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

After installing, run one command to see the product difference from a raw fetch:

touch-browser quick https://www.iana.org/help/example-domains \
  --claim "As described in RFC 2606 and RFC 6761, a number of domains such as example.com and example.org are maintained for documentation purposes."

Decision path for the output:

1. Check verdict.
2. Check reviewRecommended.
3. Quote primarySupportSnippet when you need one evidence line.
4. Read verdictExplanation when the verdict is not obvious.
5. Use citation.url and citation.retrievedAt when handing evidence to another system.

Expected shape:

{
  "verdict": "evidence-supported",
  "confidenceBand": "high",
  "reviewRecommended": false,
  "primarySupportSnippet": {
    "snippet": "..."
  },
  "citation": {
    "url": "https://www.iana.org/help/example-domains",
    "retrievedAt": "..."
  }
}

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:

{
  "statement": "The Starter plan costs $29 per month.",
  "verdict": "evidence-supported",
  "confidenceBand": "high",
  "reviewRecommended": false,
  "supportSnippets": [
    {
      "blockId": "b4",
      "stableRef": "rmain:table:plan-monthly-price-snapshots-starter-29-10-000-t",
      "snippet": "Starter | $29 | 10,000"
    }
  ],
  "verdictExplanation": "Matched direct support in 3 page block(s). Review the attached snippets before reusing the claim."
}

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:

{
  "mcpServers": {
    "touch-browser": {
      "command": "npx",
      "args": ["-y", "@nangman-infra/touch-browser-mcp"]
    }
  }
}

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.

Use this package when you want a local MCP host such as Claude Desktop, Cursor, or Codex to attach without a separate manual runtime install.

Standalone Bundle

Tagged v* pushes now build standalone macOS and Linux bundles in the Standalone Release workflow. Each bundle includes:

• bin/touch-browser
• the optimized Rust binary under runtime/touch-browser-bin
• a bundled Node runtime and Playwright adapter
• the default semantic runner scripts with lazy semantic/NLI model download

Browser actions keep the Playwright adapter as the zero-config compatibility default. For new CLI deployments, the main recommended browser engine is the Rust CDP path. Enable it with:

TOUCH_BROWSER_BROWSER_ADAPTER=cdp-rust touch-browser open <target> --browser --session-file /tmp/tb.json
pnpm run fixtures:browser-adapter-parity

The CDP adapter reports browser-backed captures as sourceType: "cdp-rust", reuses persistent browser context directories, applies the search identity profile for search-result captures, and is covered by parity fixtures plus CLI E2E validation for follow, click, type, submit, pagination, expand, iframe, shadow DOM, SPA updates, download clicks, persistent-session actions, and cross-origin nested shadow interactions. It still requires Chrome or Chromium. Set TOUCH_BROWSER_CDP_BROWSER=/absolute/path/to/chrome when it is not installed in a standard location.

The slim bundle profile is now the default release profile. It skips prebundled Playwright Chromium and semantic model caches, then downloads semantic/NLI models lazily on first use. This reduces release asset and model-cache weight. It does not make first launch instant, because the npm path can still download the standalone runtime and install browser dependencies:

pnpm run build:standalone-bundle -- local-dev

Use the full profile only when an offline bundle must include warm model caches and prebundled Playwright Chromium:

pnpm run build:standalone-bundle:full -- local-dev

When a tagged release is published, download the matching tarball from GitHub Releases, unpack it, and run:

./touch-browser-<version>-<platform>-<arch>/install.sh
touch-browser telemetry-summary
touch-browser update --check

To build the same portable bundle locally:

pnpm install --frozen-lockfile
pnpm run build:standalone-bundle -- v0.1.0-rc1

# Then install the bundled command into PATH
./dist/standalone/touch-browser-v0.1.0-rc1-<platform>-<arch>/install.sh
touch-browser telemetry-summary
touch-browser update --check

The standalone path is still the official CLI, operations, offline, and fallback install path:

1. unpack a standalone bundle
2. run install.sh
3. use the installed touch-browser command for every CLI and serve operation

The installer now performs a managed install:

• managed versions live under ~/.touch-browser/install/versions/<bundle-name>
• the active version is ~/.touch-browser/install/current
• the PATH command points at ~/.touch-browser/install/current/bin/touch-browser
• touch-browser update swaps current to a newly verified release bundle
• touch-browser uninstall --purge-all --yes removes the managed install and all stored data

First Run

This is the command-only proof path that matches the installed user experience:

touch-browser open https://www.iana.org/help/example-domains --browser --session-file /tmp/tb-first-run.json
touch-browser session-read --session-file /tmp/tb-first-run.json --main-only
touch-browser session-extract --session-file /tmp/tb-first-run.json \
  --claim "As described in RFC 2606 and RFC 6761, a number of domains such as example.com and example.org are maintained for documentation purposes."
touch-browser session-synthesize --session-file /tmp/tb-first-run.json --format markdown
touch-browser session-close --session-file /tmp/tb-first-run.json

Installed search now keeps an engine-level persistent trust profile by default:

• Google: ~/.touch-browser/browser-search/profiles/google-default
• Brave: ~/.touch-browser/browser-search/profiles/brave-default
• profile state metadata: ~/.touch-browser/browser-search/<engine>.profile-state.json

Contributor Build From Source

Prerequisites: rustup, Node.js 18+, pnpm.

bash scripts/bootstrap-local.sh
cargo build --release -p touch-browser-cli
pnpm run build:standalone-bundle -- local-dev
./dist/standalone/touch-browser-local-dev-<platform>-<arch>/install.sh

Source checkout is a contributor workflow. Release and operations docs still assume the installed touch-browser command from the standalone bundle.

bootstrap-local.sh installs the default semantic models under:

• ~/.touch-browser/models/evidence/embedding
• ~/.touch-browser/models/evidence/nli

Use TOUCH_BROWSER_EVIDENCE_EMBEDDING_MODEL_PATH or TOUCH_BROWSER_EVIDENCE_NLI_MODEL_PATH only when you need to override those default locations.

