Detailed Analysis
A developer operating under the handle alexvilelabah has released Bah, an open-source, MIT-licensed browser that positions itself as a free alternative to Perplexity's Comet browser. Like Comet, Bah is built around an AI agent capable of executing multi-step tasks directly on web pages rather than simply summarizing or chatting about content. The stated differentiator is accessibility: while Comet is free to download, its agent's multi-step task capability is rate-limited on the free tier, pushing users toward a $20/month Pro subscription. Bah, by contrast, ships with a keyless AI provider (Pollinations) baked in, meaning the agent works immediately after installation with no signup, API key, or credit system required. Users who want more power can add a cheap API key (DeepSeek is cited as costing mere cents per task, with Mistral and NVIDIA NIM also supported), or run everything locally through Ollama for full privacy. Notably, the developer states the browser was built "essentially all" using Claude Code, Anthropic's agentic coding tool, making this a concrete example of Claude being used to construct a consumer-facing AI product rather than just backend scripts or internal tooling.
The technical approach is notable for how the agent perceives and interacts with web pages: it reads the accessibility tree rather than relying on screenshot-based vision models, which is generally faster, cheaper, and more precise for structured interaction with page elements like buttons, forms, and links. This mirrors a broader industry shift among agentic browsers and computer-use tools toward accessibility-tree parsing as a lighter-weight alternative to pixel-based screen interpretation. Beyond core task automation, Bah bundles privacy and productivity features that are increasingly expected in modern browsers: built-in ad and tracker blocking via EasyList, EasyPrivacy, and uBlock filter lists, automatic stripping of tracking parameters (like fbclid and gclid) from URLs, document upload and Q&A for PDFs and Word files, background page-monitoring with change notifications, and locally-transcribed voice input. Built on Electron, React, and TypeScript, the project is a fairly ambitious full-stack application for what appears to be a solo or small-team effort.
This release is significant less for its polish (the developer notes the macOS build is still experimental) and more for what it represents: the democratization of agentic browsing capability that major AI labs and well-funded startups like Perplexity are trying to monetize through subscription gating. By using a no-cost inference backend and supporting ultra-cheap or fully local alternatives, Bah undercuts the monetization model that companies like Perplexity, OpenAI, and others are betting on for agentic browser products. This reflects a familiar pattern in software history — proprietary paid products spawning free, open-source clones once the underlying techniques (accessibility-tree navigation, LLM-driven task planning, tool-calling agents) become well understood and cheaply accessible via commodity or open-weight models like DeepSeek.
For Anthropic and Claude specifically, this project is a small but telling data point in the growing narrative around Claude Code as a tool that enables individual developers to build complex, multi-technology applications (agentic AI, browser automation, Electron desktop apps, privacy tooling) largely through AI-assisted development rather than large engineering teams. It reinforces the broader trend of AI coding assistants compressing development timelines for ambitious software projects, while also illustrating how quickly capabilities pioneered by well-funded AI companies (like Perplexity's Comet) get reverse-engineered and redistributed as free, open alternatives — a dynamic that could pressure commercial AI browser products to reconsider pricing and feature-gating strategies as the technical barrier to building competent agentic browsers continues to fall.
Read original article →