← Ghostlight
Ghostlight · Field guide

Choose the right browser context.

AI tools can work in an isolated browser, an existing signed-in session, a managed cloud browser, or a capability already included with an AI plan. Those choices change which work is possible, where data travels, what your team owns, and how easily you can change direction later.

This decision aid turns your environment into an options-considered memo. It is for developers choosing an MCP surface, local users deciding how much browser access to grant, and teams evaluating public, internal, or vendor-provided tools.

01
Begin with intent

What interests you?

Choose anything you want the comparison to pay particular attention to. These are priorities, not hard requirements; the questions below establish the real boundaries.

Development
Operations
Organization
02
The load-bearing distinction

Which AI channels do you already use?

Browser capability can be bundled with one account and absent from another channel carrying the same model. Select every channel that belongs in the real environment.

AI vendor and deployment channels
AnthropicDirect plans, cloud channels, and API access.Choose
OpenAIChatGPT plans, OpenAI API, or Azure OpenAI.Choose
GoogleGoogle AI plans, Gemini API, or Vertex AI.Choose
Local or otherSelf-hosted models and other API providers.Choose
In-house toolingA browser MCP surface your organization owns.Choose

Why this matters: direct claude.ai plans include Claude in Chrome; Bedrock, Agent Platform, Foundry, and Console/API channels do not. That difference can create a browser-capability gap without changing the model your teams use.

Where must browser control work?
03
Name the boundary

What must be true in operation?

Preferences shape the comparison. Policy answers remove options outright.

Add organization detailsOptional · regulation, license basis, and planning horizon
Regulatory or review context

Select only contexts that materially affect browser data, evidence, or processor review. Leaving this blank means no special context was supplied.

04
Enough to form a view

Your current direction

This is a working recommendation, not a verdict. The sections below let you refine scale, cost, and every assumption when the decision needs more rigor.

Current best fit

Gathering context

Why this fits

The recommendation will respond as you describe the environment.

Copy-paste brief

Decision summary

Complete the essential questions to create an options-considered summary.

05
Optional rigor

Refine the scale.

The recommendation already has a useful shape. Add the number of people and likely browser workload when you want more realistic planning totals.

Estimated browser users30
Tune the inference mixOptional · planning defaults are already applied
Inference mix for browser work

Simple navigation and form work may not need frontier reasoning. Use the mix you would actually permit or operate.

Current mix: 100%. Costs normalize the mix if it does not equal 100%.

06
Optional rigor

Cost, if it matters here.

Planning estimates are already applied. Refine them when the decision is close or when you know the rates that govern your organization.

Planning estimates are shown for context only.

Tune the estimateOptional · licensing, inference, operations, and isolation
Licensing and inference per browser user / month
Operations and isolation per month
07
Four legitimate operating models

Options compared

Every option keeps its advantages and risks visible. Cost is one input to fit, not an automatic winner.

Scenario AViable

Bundled first-party

Use browser capability included with a direct AI plan, with the vendor's controls and support path.

Annual
—
Horizon
—
Per browser user
—

Advantages, tradeoffs, and math
Advantages
  • The shortest supported path when the right plan is already held.
  • Vendor administration, product integration and a support relationship.
Risks
  • Browser policy and capability follow the selected vendor and account channel.
  • Separate accounts may add a processor and split governance from an existing cloud deployment.
Choose inputs to calculate this option.
Scenario BViable

Governed local and portable

Use Ghostlight with the inference channels and MCP clients that already fit the organization.

Annual
—
Horizon
—
Per browser user
—

Advantages, tradeoffs, and math
Advantages
  • Works in existing authenticated Chromium context across MCP clients.
  • Local, portable browser policy and per-action evidence survive a model change.
Risks
  • The organization owns updates, policy care and incident response.
  • Published support acknowledgments are not the same operating model as a bundled first-party vendor SLA.
Choose inputs to calculate this option.
Scenario CViable

Existing or ungoverned

Continue with an internal or community bridge while treating governance and evidence as organizational responsibilities.

