MCP Browser Control

What is it

MCP Browser Control is the practice of letting an AI agent drive a real, logged-in browser session through the Model Context Protocol (MCP) or a plugin bridge. Instead of scraping with a headless bot that gets blocked, the agent attaches to the browser you already use — cookies, sessions, extensions and all — and clicks, types, and navigates on your behalf. Technically it is a thin protocol layer (MCP server) plus a browser automation engine (Playwright, CDP, or a Rust driver) that exposes tools like navigate, click, read_dom, and fill_form to any LLM client.

The business significance is bigger than the tech. Once an agent can operate a logged-in browser, it can complete real workflows — booking, purchasing, applying, filing, posting — that pure API agents cannot touch because no API exists. That unlocks the long tail of the web that never shipped an API. For indie developers this is a fresh protocol land-grab: the MCP server layer is unowned, standardizing fast, and every vertical (CRM, e-commerce, gov forms) needs its own connector. Projects like Skyvern and moli are already fighting for the agent-browser slot, which tells you the market is real but the winner is undecided.

Why now

Three curves crossed in 2025-2026. First, MCP became the de facto tool-calling standard after Anthropic open-sourced it and OpenAI, Google, and a wave of IDE vendors adopted it — so a browser-control server now plugs into every major client for free. Second, browser automation matured: Playwright and CDP are stable, and Rust-based drivers (like the ones moli and similar projects use) cut latency and memory enough to run agents on a laptop. Third, LLM vision and DOM-reasoning got good enough that agents can recover from unexpected popups and layout changes, which was the wall that killed earlier RPA.

Demand-side, users are tired of copy-pasting between ChatGPT and their browser. They want the agent in the browser, using their session. The 100% growth rate and nascent stage in the data reflect exactly this: the category is being named right now. It is not last year because MCP was not ubiquitous and agents could not reliably see the page. It is not next year because the protocol window — where a small team can own a connector — closes once incumbents ship first-party servers. You have roughly 6-12 months of greenfield.

Market Evidence

The signal here is thin but directional: 2 independent sources (github and v2ex), 3 total mentions, 100% growth rate, stage marked nascent. Three mentions is not a market — it is a spark. Treat it as an early indicator, not proof. The GitHub side matters more than the forum side: repos like Skyvern and moli appearing and competing for the "agent browser" label is a leading indicator, because developers vote with commits, not posts. V2EX chatter confirms Chinese-speaking indie devs are watching the same space, which usually precedes tooling supply.

My position: this is real demand in an early, noisy form. The 100% growth rate off a tiny base is mathematically meaningless on its own — 3 mentions becoming 6 is not traction. What makes it credible is the convergence: two unrelated platforms, two different languages/communities, same week, same concept. That pattern historically precedes a category. The risk is timing — nascent categories can stay nascent for 18 months. Validate by watching whether MCP browser servers get starred and forked weekly, and whether any of them charge money. If nobody is charging in 90 days, the demand is curiosity, not budget.

Who's Behind It

The visible players are small teams and solo builders, not yet whales. Skyvern is the most cited headless-browser agent project, positioning around vision-based web automation. moli is the other name in the summary, competing for the same "let the agent use your browser" slot, and its Rust angle suggests a performance-first play. On the protocol side, Anthropic owns MCP and therefore sets the rules; OpenAI, Google, and IDE vendors are adopters, not owners. That asymmetry matters: Anthropic controls the standard, so any browser-control business is a tenant, not a landlord.

The community layer is indie hackers on GitHub and V2EX, plus the broader "AI agent" crowd on X and Hacker News. No enterprise vendor has shipped a credible first-party MCP browser server yet — that is the gap. The whales to watch are Browserbase, Playwright's Microsoft, and the browser vendors themselves (Chrome/Edge). If Chrome ships native agent control, the indie layer gets squeezed to verticals. For now, the field is open and the incumbents are distracted fighting each other on model quality.

TAM & Market Size

The buyers split into three tiers. Tier one: AI-native developers and indie hackers building agent products who need a browser tool and will pay $20-100/month for a hosted MCP browser server rather than maintain Playwright themselves. Tier two: automation-heavy SMBs — e-commerce ops, lead-gen agencies, recruiting firms — who pay $100-500/month for vertical agents that do real logged-in work. Tier three: enterprises doing RPA replacement, where budgets are $10k-100k/year but sales cycles are brutal.

