nangman-infra/touch-browser

Ask a claim. Get page-grounded evidence, verdicts, and citations.
touch-browser is an evidence verification layer for AI agents. It does more than fetch a page or convert HTML to Markdown. It opens a page, compiles a structured snapshot, and tells you whether the current page supports a claim, contradicts it, or still needs more browsing.
Use it when you need:
- source-linked evidence instead of raw HTML dumps
- support snippets and verdict explanations that an agent can inspect before answering
- a safe unresolved path for borderline claims instead of bluffing
- policy-gated browsing instead of blind automation
- replayable, auditable multi-page research sessions
Evidence-first, not fact-final:
- touch-browser helps an AI collect page-local evidence and trace where it came from
- a higher-level model or human still decides what is true across pages or across the wider world
Start Here: 60 Second Proof
Before the first browser-backed run, check the local runtime:
quickstart shows the safe local first-use flow, success signals, and common pitfalls. doctor verifies the installed executable, runtime resources, data root, Node runtime, Playwright Chromium, MCP bridge entrypoint, and semantic model readiness. If it reports attention-required, repair those local dependencies before judging browser behavior.
After installing, run one command to see the product difference from a raw fetch:
Decision path for the output:
1. Check top-level status.
2. Check summary.reusableClaims and reuseSummary.allClaimsReusable.
3. Check each claimOutcomes[].verdict.
4. Check claimOutcomes[].reviewRecommended.
5. Quote claimOutcomes[].primarySupportSnippet when you need one evidence line.
6. Use claimOutcomes[].citation.url and citation.retrievedAt when handing evidence to another system.
Expected shape:
The important signal is not that the page was fetched. It is that the claim was routed into a verdict, a confidence band, a reusable snippet, and a source citation.
MCP hosts do not expose the quick CLI command directly. For MCP, use the same idea as a tool loop: tb_session_create -> tb_open -> tb_extract.
What extract Returns
Abbreviated claimOutcome shape from the current extractor:
The extractor returns four verdicts:
- evidence-supported: the current page surfaced usable support
- contradicted: the current page surfaced conflicting evidence
- insufficient-evidence: the current page did not provide enough direct support
- needs-more-browsing: the current page is not specific enough yet, so the next step is another page
confidenceBand, reviewRecommended, supportSnippets, verdictExplanation, and matchSignals are there so an agent can decide what to do next without blindly trusting the first match.
Decision Tree: Reading a claimOutcome
First-time users should ignore most fields at first and use this order:
1. Read verdict.
2. If reviewRecommended is true, do not reuse the claim without a human or second-pass verifier.
3. If you need one quote-like evidence line, use primarySupportSnippet.
4. If you need the reason for the verdict, read verdictExplanation.
5. If the verdict is not enough to answer, follow nextActionHint.
Use matchSignals only when you are debugging or comparing evidence quality. It is not the first field to read.
claimOutcomes is the per-call decision log for every submitted claim. evidenceSupportedClaims, contradictedClaims, insufficientEvidenceClaims, and needsMoreBrowsingClaims are status-grouped views of the same extraction result. session-synthesize aggregates those grouped views across opened tabs in a session.
MCP Package
Primary local-host MCP path:
- npm package: @nangman-infra/touch-browser-mcp
- scope: public docs and research web
- MCP contract: headless-only, automatic search-engine selection, supervised recovery handoff for challenge/auth/MFA
Recommended host config:
On first launch, the package downloads the matching standalone runtime bundle from GitHub Releases, verifies the published .sha256, installs it under ~/.touch-browser/npm-mcp/versions/, and then starts touch-browser mcp.