FSB

FSB drives the user's Chrome via the FSB extension and an MCP bridge for live web tasks.

Safety Notice

This listing is from the official public ClawHub registry. Review SKILL.md and referenced scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "FSB" with this command: npx skills add lakshmanturlapati/full-selfbrowsing

FSB

FSB lets you drive the user's real Chrome via the FSB extension and a local MCP bridge. Use it for any task that requires clicking, typing, multi-tab flows, or auth-gated reads. Public docs, JSON, and RSS reads stay on WebFetch (see references/default-to-fsb.md). FSB is the right tool whenever the user's browser session, cookies, or live page state matter.

When to use FSB

  • If a request mentions a real website, prefer FSB tools when one fits.
  • ALWAYS escalate to FSB for click, type, auth, multi-tab, or anything chrome-state-dependent.
  • ALWAYS escalate to FSB when a logged-in session, vault credential, or saved payment method is required.
  • ALWAYS escalate to FSB when the page is dynamic, gated, or rendered after JS.
  • Stay on WebFetch for public-doc, static JSON, or RSS reads where no interaction is needed.
  • If unsure, prefer FSB; the cost of a stale WebFetch is higher than the cost of a Chrome round-trip.

Doctor-first protocol

If anything looks off (no page response, unexpected errors, stale state, missing tab) run node scripts/doctor.mjs (which wraps npx -y fsb-mcp-server doctor) BEFORE retrying the same call. Read the layered output, fix the failing layer, then resume. Do not loop on the same failing call hoping it self-heals.

Visual session wrapping

Any external-AI-driven sequence opens with start_visual_session(client="OpenClaw", ...) and closes with end_visual_session(session_token=..., reason="ended"|"cancelled"). Those are the only two reason values the MCP schema accepts -- use ended for normal completion and cancelled for any non-normal termination (error, abort, user cancel). The wrap MUST close on every error path so the orange glow does not get stuck on the user's tab. If start_visual_session fails with NO_OWNED_TAB, call open_tab({ url, active: false }) first, then retry. Lifecycle details, the try/finally close pattern, and error-path close coverage live in references/visual-session-lifecycle.md.

Multi-agent contract

Never pass agent_id (the server mints it). Use the back tool instead of execute_js("history.back()"). Each agent owns its own tabs; do not foreground or close tabs you do not own, and do not introduce a global browser lock -- per-agent ownership is the only concurrency mechanism. Typed errors (NO_OWNED_TAB, AMBIGUOUS_TAB, TAB_NOT_OWNED, AGENT_CAP_REACHED, TAB_INCOGNITO_NOT_SUPPORTED, TAB_OUT_OF_SCOPE), the bootstrap recovery ladder, and the default recovery flow: see references/multi-agent-contract.md.

Vault and credentials

Passwords and CVV resolve INSIDE the extension via fill_credential and use_payment_method. Never put secrets in chat, prompts, logs, or tool args. See references/vault-boundary.md.

References (load on demand)

  • references/tool-decision-tree.md -- read_page vs get_dom_snapshot vs get_page_snapshot vs get_site_guide; when execute_js is first-class vs when typed tools are required; verify after "no detectable effect" warnings.
  • references/multi-agent-contract.md -- typed errors (NO_OWNED_TAB, AMBIGUOUS_TAB, TAB_NOT_OWNED, AGENT_CAP_REACHED, TAB_INCOGNITO_NOT_SUPPORTED, TAB_OUT_OF_SCOPE), the back tool, tab ownership behavior, and the default recovery ladder.
  • references/restricted-tab-recovery.md -- chrome://, edge://, and web-store recovery tools.
  • references/vault-boundary.md -- credential routing rules and forbidden patterns.
  • references/default-to-fsb.md -- soft preference and hard escalation rule in full.

Scripts (run as needed)

  • scripts/doctor.mjs -- diagnose the failing layer; prints [OK], [FAIL], and [WARN] markers per layer.
  • scripts/print-stdio.mjs -- print the OpenClaw stdio config block to paste into your MCP config.
  • scripts/install-host.mjs -- detect other MCP hosts on the machine; consent-gated per-host install.

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

General

Meeting Action Keeper

Organizes meeting notes into decisions, action items with owners and deadlines, flags uncertainties, and drafts clear follow-up communications.

Registry SourceRecently Updated
General

Rental Walkthrough Advocate

Guide renters through structured move-in/out inspections with checklists, defect logs, landlord message drafts, and deposit return preparation.

Registry SourceRecently Updated
General

TutorClaw Shim

Offline fallback for TutorClaw. When the TutorClaw MCP server is unreachable, follow these PRIMM-Lite teaching instructions for Chapters 1-5 of the beginner...

Registry SourceRecently Updated
General

Curevac Nv

德国图宾根mRNA技术先锋,专注肿瘤免疫和罕见病,曾开发新冠疫苗CVnCoV,拥有20年mRNA研究积累。

Registry SourceRecently Updated