Size it bottom-up. There are roughly 30-50 million developers worldwide; the slice actively building agents is maybe 2-5 million and growing fast. Capture 0.1% at $30/month average and that is $7-18M ARR — a real indie business, not a unicorn. Price tolerance is set by the alternative: a Browserbase-style hosted browser runs $0.10-0.50 per session-hour, and Playwright maintenance costs a dev 2-5 days/month. So $29-99/month is easy to justify for a solo builder, and $299-999/month for a team. The scores here (opportunity 0/100, demand 0/100) are placeholders from sparse data — do not read them as "no market." Read them as "unmeasured market," which is exactly where indie margins live.

Competitive Landscape

Direct competitors: Skyvern (vision-based headless automation, open source with a cloud tier), moli (Rust, performance-focused), and a scattering of MCP servers on GitHub that wrap Playwright. Indirect: Browserbase and Steel.dev (hosted browser infra), Playwright/Puppeteer (raw libraries), and classic RPA (UiPath, Zapier for the no-code crowd). Big Tech threat: Microsoft owns Playwright and Edge; Google owns Chrome. If either ships native "agent mode" with MCP, the generic layer commoditizes overnight.

The gap is logged-in, local, privacy-preserving control. Headless cloud browsers cannot use your Netflix session, your bank, your internal CRM behind SSO. A local MCP server that drives your Chrome profile solves a problem cloud players structurally cannot. That is the wedge. Competition score 0/100 is unmeasured, not safe — assume 5-10 credible players within a year. Your differentiation must be a vertical or a deployment model (local-first, self-hosted, compliance-friendly), not "we also click buttons." Time before Big Tech enters: 6-12 months for the generic layer, 18-24 months for regulated verticals.

Business Model

Go freemium SaaS with a self-hosted open-core option. The MCP server itself should be free and open source — that is your distribution, because developers install it into Claude, Cursor, or their own agent and it spreads through config files. Monetize the hosted control plane: managed browser sessions, credential vaulting, audit logs, team seats, and prebuilt vertical workflows.

Pricing: Free tier — 1 browser, 100 actions/month, local only. Pro $29/month — 3 concurrent sessions, cloud relay, 5k actions. Team $99/month — 10 seats, SSO, audit logs, shared workflows. Business $499/month — unlimited seats, on-prem option, SLA. Rationale: anchored below Browserbase's usage pricing and above "free Playwright," and priced to a solo dev's card without a procurement call.

12-month forecast. Conservative: 300 paying users, $25 ARPU, ~$90k ARR. Base: 1,200 users, $35 ARPU, ~$500k ARR. Optimistic: 5,000 users plus 20 team plans, ~$2.5M ARR. CAC: near-zero for the open-source funnel (GitHub → install → upgrade), $80-150 for paid ads, which you should mostly avoid early. Payback: under 2 months on the freemium path because acquisition is organic. The open-core flywheel is the whole model — do not gate the server.

MVP Blueprint

Ship in 2-7 days. Core features only: (1) an MCP server exposing navigate, click, type, read_page, screenshot, and wait_for; (2) attachment to an existing Chrome profile via CDP so sessions and logins persist; (3) a confirmation gate for destructive actions (purchase, delete, send); (4) a local session log so the user can audit every action; (5) one-command install (npx or cargo install) plus a copy-paste MCP config block for Claude Desktop and Cursor.

Cut everything else: no cloud dashboard, no multi-user, no credential vault, no visual workflow builder, no vertical templates in v1. Those are v2 monetization, not v1 validation.

Tech stack: Rust driver (or Playwright if you want speed over performance) for the browser layer, TypeScript or Python for the MCP server since both have mature MCP SDKs, CDP for session attach. Rust is the differentiator if you can ship it — lower memory, faster start, and it matches the moli positioning. Fastest path to launch: fork an existing Playwright MCP server, add profile-attach and the confirmation gate, publish to GitHub with a 60-second demo video showing an agent booking something in a real logged-in session. Distribution beats polish. Suggested product types: SaaS, Tool, API — lead with the open-source Tool, monetize the SaaS.

Commercial Opportunities

1. Local-first "agent browser" for privacy-conscious users. Product: a desktop app that runs the MCP server locally against your Chrome profile, with a simple UI to approve actions. Target: developers and power users who refuse to hand sessions to a cloud browser. Revenue: $15-29/month, 500-2,000 users → $7.5k-58k MRR. Why it beats cloud: cloud players legally and technically cannot touch logged-in banking, healthcare, or internal SSO sessions.

2. Vertical agent for e-commerce ops. Product: an MCP-powered agent that logs into supplier portals, marketplaces, and ad dashboards to pull reports and update listings. Target: 1-5 person e-commerce teams. Revenue: $199-499/month per account, 50-200 accounts → $10k-100k MRR. Why it beats generic tools: the value is in the prebuilt workflows and portal quirks, not the browser layer.

