See who's on :3000. And which agent started it.

• Keep a still alongside the video

  Capture the Holdmap view, choose a screen or window, or borrow the active recording source without another picker. Review and save the PNG locally while video keeps running.

  Take a PNG snapshot

• Read the service graph at a glance

  Compact disconnected services, readable starting zoom and a full canvas make local relationships easier to explore. Search services, ports and PIDs while keeping observed neighbors in context.

  Explore the service graph

• Understand each capability

  Explore declared skill purposes, retain the source Map and Compare observed definitions. Purpose, Source, Evidence and Activity keep file presence distinct from loading, approval and use.

  Explore MCP & skills

0.9.0 adds PNG snapshots, the readable service graph and the capability explorer. Read the full 0.9.0 release notes.

• A safe stop, with protection

  Preview the plan, stop the whole dev-server tree (SIGTERM, then SIGKILL) and check the port is free. The OS, shells, IDEs and AI agents are protected.

  

• Why it's busy, in plain English

  A dev-server tree, a container, a system service, AirPlay Receiver, TIME_WAIT or another user.

  

• Project and branch aware

  The project, git branch and framework (Next.js, Vite, Django, Postgres…) or the Docker, Podman or OrbStack container behind each port, with real memory for the whole app.

  

• Everything from the keyboard

  A searchable command palette and predictable arrow navigation. Turn off single-key shortcuts while keeping modifier shortcuts and native controls available.

  

• The service graph

  Which local services talk to which, grouped by Compose, Kubernetes or workspace. Stop a whole stack in dependency order.

  Project stacks
  

• Computer → Footprint → Project

  Zoom from this computer and its agents to their folders, tools, ports and a single source tree. Inspect memory, CPU, parent processes and access, with each fact marked seen, inferred or unknown. Reveal a folder, open it in your editor, or stop only the services an agent started.

  The capability explorer leads with declared skill purposes in Explore, retains source relationships in Map, and shows exact observed definitions in Compare. File presence does not prove loading, approval or use.

  Explore the agent map MCP and skills explorer

  $ holdmap agents
  Claude Code  pid 51200 · 214 MB · 3% CPU · 1 stoppable
    folders   shop-api  ~/code/shop-api
    ports     :3001  shop-api (Vite)
    access    sandbox unknown · network :3001 only
  $ holdmap agents claude --stop-ports --dry-run
  Dry run: nothing was changed.

  

• Record a screen or window

  Choose what to share, then save available MP4 or GIF exports locally. Record for up to two minutes with no microphone or system audio. Close and reopen the latest preview while Holdmap stays open; nothing is uploaded.

  Each new video recording asks for a screen or window. Export formats depend on the encoder available in your app or browser.

  PNG snapshots of the Holdmap view or a selected screen/window. An active recording can supply the still without another picker or stopping the video.

  Screen recording and saving PNG snapshots
  

• See agent connections and recorded calls

  Connections maps observed processes, listeners and live sockets. Calls shows tool and agent activity reported by integrations you configure. Find failures or missing completion, sort by reported duration, and inspect the reported parent context. Optional recording keeps sanitized metadata in memory on this computer; payload fields are excluded.

  Socket relationships do not reveal messages. Session grouping does not imply communication; parent links come from supplied trace context.

  Agent Flow and recording
  

• MCP for AI agents

  Claude, Cursor and VS Code can list ports, inspect agents and tools, find a free port and stop their own services. Stops require a reviewed, one-use confirmation. Protected targets are refused.

  // tools/call parameters
  { "name": "list_agents", "arguments": {} }
  → Claude Code · Cursor · Docker Desktop
  { "name": "stop_agent_ports",
    "arguments": { "agent": "claude",
      "dry_run": true } }

• Scriptable to the core

  Read reports support --json, stop actions support --dry-run, and exit codes mean something.

  $ holdmap list --json \
      | jq '.entries[0].project.name'
  "shop-web"

• A TUI, over SSH too

  Run holdmap for a fast terminal UI. holdmap ssh devbox reads another machine without installing anything there.

  $ holdmap ssh devbox
  connected · 14 ports · linux

• Getting started Install, then list, explain, stop — and see your agents.
• What's new Home as Cards or Map, a richer agent inventory and stopping ports from Home.
• Desktop app Save a snapshot, record your screen, explore maps and review schedules.
• Agents, tools and skills Inspect folders, ports, MCP definitions and skill metadata with their evidence.
• Project stacks Start a project's services in order with .holdmap.toml.
• Shell hook Explain port conflicts after a failed command.
• MCP for AI agents Set up Claude, Cursor or VS Code in one snippet.
• Troubleshooting PATH, hidden ports, first open, uninstalling.
• CLI reference Every command and flag, generated from --help.

• Local inspection

  Scans this computer and probes localhost. SSH connects to hosts you choose; desktop update checks use GitHub Releases. Project commands and SSH configuration can make their own network requests.

• No product analytics

  No Holdmap account or product analytics. This site has no analytics scripts. Optional screen capture and agent activity recording start only when you choose them.

• Credential redaction

  Recognized credentials in command displays are redacted; arbitrary positional secrets can be missed. Default scans do not open agent chats or settings. MCP & skills inspection is explicit: bounded config files may contain credentials, whose values are excluded from results. Only allowed skill-header metadata is retained, not instructions. Activity recording excludes payload fields before retention.

• Protected by default

  Hard protection cannot be overridden. Soft protection needs the CLI's explicit --allow-protected option or a desktop review that permits it. MCP refuses all protected targets.