Why Not Markdown Alone?

touch-browser is not trying to replace every crawler or browser tool. Its job starts after page acquisition.

Need	Markdown-only fetch	touch-browser
Read the page	yes	yes
Keep source-linked block refs	partial	yes
Judge whether the page supports a claim	no	yes
Return contradiction and unresolved states	no	yes
Give support snippets and verdict explanations	no	yes
Tell the agent to escalate instead of answering	no	yes

Why Trust This?

These are the latest local benchmark rerun signals from 2026-05-05, not a promise about every future public page:

Benchmark	Current generated signal
Public web benchmark	4/4 task-proof claims evidence-supported on IANA and RFC Editor pages
Real-user research benchmark	average supported claim rate 1.00 across 3 MCP-driven public-doc scenarios
Adversarial benchmark	verified exact verdict accuracy 1.00, raw exact verdict accuracy 0.80, unsafe auto-answer count 0
Citation metrics	classification, citation, unsupported-claim, and support-ref precision/recall 1.00 across 34 fixtures
Tool comparison benchmark	extract false-positive rate 0.17 vs markdown-baseline false-positive rate 0.33 on plausible negative claims
Evidence-grounded web research benchmark	current seed 61 claim/fixture units with status seed-validated, target corpus 1000 claims

See benchmarks/README.md for the benchmark map and generated report locations.

Product Surface

Primary surface:

• extract: verify claims against the current page and return structured claim outcomes

Supporting read surfaces:

• read-view: readable Markdown for a human reviewer or verifier model
• compact-view: low-token semantic state for agent loops
• search: structured discovery before opening candidate pages

Safety and audit surfaces:

• policy: classify pages and actions as allow, review, or block
• session-synthesize: turn a multi-page session into JSON or Markdown with citations
• serve: expose the runtime over stdio JSON-RPC for MCP or agent integration

What touch-browser Is

• an evidence-first extractor
• a selective prediction surface
• a verifier-friendly routing layer

What touch-browser Is Not

• a universal truth oracle
• a generic crawler replacement
• a guarantee that every unsupported claim is false

MCP Example

Recommended MCP setup for local hosts:

{
  "mcpServers": {
    "touch-browser": {
      "command": "npx",
      "args": ["-y", "@nangman-infra/touch-browser-mcp"]
    }
  }
}

The MCP package is intentionally narrower than the full CLI:

• scope is public docs and research web
• recommended loop is tb_search -> tb_search_open_top -> tb_read_view -> tb_extract
• tb_session_create accepts an optional caller-provided sessionId for external correlation
• tb_open needs target for stateless use; with sessionId, it can omit target to reopen the active tab URL
• tb_read_view, tb_extract, and tb_policy can omit target when sessionId points at an opened active tab
• tb_extract always needs claims
• tb_session_synthesize needs sessionId and at least one opened tab
• long-running MCP calls emit notifications/progress when the host provides _meta.progressToken
• tb_cancel can reset the daemon; use MCP notifications/cancelled for an in-flight request
• engine is not exposed over MCP
• headed is not exposed over MCP
• if the page indicates challenge, auth, MFA, or other supervised recovery, stop and hand off to a human instead of retrying with different browser settings

Alternative MCP bridge setup for an installed standalone command:

{
  "mcpServers": {
    "touch-browser": {
      "command": "touch-browser",
      "args": ["mcp"]
    }
  }
}

The bridge starts touch-browser serve underneath and exposes tools like tb_search, tb_search_open_top, tb_open, tb_read_view, tb_extract, tb_tab_open, and tb_session_synthesize.

Repository checkout integration asset:

{
  "mcpServers": {
    "touch-browser": {
      "command": "node",
      "args": ["integrations/mcp/bridge/index.mjs"]
    }
  }
}

The standalone bundle ships touch-browser mcp and touch-browser serve. The checked-in Node launcher remains a repository integration asset for repo checkouts or container images.

By default the bridge prefers an explicit TOUCH_BROWSER_SERVE_COMMAND, then an explicit binary path, then an installed or packaged touch-browser binary, then repo-local target/{release,debug} binaries. If none are available, it fails fast with an install/build instruction instead of dropping back to cargo run.

Use TOUCH_BROWSER_SERVE_COMMAND if you want to force a specific built binary or wrapper command.

Architecture

Query / URL / fixture / browser tab
  -> browser-first search result parsing
  -> Acquisition
  -> Observation compiler
  -> read-view / compact-view
  -> extract (evidence + citations + optional verifier)
  -> policy
  -> session synthesis / replay
  -> CLI / JSON-RPC serve / MCP

Docs And Proof

• quick start and operations: doc/INSTALL_AND_OPERATIONS.md
• command surface: doc/CLI_SURFACE_SPEC.md
• evidence operating model: doc/EVIDENCE_OPERATING_MODEL.md
• npm MCP package: packages/mcp/README.md
• examples: examples/README.md
• integrations: integrations/README.md
• benchmarks and positioning: doc/README.md
• pilot and operations package: doc/PILOT_PACKAGE_SPEC.md, doc/OPERATIONS_SECURITY_PACKAGE_SPEC.md

License

This repository now uses MPL-2.0.

• commercial and non-commercial use are allowed
• if you distribute modified MPL-covered files, those covered files stay under MPL-2.0
• separate files in a larger work can use different terms
• full legal text: LICENSE
• plain-language policy: LICENSE-POLICY.md