Annual
—
Horizon
—
Per browser user
—

Advantages, tradeoffs, and math
Advantages
  • Low license cost and a fast path for contained evaluation.
  • Internal tools can fit a known workflow precisely.
Risks
  • Authenticated actions may lack independent policy and incident evidence.
  • Maintenance, browser drift and prompt-injection exposure remain with the adopting team.
Choose inputs to calculate this option.
Scenario DViable

Dedicated or isolated

Use Playwright, DevTools or managed browsers in a dedicated profile rather than a person's daily session.

Annual
—
Horizon
—
Per browser user
—

Advantages, tradeoffs, and math
Advantages
  • Strong separation from unrelated personal browser state.
  • Well suited to repeatable, unattended, parallel and cross-platform automation.
Risks
  • Existing SSO, passkeys and user-mediated sessions may not be available.
  • Cloud variants introduce another data processor; local variants still need secure operation.
Choose inputs to calculate this option.
Decision pointBundledGoverned localExisting toolsIsolated
08
Leave a reviewable trail

Assumptions and questions

Defaults are dated, sourced, and visibly separated from your overrides. Verify commercial numbers against the agreements that actually govern your organization.

AssumptionValueSourceBasis
Questions for the bundled vendor
  • Which of our deployment channels include browser control under our agreement?
  • What policy, audit, retention, and administrative controls apply to browser actions?
  • What changes if we move inference to a cloud provider or another model vendor?
Questions for Ghostlight
  • How are trained browser schemas tracked when upstream interfaces change?
  • What does the Continuity Promise guarantee if active maintenance stops?
  • Where does injection containment end, and what detection is deliberately not claimed?
  • What support or SLA arrangements are available beyond the published acknowledgment commitments?
Questions for your own team
  • Which workflows genuinely require a person's existing session?
  • May attended agents act in user context, and who can stop them?
  • Which SIEM or evidence store receives browser-action records?
  • Who owns browser-agent incidents, upgrades, and uncertain outcomes?
View the dated claim register
  • CLM-001Claude in Chrome is available on paid claude.ai plans.As of 2026-07-15 · HIGH volatility
  • CLM-002Claude Code browser integration supports Chrome and Edge, not other Chromium browsers, WSL or mobile.As of 2026-07-15 · MED volatility
  • CLM-003Claude in Chrome requires a direct Anthropic plan and is unavailable through Bedrock, Google Cloud Agent Platform or Microsoft Foundry without a separate claude.ai account.As of 2026-07-15 · HIGH volatility
  • CLM-004Claude Desktop can route Chat, Cowork and Claude Code inference through Amazon Bedrock.As of 2026-07-15 · HIGH volatility
  • CLM-005Computer Use is excluded from Claude Desktop on Bedrock because it requires Anthropic-hosted inference.As of 2026-07-15 · HIGH volatility
  • CLM-006Claude subscription features, including the Chrome extension, are not reachable with a Console API key or a third-party provider.As of 2026-07-15 · HIGH volatility
  • CLM-060Playwright MCP is open source, supports dedicated or isolated browser profiles, and explicitly is not a security boundary.As of 2026-07-15 · LOW volatility
  • CLM-061Chrome DevTools MCP collects usage statistics by default and provides an opt-out flag.As of 2026-07-15 · MED volatility
  • CLM-062Playwright automates Chromium, Firefox and WebKit, while its patched Firefox and WebKit builds are not the installed branded Firefox or Safari browsers.As of 2026-07-15 · MED volatility
  • GL-V1-001The comparison uses Ghostlight's intended V1 operating profile rather than transient release-readiness status.As of 2026-07-15 · MED volatility
  • GL-V1-002Ghostlight intended V1 requires a Chromium browser version 116 or newer.As of 2026-07-15 · MED volatility

The link includes your selected priorities, vendors, browser estate, and governance posture. It can disclose information about your technology environment.