v3.2.0 — Multi-model consensus and deep-research orchestration over your own browser

Auto_Browser v3.2.0 — Multi-model consensus and deep-research orchestration over your own browser

Auto_Browser is an MCP server that drives Claude, ChatGPT, and Gemini through their web UIs in parallel — a real Chrome session over the DevTools Protocol, using the subscriptions you already pay for. It runs cross-pollinated consensus rounds between the models, and as of this release, orchestrates multi-provider deep-research batches end to end: submit a set of research prompts, route them across providers under quota limits, harvest the reports (including the tricky ones — see below), and synthesize a final report per task from all providers' outputs.

Highlights

Deep-research pipeline (new). Batch submission with per-provider routing, a persisted quota ledger (e.g. Gemini's daily deep-research cap), a headless queue runner that survives restarts, and spend-safety as a design invariant: a task that sent is never re-sent — paid research runs are never doubled, and chat URLs are preserved so a completed report can be re-harvested for free.

Report harvesting that handles real provider UIs (new). ChatGPT renders deep-research reports inside a sandboxed cross-origin iframe — invisible to naive DOM scraping; Auto_Browser extracts them frame-aware. Claude's full reports live in an expandable Document artifact; harvesting captures the full text, not the preview. Gemini's plan-confirmation step is cleared automatically. All three providers were verified generating and fully harvesting their reports live during development on the tested accounts.

Per-task model selection (new). Every send resolves an explicit model — per-call → per-task → configured default — and never silently inherits whatever the tab last used. Model pickers are discovered live from each provider's UI, verified after selection, and an unavailable model degrades to a typed warning + default, never a guess. Testing pins each provider's cheapest model by design.

Consensus with a structural verdict protocol. Rounds end with line-anchored VERDICT: AGREE|DISAGREE votes: consensus needs ≥2 unhedged AGREEs and zero DISAGREEs among models that succeeded; dissent is sticky across a dissenter's timeout, peer verdict lines are stripped before cross-pollination, and failures are quarantined — an error string is never presented to another model as a peer's answer. Oversized peer text is compressed before embedding.

Operational hardening. Crash-safe atomic state with corrupt-file quarantine, single-flight run guard, per-model and per-call response timeouts, per-round login gating (an expired session fails fast as login_expired instead of burning a timeout), Chrome auto-launch with the profile and state kept outside the repo, and a stats CLI for quotas, latencies, and batch status.

Tested how

A 22-file Chrome-free unit suite runs in CI (Node 20/22). Live end-to-end gates — parallel-send isolation, consensus, cold-start, double-start, crash-boot, deep-research flows per provider — were run against the real sites during development from the gate scripts in mcp-orchestrator/scripts/e2e/; they require your own logged-in Chrome and never run in CI.

Honest limitations

• This automates provider web UIs. It is not affiliated with Anthropic, OpenAI, or Google, and browser automation may be restricted by each provider's terms of service. Use your own personal, paid accounts, at your own risk.
• Provider UI changes can break selectors until the registry is updated (~/.auto-browser/registry.json overrides, no code change needed).
• Deep research spends real quota and money; the queue is built to respect caps and never double-spend, but the spending is real.
• Platform status: macOS is live-proven end to end. Windows and Linux Chrome paths are configured but currently untested.
• Never commit or share your Chrome profile directory — it contains your logins.

Install

Node ≥ 20. git clone → cd mcp-orchestrator → npm ci → npm test, then point Claude Desktop (or any MCP client) at server.js — config snippet in the README. Chrome auto-launches on first connect; sign into your providers once in that window.

Support

Bugs and feature requests via the issue templates; security concerns via GitHub's private security advisories (see SECURITY.md) — please never include profile paths, cookies, or session data in public issues.

Roadmap

Prompt-injection fencing of peer outputs · API fallback driver · new providers as registry descriptors · Windows/Linux re-validation.