / 1000) - Number(parts.t));
if (!Number.isFinite(age) || age > toleranceSeconds) return false;
const expected = createHmac('sha256', secret).update(`${parts.t}.${rawBody}`).digest('hex');
const a = Buffer.from(expected, 'hex');
const b = Buffer.from(parts.v1 ?? '', 'hex');
return a.length === b.length && timingSafeEqual(a, b);
}import hashlib, hmac, time
def verify(raw_body: bytes, header: str, secret: str, tolerance_seconds: int = 300) -> bool:
parts = dict(p.split("=", 1) for p in header.split(","))
if abs(int(time.time()) - int(parts["t"])) > tolerance_seconds:
return False
expected = hmac.new(
secret.encode(), f"{parts['t']}.".encode() + raw_body, hashlib.sha256
).hexdigest()
return hmac.compare_digest(expected, parts["v1"])Rotating the secret takes effect immediately, so update your receiver first or accept a short gap. If your receiver can check two secrets at once, add the new one before rotating.
What is never sent
- Secret values. Tokens, passwords, client secrets, private keys, and API keys are replaced with
[REDACTED]wherever they appear, at any depth. - Encrypted secret columns. A change to a stored secret is reported as
{"changed": true}and never decrypted. Encrypted configuration, such as a server's transport settings, is included so that you can see what changed, with any embedded credentials redacted inside it. - Anything Gatana cannot decrypt appears as
<DECRYPT_FAILED>rather than blocking the stream.
A details payload larger than 64 KB is replaced with {"_truncated": true, "_originalBytes": …, "preview": "…"}. This mainly affects verbose MCP logging, where a whole tool response is recorded.
Receiver setup
HTTPS endpoint
The endpoint URL must use HTTPS. In Gatana Cloud it must also resolve to a public address: internal hostnames and private address ranges are refused, because a shared platform must not be pointed at infrastructure it should not reach. Self-hosted deployments have no such boundary, so they may stream to a collector on their own network, such as https://splunk.internal/… or a private address. HTTPS is required either way, so audit data never travels in the clear.
Use the optional auth header to carry whatever credential your collector expects, for example Authorization: Splunk <token> for Splunk HTTP Event Collector, or Authorization: ApiKey <key> for Elastic. The value is stored encrypted and never returned by the API.
Splunk HTTP Event Collector. Point the URL at https://<host>:8088/services/collector/raw and set the auth header to Authorization with the value Splunk <hec-token>. The raw endpoint accepts newline-delimited JSON directly. Set sourcetype on the token or its input.
CrowdStrike Falcon LogScale. Use the HEC-compatible ingest endpoint of your repository, https://<host>/api/v1/ingest/hec/raw, with Authorization: Bearer <ingest-token>.
Elastic. Send to an HTTP input on your ingest pipeline, then split lines and parse each as JSON.
Anything else. Any endpoint accepting an authenticated HTTPS POST of newline-delimited JSON works. Verify the signature, split on newlines, parse each line, and deduplicate on id.
S3 bucket
Create a bucket, or choose a prefix in one you have, and an IAM user whose only permission is to put objects under that prefix. Gatana needs no list or read access. Enter the bucket, region, prefix, access key id, and secret access key in the dashboard. The secret is stored encrypted and never returned.
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": "s3:PutObject",
"Resource": "arn:aws:s3:::acme-audit/gatana/*"
}
]
}For an S3-compatible store such as MinIO, OVH Object Storage, or Cloudflare R2, set the store's HTTPS endpoint and turn on path-style addressing. The endpoint follows the same rule as an HTTPS destination: public in Gatana Cloud, any address in a self-hosted deployment. Checksum headers that only Amazon S3 understands are left out when an endpoint is set.
Bucket encryption settings apply. Configure Object Lock, lifecycle rules, and replication in your storage provider.
Cloud Storage bucket
Create a service account, grant it the Storage Object Creator role (roles/storage.objectCreator) on the bucket, and create a JSON key for it. Paste the key file into the dashboard together with the bucket and prefix. The key is stored encrypted and never returned; the dashboard shows the service account email so you can tell which identity is in use.
gcloud iam service-accounts create gatana-siem --project acme-logs
gcloud storage buckets add-iam-policy-binding gs://acme-audit \
--member serviceAccount:[email protected] \
--role roles/storage.objectCreator
gcloud iam service-accounts keys create gatana-siem.json \
--iam-account [email protected]Object Creator cannot overwrite objects. Gatana creates each object only if its name is unused. A retry counts an existing object as delivered.
Reading objects into your SIEM
| Platform | How |
|---|---|
| Splunk | Splunk Add-on for AWS, SQS-based S3 input with an S3 event notification on the prefix. Source type: a custom JSON type with line breaking on newline |
| Microsoft Sentinel | Amazon S3 connector with an SQS queue on the prefix, or the Codeless Connector Framework for custom NDJSON |
| Datadog | Datadog Forwarder Lambda triggered by S3 object creation; it handles .gz |
| Elastic | Elastic Agent aws-s3 input with SQS notifications, or the gcs input for a Cloud Storage bucket; both decode gzip and split lines |
| Google SecOps (Chronicle) | A Cloud Storage feed on gs://<bucket>/<prefix>, log type JSON |
| BigQuery or Athena | An external table over the prefix with Hive partitioning on dt, format newline-delimited JSON |
In every case, verify gatana-signature from the object metadata if your pipeline can, and deduplicate on id if you ever ingest the same object twice.
Testing a destination
Use Send test event under Settings to deliver a single gatana.test event. For an HTTPS endpoint the dialog shows the HTTP status and the beginning of the response body, which is usually enough to identify a rejected credential or a wrong path. For a bucket, it shows the test object under <prefix>gatana-test/ or the storage error.
Last updated on