## Overview Security and compliance teams now have visibility into coding agents that run **outside** GitLab — Claude Code, Cursor, Codex, and others — through the [AI Governance Dashboard](https://docs.gitlab.com/user/ai-governance/governance-dashboard/). Previously, the AI Governance Dashboard showed activity only from GitLab Duo Agent Platform (DAP) agents. With GitLab 19.5, after a one-time `glab agent setup`, every external agent session is captured automatically and surfaced alongside your DAP agents on the same pane of glass. This feature is in **beta** and available on **GitLab Premium and Ultimate** (GitLab.com, Self-Managed, and GitLab Dedicated). --- ## Problem Developers increasingly use third-party coding agents — Claude Code, Cursor, Codex — that operate entirely outside GitLab. These agents read files, write code, run shell commands, and push commits, but security and compliance teams had no visibility into what they were doing, who was running them, or which projects they were touching. This created a blind spot in AI governance: DAP agents were fully audited, but any agent running locally or via a third-party tool was invisible. --- ## Proposal Connect external coding agents to GitLab AI Governance using the GitLab CLI (`glab`) as the bridge. The design has two parts: **1. Agent detection and registration** `glab` detects when it is being invoked inside a coding agent session by reading environment variables set by Claude Code, Cursor, Codex, and other agents. On first detection within a new session, `glab` silently registers the agent against the GitLab AI Catalog in the background. This requires no developer action — any developer already using `glab` alongside a supported agent gets registration automatically. **2. Session capture and audit** Claude Code's `SessionEnd` lifecycle hook is configured (via `glab agent setup`) to POST session data directly to a GitLab API endpoint when the session ends. GitLab receives the payload, creates a session record, and generates audit events for each tool call. For sessions where the hook does not fire (agent crash, abnormal exit), a periodic `glab agent audit sync` provides a fallback by scanning for unsynced JSONL files and uploading them in batch. Tool call audit events include: tool name, input arguments, timestamp, agent identity, authenticated GitLab user, and project. **Tool results are deliberately excluded** to protect sensitive file contents and command output. --- ## What This Delivers - A record of every external agent session: who ran it, which agent, which project, and when - A full tool call timeline for each session (file reads, file writes, bash commands, MCP tool calls) - Audit events in the existing GitLab audit stream, available for SIEM export via audit event streaming - Sessions and tool calls visible in the AI Governance Dashboard alongside DAP agent sessions --- ## Setup | User group | Setup required | |---|---| | Claude Code (non-Enterprise) | One-time `glab agent setup` | | Claude Code Enterprise | Admin pushes managed settings via Claude.ai admin console | | Cursor / Codex | `glab agent setup` (hook configuration) | --- ## Scope and Limitations (Beta) - Initially scoped to Claude Code, Cursor, and Codex - Developers not using `glab` are not captured in Stage 1 - Windows support is out of scope for the initial beta - Developers using raw API clients, n8n, Bedrock, or custom harnesses bypass this entirely — accepted for Stage 1 --- ## Technical Implementation The backend foundation is fully shipped in 19.5: - **Agent identity registration API** — `POST /api/v4/projects/:id/ai_agent/identities` registers an external agent against a project - **External agent session API** — `POST /api/v4/projects/:id/ai_agent/sessions` and `PATCH /api/v4/projects/:id/ai_agent/sessions/:id` create and complete session records, stored in `ai_governance_sessions` (migrated from `duo_workflows_workflows` to a dedicated AI Governance-owned table) - **Audit events API** — `POST /api/v4/projects/:id/ai_agent/audit_events` writes structured audit events using the existing `AuditEvents::AiAuditEvent` model, with `cloud_event_id` deduplication - **`glab agent` CLI command group** — `glab agent setup`, `glab agent audit sync`, `glab agent list`, `glab agent doctor` Supported audit event types: `ai_agent_session_started`, `ai_agent_session_ended`, `ai_tool_invoked` (covering `Bash`, `Read`, `Write`, `Edit`, `WebFetch`, `mcp_tool_call`, and subagent calls). --- ## Documentation - [AI Governance Dashboard](https://docs.gitlab.com/user/ai-governance/governance-dashboard/) --- ## Feedback This feature is in beta. Please share feedback, bug reports, and use cases in the comments below. --- > **Note:** This is a public tracking issue for the feature shipped in GitLab 19.5. The internal epic with full implementation details is confidential. <!-- triage-serverless v3 PLEASE DO NOT REMOVE THIS SECTION --> *This page may contain information related to upcoming products, features and functionality. It is important to note that the information presented is for informational purposes only, so please do not rely on the information for purchasing or planning purposes. Just like with all projects, the items mentioned on the page are subject to change or delay, and the development, release, and timing of any products, features, or functionality remain at the sole discretion of GitLab Inc.* <!-- triage-serverless v3 PLEASE DO NOT REMOVE THIS SECTION -->