Inquiso

Chat with any web page. Ask questions, get summaries, check trustworthiness, and let an agent take actions for you — using your own AI keys or your browser's built-in AI, without ever leaving the page.

Inquiso is an open-source, cross-browser extension (Chrome, Firefox, and Chromium-based browsers) that adds an AI sidebar to your browser. It reads the page you're on — or a whole tab group, or every tab in a window — and lets you:

• Ask — "What does this page say about refunds?"
• Summarize — "Give me the 5 key points."
• Assess — "Is this source trustworthy? What's the bias?"
• Act — "Fill this form with my details" / "Find the cheapest option and add it to cart."

You watch the agent think and reason in real time as it browses the page and works toward an answer. Everything runs locally in your browser — there is no Inquiso server, no telemetry, and no place your data is sent except directly to the AI provider you choose.

Beyond completing tasks, Inquiso remembers how they were completed — locally, inspectably:

• Cross-session memory — typed facts, preferences, episodes, and failure lessons with scope, provenance, confidence, and temporal validity; retrieved per task with trust labels and shown live as "Using memory: …".
• Learned workflows — repeated verified tasks become versioned browser skills that are shadow-tested, execute through the normal confirmation gates, degrade when a site changes, and repair themselves into new versions.
• Evidence, not vibes — an append-only event ledger records every action with verified outcomes (URL/element/download evidence); only verified success reinforces memory, and interrupted tasks resume from checkpoints with revalidation.
• Deterministic safety — secrets are never stored, page-planted "memories" are quarantined (0% injection success in the adversarial suite), and memory can never authorize an action: a capability policy engine + confirmation floors always stand between memory and side effects.
• The Memory Center — inspect, edit, confirm, forget, export, or delete everything, including blocked write attempts and per-run measured metrics.

Start at docs/memory-agent/overview.md; the measured results live in evaluation-results.md.

• Bring your own model. Use a free local model (Chrome Built-in AI / Gemini Nano), or bring your own API key for OpenAI, Anthropic, Google, Qwen, or Kimi. You stay in control of cost and data.
• Transparent agent. Every reasoning step and every action the agent takes is shown to you. Sensitive actions require your confirmation.
• Fast. Built on WXT + Vite + React with aggressive, memory-aware caching so the sidebar opens instantly and answers stream immediately.
• Private & secure by design. Manifest V3, strict CSP, least-privilege permissions, encrypted key storage, and first-class prompt-injection defenses.

🚧 Pre-1.0, actively built. Milestones M0–M5 are implemented; the agent and cloud paths need real-browser testing before a tagged release.

pnpm install
pnpm dev            # Chrome with HMR (load .output/chrome-mv3 if prompted)
pnpm dev:firefox    # Firefox
pnpm build          # production builds for Chrome + Firefox
pnpm check && pnpm typecheck && pnpm test   # quality gates

Then load the built extension in your browser and open the sidebar. The step-by-step install, model setup (BYOK), and how to use every feature are in the User Guide.

New here? Start with the User Guide — installing from source, picking a model (free on-device or your own key), asking & acting, autonomy and permissions, attachments, voice, MCP tools, and moving/backing up your chats (export/import and bring-your-own-cloud backup).

Read the design before contributing code:

MIT — see also SECURITY.md and CONTRIBUTING.md.