Docs

AI Coding Agents

Agent Keeper monitors AI coding agents across the major IDEs. Shell commands, file reads and writes, MCP tool calls, and prompts are intercepted before execution so you can block, warn, or audit every action.

Coverage and enforcement capabilities differ by IDE based on what each platform exposes through its hook system.

Capability matrix

CapabilityClaude CodeCodexGemini CLICursorVS Code CopilotWindsurf
Shell command blockingYesYesYesYesYesYes
File read auditYesYesYesYesYesYes
File write blockingYesYesYesAudit onlyAudit onlyYes
Prompt blockingYesYesYesYesPreviewYes
MCP call monitoringYesYesYesYesLimitedYes
Session check-inYesYesYesFirst eventFirst eventFirst event
Hook config location~/.claude/settings.json.codex/config.toml.gemini/settings.json.cursor/hooks.json.github/hooks/agentkeeper.json.windsurf/hooks.json

Setup guides

  • Claude Code Setup: plugin-based setup with full enforcement and MCP monitoring.
  • Codex Setup: OpenAI Codex hooks with shell, patch, prompt, and MCP policy coverage.
  • Gemini CLI Setup: Gemini command hooks with file, shell, prompt, and MCP policy coverage.
  • Cursor Setup: project hooks for Cursor with shell, prompt, file, and MCP coverage.
  • VS Code Copilot Setup: GitHub Copilot agent hook configuration where Preview hooks are available.
  • Windsurf Setup: Windsurf Cascade hooks with strong file write enforcement.

How monitoring works

Each IDE integration works by registering Agent Keeper as an HTTP hook receiver. When the agent is about to execute a tool call, the IDE sends the event to Agent Keeper before the action runs. Agent Keeper evaluates the event against your org's security policies and returns a verdict:

  • Block: the tool call is denied; the agent receives a rejection message.
  • Warn: the call proceeds but is flagged in the audit trail.
  • Pass: the call proceeds and is logged.

All events are recorded in the Activity page regardless of verdict.

Runtime Shield

The Runtime Shield runs across all IDE integrations. It detects credential exfiltration patterns, reverse shells, prompt injection, and 55+ threat patterns in prompts and tool call inputs. Shield events appear in the Security dashboard and can trigger email alerts (Pro+) or webhook alerts (Team+).