Agent Tooling Browser Capabilities Matrix — 2026

Executive synthesis

The browser-agent landscape splits into five practical buckets.

1. Integrated agent workbenches: Hermes, Codex, Claude Code, and ChatGPT Agent/Atlas combine model/tool orchestration with browser affordances. Hermes is the broadest provider-agnostic research workbench. Claude Code’s Chrome integration is the strongest captured option for real logged-in browser QA/debugging. Codex’s app browser is intentionally scoped to unauthenticated public/local/file pages.
2. Open local browser control: Playwright MCP, Browser MCP variants, and BrowserBash are the best free or BYOM candidates. Playwright MCP is the deterministic baseline; Browser MCP variants are attractive for real Chrome-profile reuse but require serious permission review; BrowserBash is useful when agents need a reviewable browser-flow verdict.
3. Cloud browser infrastructure and frameworks: Browserbase/Stagehand, Browser Use, and Steel provide managed browser runtime, sessions, recording, proxy, stealth, and captcha features. These claims are plan-sensitive and should be rechecked before purchase.
4. Production workflow/RPA browser agents: Skyvern and MultiOn target multi-step business workflows. Skyvern has the clearest artifact, credential-vault, profile/session, file-block, and human-interaction surface in the captured docs.
5. Consumer AI browsers and agents: ChatGPT Agent/Atlas, Perplexity Comet, and Gemini/Mariner are worth monitoring and can be useful personally, but they are weaker Hermes integration targets unless a developer API, MCP endpoint, or local automation bridge exists.

Hermes already has a credible research/browsing stack: search, extraction, JS-rendered browser actions, screenshots/vision, console inspection, file/terminal/wiki workflows, skills, cron, memory, profiles, MCP support, and configurable local/cloud browser backends. The main captured browser-specific gaps are file download/upload support, richer first-party network/CDP tools in the live schema, persistent authenticated browser-profile reuse, and native anti-bot/proxy/captcha infrastructure.

Recommendations for Hermes

Keep Hermes as the default research workbench

Hermes should remain the first tool for source gathering, synthesis, and knowledge-base publication because it can move from web search to rendered browser inspection to file/wiki output without leaving one agent runtime. Dedicated browser products rarely include that curation and workflow layer.

Standardize two free add-ons first

1. Playwright MCP should be the default free/open browser-control MCP. It fills deterministic browser automation gaps around file upload, persistent storage state, console/network inspection, tracing, video, PDF, and cross-browser control.
2. BrowserBash should be added as an agent-callable verification harness. Markdown tests, NDJSON --agent output, exit-code verdicts, and optional screenshots/video make it well suited to “prove this web flow works” tasks.

Pay only when the task needs paid infrastructure

• Use Browserbase/Stagehand when Hermes needs reliable managed browsers, session replay/debugging, contexts, proxies/stealth/captcha, and MCP/SDK integration.
• Use Skyvern for repeatable business-process browser automation with credentials, profiles, downloaded/uploaded files, workflow blocks, human-interaction steps, and audit artifacts.
• Use Steel when Chad/Hermes wants to build its own browser-agent control loop but outsource cloud browser infrastructure, proxies, captcha solving, and live viewing.
• Use AgentQL when the bottleneck is resilient structured extraction or semantic selectors across changing pages, especially with Playwright.
• Treat Browser Use Cloud as promising but require current license/pricing verification and a small hands-on test before relying on it.

Suggested operating defaults

• Read-only research: start with web_search and web_extract; switch to the Hermes browser when extraction returns a shell, the page is JS-rendered, or exact rendered state matters.
• Local web-app QA/debugging: use Hermes browser for quick checks; use Playwright MCP for deterministic flows, uploads, storage, console/network artifacts, traces, video, or PDF; use BrowserBash when a CI-friendly pass/fail artifact is required.
• Authenticated personal accounts: only use real Chrome/Edge profile reuse when the user explicitly authorizes the trust boundary. Claude Code Chrome or a reviewed Browser MCP/local-CDP route are the captured options.
• Anti-bot, proxy, or CAPTCHA-heavy targets: do not pretend open local tools solve this natively. Evaluate Browserbase, Steel, Browser Use Cloud, Skyvern, or AgentQL/Tetra based on whether the need is infrastructure, full workflow automation, or extraction.
• Enterprise repeatable workflows: evaluate Skyvern first, then Browserbase/Stagehand or Steel if the team wants more control over the agent loop.

Capability bucket notes

Hermes versus dedicated browser products

Hermes is strongest when the goal is not just to browse but to produce an auditable artifact: a ResearchHub report, wiki page, code change, scheduled digest, or multi-agent Kanban handoff. Dedicated browser products are stronger at narrow browser mechanics such as persistent sessions, file transfer, anti-bot/proxy/captcha tiers, or enterprise workflow artifacts.

Local browser control options

Playwright MCP is the safest default because it is open, deterministic, and purpose-built for automation. Browser MCP variants that control a real logged-in Chrome profile are powerful but should be treated like giving an extension broad access to the user’s personal browser. BrowserBash is complementary: it turns natural-language web-flow checks into reviewable tests and machine-readable verdicts.

Cloud infrastructure options

Browserbase/Stagehand and Steel are infrastructure layers. They are not the whole agent brain; they provide browsers, sessions, debugging, proxies, captcha, and compatibility with automation frameworks. Skyvern is closer to a workflow/RPA system with business-process primitives and audit artifacts.

Unresolved verification gaps

• OpenAI Codex Chrome-extension details outside the in-app browser docs, and richer current ChatGPT Agent/Atlas FAQ details, remain thin.
• Browser Use’s current license, cloud pricing, and exact file/console/network support need direct verification.
• MultiOn pricing and whether Agent V1 Beta remains current or broadly available need verification.
• Project Mariner’s current availability and capabilities need stronger official source-of-record evidence.
• Browser MCP variants beyond Microsoft Playwright MCP should be installed and tested before relying on README claims, especially for CAPTCHA/human fallback, network tools, and file upload/drop.
• CAPTCHA, proxy, stealth, and anti-bot claims are vendor- and plan-sensitive; recheck before buying or promising them operationally.