Are Replit Agent audit logs available on every plan?

Replit documents audit logs as exclusive to Enterprise customers. An account administrator enables them in the organization's advanced settings. Teams on another plan should verify their available activity history directly rather than assuming the Enterprise portal, event set, export, or SIEM stream is present.

How long does Replit retain audit logs?

The documented default is 30 days. Replit recommends SIEM integration for longer customer-controlled retention and directs customers to Support for longer in-portal retention. Security teams should test retrieval beyond day 30 instead of treating stream configuration as proof that historical events arrived and remained intact.

Does an audit event failure block Agent activity?

Replit says the underlying action still completes if its audit event fails to record. That fail-open property belongs in the control narrative and the test plan. Monitoring should detect delivery gaps, while prevention decisions need to happen on the action or request path itself.

Can DeepInspect audit every action Replit Agent takes?

Coverage applies to HTTP LLM requests explicitly routed through DeepInspect. Replit workspace actions and local code execution remain outside that boundary. Replit-managed internal model traffic also remains outside it unless the architecture exposes and routes the relevant HTTP hop. The integration diagram must show the actual request path before anyone claims coverage.

Replit Agent Audit Logs: The 30-Day Evidence Window

Replit Enterprise audit logs give an investigator more than 50 event types, including Agent activity, with filters by actor and date as well as target and event type. One detail changes the review plan: Replit retains those records for 30 days by default. Another detail changes the control claim: if an audit event fails to record, the underlying action still completes. I want to separate a useful workspace activity feed from the evidence required to reconstruct one AI request.

TL;DR

Replit Enterprise audit logs include Agent activity among more than 50 event types and can be searched and filtered. They can also be bulk-exported or streamed to a SIEM.
Default retention is 30 days. Longer evidence preservation requires SIEM streaming or customer storage, with a separate arrangement available through Replit Support.
Audit recording is fail-open: Replit says an action still succeeds if its audit event fails to record.
A defensible AI decision record also needs the originating principal, relay identity, model route, prompt classification, policy version, outcome, timestamp, and correlation ID.

The workspace record Replit provides

Replit documents audit logs as an Enterprise-only capability that an account administrator enables under Settings, Advanced, Audit Logs. Once enabled, the portal covers deployments, identity and access changes, workspace administration, project activity, secrets, connectors, domains, and Agent activity. The public schema catalog defines approved event contracts with action versions and metadata for each target.

That scope is valuable. An administrator can establish that a member changed a connector or touched a secret. The record can also show that the member initiated an Agent-related event at a particular time. Search and filters narrow a noisy account history to one actor or target. Bulk export gives an incident team a local copy. Admin-only access also keeps the audit data away from ordinary workspace members.

The record answers an administrative question: who did what inside the Replit organization? It should be preserved and reviewed as such. I would reject any assessment that dismisses it because it lacks fields built for a different layer. The problem begins when the workspace event is presented as the complete record of a model decision.

Retention and delivery define the usable window

Thirty days is a short evidence window for a quarterly access review or an incident discovered after a billing cycle. Replit offers two paths beyond it. Administrators can bulk-export portal data, or they can stream events through the WorkOS-backed integration to Datadog, Splunk, Amazon S3, or a generic HTTP endpoint. Replit also directs customers seeking longer in-portal retention to Support.

A security team should decide the destination before enabling Agent for regulated work. The practical test is simple: pick an Agent event from 31 days ago and retrieve it without relying on a person who remembers exporting a file. If the result is an empty search screen, the retention design failed even though logging worked on the day of the action.

The SIEM route needs its own checks. Confirm delivery timestamps, alert on stalled streams, restrict deletion in the destination, and document the timezone used during exports. A CSV opened on a 13-inch laptop should line up with the SIEM event to the second. That visual reconciliation catches quiet parser and timezone errors.

