Touch Browser

What the nangman-infra/touch-browser MCP server does

The nangman-infra/touch-browser MCP server evaluates claims against the content of web pages. Rather than returning only raw HTML or converted Markdown, it opens a page, creates a structured snapshot, and maps each submitted statement to an evidence result. The result can indicate that the page supports the claim, contradicts it, lacks enough evidence, or requires another page.

Each claim outcome can include a confidence band, a review recommendation, support snippets, a verdict explanation, match signals, and citation details. Citations include the source URL and retrieval time. Reusable-claim indicators help an agent determine whether evidence may be passed into a later answer, while unresolved outcomes provide a safer alternative to presenting a weak match as fact.

The tool is evidence-grounded rather than fact-final. It records page-local support and provenance, but a model or human still needs to decide what is true across several pages or beyond the collected sources.

How it works

For MCP use, an agent creates a session, opens one or more pages, and extracts claim results. The documented sequence is tb_session_create → tb_open → tb_extract. Multi-page session synthesis aggregates grouped results from the pages opened in that session.

The extraction response exposes per-call claimOutcomes, along with grouped views for supported, contradicted, insufficient, and more-browsing-needed claims. An agent can first inspect verdict, then check reviewRecommended, read primarySupportSnippet when it needs an evidence line, and follow nextActionHint if the result does not resolve the question.

Browsing is headless-only and uses automatic search-engine selection. If a site presents a challenge, authentication step, or MFA flow, the documented behavior is a supervised recovery handoff rather than blind automation. The project also describes replayable, auditable multi-page research sessions.

Setup and configuration

The primary local MCP installation uses the public npm package @nangman-infra/touch-browser-mcp. A host such as Claude Desktop or Cursor can launch it with npx -y @nangman-infra/touch-browser-mcp. On first launch, the package downloads a matching standalone runtime from GitHub Releases, verifies its SHA-256 file, installs it under the user’s touch-browser npm runtime directory, and starts the MCP service.

The repository also provides standalone macOS and Linux bundles. These include the touch-browser command, the optimized runtime, a Node runtime, and the Playwright adapter. A bundle is installed with its install.sh script. The slim profile includes Playwright Chromium and downloads semantic or NLI models lazily; a full profile can include warm model caches for offline use.

Browser actions use the Playwright adapter by default. New CLI deployments can select the Rust CDP adapter with TOUCH_BROWSER_BROWSER_ADAPTER=cdp-rust; that path still requires Chrome or Chromium, and TOUCH_BROWSER_CDP_BROWSER can specify a nonstandard browser location.

Tools and capabilities

The nangman-infra/touch-browser MCP server supports:

• Creating browser research sessions.
• Opening pages for structured capture.
• Extracting and classifying claim evidence.
• Returning support snippets with stable references.
• Reporting confidence bands and review recommendations.
• Providing source URLs and retrieval timestamps.
• Synthesizing evidence across pages in a session.
• Handling browser-backed actions through Playwright or the optional Rust CDP path.

Use the standalone CLI for commands such as quickstart, doctor, and quick; MCP hosts do not expose the quick CLI command directly. The local diagnostics check the executable, runtime resources, data root, Node runtime, Playwright Chromium, MCP bridge entrypoint, and semantic model readiness.

Limitations and notes

Evidence is scoped to the pages that were opened and the claims submitted. An evidence-supported result does not establish universal truth, and an unresolved result may require another page or human review. Claims marked for review should not be reused without that review or a second verification pass.

The MCP path is intended for public documentation and research web content. Challenge, authentication, and MFA situations use supervised recovery handoff. The Rust CDP browser path requires Chrome or Chromium, while semantic and NLI models may be downloaded lazily depending on the bundle profile.