Runline

Live demo

Agent capabilities, terminal commands, and console destinations for a 5-minute narrative.
Open alerts

Recommended flow

  1. Start console: RUNLINE_AUTH_DISABLED=1 pnpm dev · optional pnpm seed:reset
  2. Enroll agent, then confirm with runline-agent status
  3. runline-agent classify — prove AI session
  4. Run scenario #1 below, then click alert row for drawer

Full walkthrough: docs/demo-live.md in the repo

Demo scenarios

  • 1. AI agent blocked on destructive S3

    The headline demo — deny in terminal, alert in console in seconds.

    export CLAUDECODE=1 && aws s3 rb s3://runline-demo-bucket --force

    Rule hint: s3-remove-bucket-deny

    deny
    Console ai-agent
  • 2. AI blocked on kubectl delete namespace

    Proves multi-tool coverage beyond AWS.

    export CLAUDECODE=1 && kubectl delete namespace prod

    Rule hint: kubectl-delete-namespace-ai

    deny
    Console ai-agent
  • 3. AI cannot create IAM users

    Shows identity-plane protection, not just data plane.

    export CLAUDECODE=1 && aws iam create-user --user-name rogue-agent

    Rule hint: iam-create-role-no-ai-agent

    deny
    Console ai-agent
  • 4. Same command, human gets approval

    AI is denied (above); a human running the exact same command is routed to approval, not blocked.

    aws s3 rb s3://runline-demo-bucket --force

    Rule hint: human-aws-bucket-or-data-destroy-approve

    approve
    Console human
  • 5. Audited emergency bypass

    Break-glass without turning off the agent.

    RUNLINE_BYPASS=1 aws s3 rb s3://runline-demo-bucket --force
    bypass
    Console human

What the agent does today

  • AWS CLI shimlive

    Every aws invocation is evaluated by runlined before exec — no wrapper scripts.

  • Human vs AI actor classificationlive

    TTY, process ancestry, and env markers (CLAUDECODE, CURSOR_AGENT, …) stamp every audit record.

  • Embedded policy (45+ rules)live

    Destructive IAM, S3, EC2, kubectl, git, terraform, MCP tools/call, and exfiltration patterns.

  • TTY approval flowlive

    High-risk human actions can require on-terminal approval; AI agents never get prompted.

  • Governed bypasslive

    RUNLINE_BYPASS is audited; bypassed events surface in Logs and alert tabs.

  • Tamper detectionlive

    Binary manifest verified at boot and every 60s; mismatch exits the daemon.

  • Enrollment + mTLSlive

    Install token is exchanged for a per-endpoint cert; heartbeats and ingest share the same identity.

  • Audit shipperlive

    Local JSONL is authoritative; batches POST to the console ingest API.

  • Remote rule tuninglive

    Console pushes tuning.yaml; daemon hot-reloads without restart.

  • Multi-tool CLI shimslive

    aws, kubectl, git, terraform, gcloud, psql — one daemon, composed policy.

  • MCP runtime proxylive

    runline-mcp-proxy enforces tools/call for Cursor and other MCP clients.

  • Remote policy publishlive

    Console pushes policy.yaml; agents pull via refresh_policy with SHA256 verify.

Agent CLI

  • runline-agent statusEnrollment, endpoint ID, platform URL, mTLS, live rule count
  • runline-agent classifyHow this shell would be classified (human vs AI)
  • runline-agent doctorPass/fail health checks before presenting
  • runline-agent simulate --as-ai aws s3 rb s3://x --forceDry-run policy without executing AWS
  • runline-agent simulate mcp delete_repositoryDry-run MCP tools/call policy
  • runline-agent demoPrint the full terminal demo script
  • runline-agent policy showEvery loaded rule in evaluation order (embedded + packs)

Policy highlights (0 rules · 0 actor-scoped)

Full catalog

Actor-aware rules fail shut for AI agents; humans may see approve or allow paths.

    Fleet healthAI inventoryAudit logs