Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add commands/openclaw/peekaboo/mcpgit clone --depth 1 https://github.com/openclaw/PeekabooWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/commands/openclaw/peekaboo/mcp)<a href="https://agentmods.dev/commands/openclaw/peekaboo/mcp"><img src="https://agentmods.dev/badge/commands/openclaw/peekaboo/mcp.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00000 | $0.02144 |
| Opus 5 | $0.00000 | $0.01072 |
| Sonnet 5 | $0.00000 | $0.00429 |
| Haiku 4.5 | $0.00000 | $0.00214 |
Grade A, and why
mcp scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 123 lines — stays where its author put it; the contents beside it link to each section on GitHub.
peekaboo mcp
mcp runs Peekaboo as a Model Context Protocol server. peekaboo mcp defaults to serve, so you can launch the server without specifying a subcommand.
Subcommands
| Name | Purpose | Key options |
|---|---|---|
serve |
Run Peekaboo’s MCP server over stdio. | --transport stdio (default); --allow-foreground explicitly authorizes foreground/global UI, including browser connection setup, for this server process; global --bridge-socket <path> attaches to an existing Bridge host. HTTP/SSE names and --port are reserved for future support and currently fail with an actionable error. |
Implementation notes
serveinstantiatesPeekabooMCPServerand maps the transport string toPeekabooCore.TransportType. Stdio is the default for Claude Code integrations.- Public MCP servers are background-only by default. Foreground actions, shared desktop input, and browser connection
setup fail before dispatch unless a human starts that server process with
--allow-foreground. Each server whose catalog includesbrowser, or which consumes an explicit browser handoff, owns a fresh browser child and does not borrow another CLI, daemon, or MCP caller's connection. Scoped MCP page actions never ambiently auto-connect, even with foreground authority.--allow-foregroundexposes the explicitbrowserconnectaction for that exact child and browser routes that can enter Puppeteer evaluation. The pinned provider marks those evaluations as user gestures even for background or headless pages, so default servers hide and pre-dispatch refuse page discovery, snapshots, navigation, waits, element interaction, and arbitrary script evaluation. Foreground-authorized calls reuse the scoped connection and truthfully report foreground browser-protocol delivery; this classification does not claim the page was visibly fronted. Authenticated sessions rejectPEEKABOO_BROWSER_MCP_ISOLATED=1before provider startup because that child has no pinnable browser identity. For headless use, launch Chrome separately and pass its exact loopbackbrowser_url. This explicit authority never exposes Shell, and a nested Agent remains background-only. A background-only Bridge-backed opaque browser session can start connected only through--browser-handoff <absolute-private-path>together with exactly one matching--bridge-socket. The receipt must first be created by an explicit foregroundpeekaboo browser connect --handoff-fileagainst that same socket. A current Bridge consumes the signed receipt once, authenticates its caller, listener generation, exact target, claim, and provider epoch, then gives this MCP server a distinct scoped provider child. The background server starts with that inherited exact connection and does not exposebrowser connector fall back to the Bridge root. Missing, stale, mismatched, consumed, or downgraded handoffs fail before provider dispatch. - Direct-text
pasteis admitted only with an exact generation-pinned app/PID/window authorization and a canonical background result. Targetless, foreground, current-clipboard, and binary paste are refused before dispatch. The nestedagenttool likewise retains immutable background-only authority and never exposes Shell. - HTTP/SSE server transports are reserved but not implemented. Selecting either fails before daemon startup and emits a structured error in JSON mode.
- The MCP process owns its stdio lifecycle and never hosts a Bridge listener. Support stays process-local by default;
an explicit
--bridge-socket <path>uses that existing Bridge host and skips the embedded daemon. - An explicit
--bridge-socketbinds caller-side capture preflight and every later request to that socket's authenticated process generation; unrelated auxiliary sockets cannot freeze the MCP session. The selected host still must advertise process ownership, then enforces the canonical process-lifetime ScreenCaptureKit lease at every SCK leaf by scanning all same-user potential Peekaboo processes. Reentry refuses owner-unaware processes discovered later. Modern refusal remains a signedCAPTURE_FAILED/runtime_incompatible, retry-safe, not-dispatched result instead of being rewritten as a target-attribution error.seecan usecapture_engine: "classic"without entering ScreenCaptureKit. Caller-local MCP keeps the broader startup scan because it has no external Bridge generation to own capture. - The native tool catalog includes bounded
capturefor live screen/window/region recording or video ingest. It writes retained frames,contact.png,metadata.json, and optional MP4 output. Frame/contact metadata carries capture-session-authored SHA-256 custody, projected asartifact_sha256, and success revalidates those bytes plus semantic metadata. Use tool allow/deny filters when exposing MCP to untrusted clients. - An explicit environment allowlist containing only tools proven unable to capture pixels, such as
PEEKABOO_ALLOW_TOOLS=browser, skips ScreenCaptureKit owner selection at startup. Missing allowlists, unknown tools, nestedagent, and any capture-capable tool remain fail-closed behind the normal owner preflight. - MCP resolves its filtered tool catalog before browser bootstrap. Filtering out
browsertherefore creates no browser child and lets browser-free catalogs start against legacy remote providers. An explicit--browser-handoffis still authenticated, consumed, and opened even whenbrowseris filtered out, so one-shot authority is never silently ignored; the filtered browser tool remains unavailable. - UI automation tools include action-first additions:
set_valuedirectly mutates a settable accessibility value, andactioninvokes a named accessibility action on an element fromsee. verify_statereplaces fixed sleeps with bounded native polling. It resolves an app or PID to one exact window, evaluates 1–8 AND predicates for window existence/bounds or exact AX element existence/value/enabled/selected state every 100 ms, and reportssatisfied,unsatisfied, or conservativeunknownafter at most 10 seconds. Explicit PIDs and app-name selectors are pinned to the first resolved PID/process-start generation for the whole invocation; relaunch, PID reuse, and selector drift areunknown. Exact-window ownership is rechecked on every sample and before an optional screenshot, whose capture metadata must confirm the same PID and window ID. A directly read value matching a unique exact AX identifier can satisfy anelement_valuepredicate when unrelated AX siblings are unreadable; missing, mismatched, non-identifier, or ambiguous partial-tree evidence remainsunknown. A WindowServer miss is corroborated with a complete app-scoped window inventory before Peekaboo reports absence, preserving minimized AX windows. Ownership ambiguity, partial enumeration, or identity changes areunknown. It never focuses or replays actions.clickpreserves element IDs and queries when forwarding to automation, so action-first policy can use accessibility actions before synthetic fallback.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 123 lines · 0 tokens per session scan A acdf9ed537e0
mcp is a command published in the GitHub repository openclaw/Peekaboo (5,105 stars, last pushed today), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 2,144 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other commands, from other repositories
upload
Archive, upload, and submit to external TestFlight (full loop with monitor).
release
Create a new Remarc release using the automated GitHub Actions workflow.
imark-review
Open a markdown document in Imark for review and wait for the reviewer's notes.
imark-notes
Read the Imark comments out of a markdown file.
archive
Create xcarchive for TestFlight.
build
Build debug and launch HomeClaw.