Failed logging leaves the action intact

Replit states that audit recording is designed never to interfere with the underlying operation. If an event fails to record, an action such as user provisioning still completes, while the failed event is logged internally for investigation. Availability wins over audit completeness at that point.

That is an explicit design choice, and it sets the evidentiary ceiling. A workspace audit feed can show recorded actions. It cannot prove that no unrecorded action occurred during a logging failure. A reviewer should capture this property in the control description instead of writing that every action always produces an event.

The distinction matters more for automated work. Replit Agent can write code, set up infrastructure, test work, and prepare deployment. One user instruction may produce several intermediate steps. A single workspace event can establish that Agent activity occurred while still leaving the internal request sequence and policy basis outside the exported row. The model route may also be absent. The audit claim has to match the event contract rather than the product's visible outcome.

The per-decision evidence layer

An AI request review starts with a narrower unit: one HTTP exchange between an authenticated caller or agent and an LLM endpoint. The record for that exchange should carry the originating principal, the application or agent relay, the destination model route, a classification of prompt content, the versioned policy evaluated, the resulting allow or block decision, a precise timestamp, and a correlation identifier.

Replit's workspace event can be one side of that evidence package. The model provider's telemetry may be another. An independent request-boundary record supplies the policy decision between them. Correlation then lets an investigator move across all three without assuming that an actor field in one system means the same thing as a credential in another.

This is the post-authentication gap at audit resolution. A named Replit member can be authenticated while a downstream model call uses a shared service credential. The workspace knows the member. The model endpoint sees the service. The missing record is the lineage that binds both identities to one request and its content decision.

A repeatable evidence test

Run a controlled test with synthetic data before approving the integration. Have a named Replit Enterprise member run one documented Agent task using a unique marker such as AUDIT-2026-09-07-017. Record the start time and expected model route. Then retrieve every related workspace event through the portal and the SIEM destination.

Compare the output against the field set in the AI audit log schema. The test should answer these questions:

Which Replit identity initiated the work, and which agent or application relayed it?
Which HTTP model route received the request?
Which data classification applied to the synthetic marker?
Which policy version produced the decision?
Did the model call proceed, and what correlation ID joins the records?
Can the same package be retrieved after the 30-day Replit window closes?

Repeat the exercise with the SIEM stream paused. Replit's documented failure behavior means the underlying action may continue. The resulting gap should create an alert and an incident ticket. That test tells the CISO exactly which evidence disappears when delivery breaks.

Scope outside the request boundary

Workspace membership, SAML SSO, SCIM, secret handling, connector permissions, generated-code review, and local execution remain adjacent controls. Replit's shared responsibility model assigns workspace audit-log provisioning to Replit and review to the customer. It also assigns review of Agent-generated code, approval of sensitive actions, prompt hygiene, and third-party Skill or MCP vetting to the customer.

DeepInspect operates only on HTTP AI traffic deliberately routed through its gateway. It cannot retroactively fill missing Replit workspace events or govern local shell execution. It also cannot intercept Replit-managed internal model calls whose route a customer cannot place behind an external proxy. Those controls need their own owners and evidence.

This boundary keeps the audit design honest. Replit events establish platform activity. Request records establish the policy decision on a routed model call. Git history establishes code changes. IAM and endpoint tools establish access and device state. Combining those artifacts is stronger than relabeling any single feed as a complete AI audit trail.

DeepInspect

This is the gap DeepInspect closes for customer-routed HTTP model calls. DeepInspect sits inline between an authenticated user or agent and an LLM endpoint, evaluates application-supplied identity context and prompt classification against a versioned policy, and records the decision before forwarding permitted traffic.

Each routed request produces a signed, tamper-evident record with the principal, relay, model route, classification, policy version, outcome, timestamp, and correlation ID. Replit Enterprise audit logs continue to provide the workspace activity record. DeepInspect supplies the independent decision record described in signed audit logs for AI requests. Book a technical deep dive at deepinspect.ai.