Build Pipeline as Agent Context Window

Frontman does something most AI coding tools get wrong: instead of working from screenshots or pasted code, it hooks directly into the framework’s build pipeline — Next.js, Astro, Vite — so the agent actually understands components, routes, and compilation errors. You click an element in the browser, describe what you want changed, and Frontman edits the real source file. Not a sandbox. Not a copy-paste suggestion. The actual code, hot-reloaded, ready for review.

The architecture is where it gets interesting. The libs/ directory splits into frontman-protocol, frontman-client, frontman-core, and per-framework adapters (frontman-nextjs, frontman-vite, frontman-astro). There’s a context-loader that feeds codebase understanding into the agent, and their docs reference both MCP schemas and what looks like an “ACP” protocol layer (acp-plan.md, acp-toolcalls.md). The client, server, and framework adapters are fully decoupled — they call it an “open protocol” and they mean it structurally.

The split licensing is smart too. Client libraries and framework integrations are Apache 2.0, the server is AGPL-3.0. Encourages adoption of the protocol while protecting the core. The pattern of “intent capture surface → protocol layer → framework-specific execution” is the same shape as a gateway that routes from Telegram or Discord through an event bus to domain-specific handlers. Different domain, same decomposition.

Key Ideas

• Build pipeline as context — hooking into the framework’s compiler gives the agent structural understanding (components, routes, errors) that screenshot-based tools can’t match
• Protocol-first agent architecture — frontman-protocol decouples intent capture (browser overlay) from execution (framework adapter), making the system extensible without rewriting the core
• Context loader as a first-class concern — dedicated context-loader library for feeding codebase structure to the agent, separate from the protocol and client
• MCP/ACP protocol surface — MCP schema and ACP planning docs suggest they’re building toward standardized agent communication, not a proprietary API
• Split open-source licensing — Apache 2.0 for client/adapters, AGPL-3.0 for server — grow the protocol ecosystem, protect the business
• Browser overlay as intent surface — the “click and describe” interaction model is a clean separation of user intent from implementation detail

Links

• Frontman GitHub
• Frontman website
• Frontman docs
• Demo video
• Discord community
• Changelog
• npm: @frontman-ai/nextjs