3. RPA replacement for regulated SMBs. Product: self-hosted browser agent with audit logs and role-based approvals for finance, legal, and healthcare back-office. Target: 20-200 employee firms with compliance needs. Revenue: $500-2,000/month, 20-50 accounts → $10k-100k MRR. Why it beats UiPath: 10x cheaper, LLM-native, no six-month implementation.

Product Ideas

🥇 LocalBrowser MCP — "Let Claude and Cursor use your real logged-in browser, locally." One-line value prop: an open-source MCP server that attaches to your existing Chrome profile so any agent can act on sites that have no API. Target user: AI-native developers and power users. Why now: MCP is standard, cloud browsers cannot touch logged-in sessions, and no credible local-first player owns this slot. Monetize the hosted relay and team features. This is the wedge — build it first.

🥈 FormPilot — "The agent that fills, files, and submits the web forms you hate." Value prop: a vertical MCP agent for repetitive form work — visa applications, insurance claims, government portals, expense reports. Target user: freelancers, ops staff, small agencies. Why now: these portals have no APIs, are high-friction, and users already trust an agent with the data. Charge $19-49/month per user or per completed submission. High retention because the workflows recur monthly.

🥉 AgentOps Console — "Observability and guardrails for browser agents." Value prop: a dashboard that logs every agent browser action, flags risky ones, and enforces human approval for purchases and deletions. Target user: teams running agents in production. Why now: as browser agents touch money and PII, audit trails become mandatory, and nobody has built the Datadog for this yet. Price $99-499/month per team. This is the picks-and-shovels play — later to market, but defensible once agents go enterprise.

SEO Opportunity

Search volume is small but compounding — terms like "MCP browser server" and "Claude control my browser" are early and rising. SEO difficulty is effectively unmeasured (0/100 in the data), which means low competition. Target long-tail keywords: "MCP browser automation server," "let Claude use my browser session," "Playwright MCP logged-in browser," "local browser agent open source," "Skyvern alternative." Content strategy: ship a comparison page against Skyvern and moli, a "how to add browser control to Claude Desktop in 5 minutes" tutorial, and a GitHub README optimized for the same phrases. Own the category name before it is named.

Risk Assessment

The thesis breaks if MCP fades as a standard, if Chrome/Edge ship native agent control and commoditize the layer, or if the "nascent" stage never converts to budget. Top three risks: (1) Tech — browser vendors lock down CDP or profile attach, or anti-bot systems detect agent traffic and break workflows. (2) Market — enterprises refuse to let agents touch logged-in sessions for compliance reasons, capping you at prosumer spend. (3) Execution — you build a generic layer that Skyvern or a Big Tech first-party server erases in one release.

Validate cheaply before building: publish the MCP config and a demo video, count GitHub stars and installs, and DM 20 developers asking "would you pay $29/month for the hosted version?" If fewer than 5 say yes, walk away. Also check whether any competitor has paying customers within 90 days — if the whole category is free, the demand is curiosity. Walk away if Chrome ships native MCP browser control before you have 100 paying users, or if your CAC exceeds $200 with no organic funnel.

Action Plan

Today: fork an existing Playwright MCP server, add Chrome profile attach and a confirmation gate, and publish it to GitHub with a 60-second demo of an agent completing a real logged-in task. That is one evening of work and it is your entire validation instrument.

Week 1: post the repo to Hacker News, V2EX, and the MCP community; measure stars, installs, and inbound DMs. Set up a waitlist for the hosted version with a price anchor ($29/month) to test willingness to pay. Month 1: if you clear 200 stars and 10 waitlist signups, ship the hosted control plane (session relay, logs, team seats) and onboard 20 beta users free, then convert to paid. Month 3: pick one vertical (e-commerce ops or forms) and ship a prebuilt workflow; target $2k-5k MRR. If by month 3 you have under 50 active installs and zero payment intent, stop and redeploy the skills to a vertical agent on top of someone else's browser layer.

Related Terms

AI Agent tool-calling / MCP ecosystem — the protocol layer this whole opportunity rides on; if MCP wins, browser control is one of its highest-value servers. Headless browser automation (Skyvern, Browserbase) — the cloud-native cousin; competes on scale but structurally cannot do logged-in local sessions, which is your wedge. Computer-use agents (Anthropic, OpenAI) — screen-level control that overlaps with browser control; it is a substitute threat but currently slower and less reliable than DOM-level MCP tools, so it validates the category more than it kills it.