A local agent that watches your screen and writes its own playbook

What it does

Photo Agents is a Python package that runs a local perceive-reason-act loop to let tool-calling LLMs operate your computer. It bundles browser automation via a Chrome DevTools Protocol bridge, sandboxed code execution, file I/O, and a layered memory system that archives observations, working context, and long-term facts to disk. A multi-provider router switches between Anthropic Claude and OpenAI GPT sessions with an optional failover mixin.

The interesting bit

The memory model is explicitly biological: vision goes in, bound observations get stored in layers, and the agent authors its own L3 SOPs and helper skills after successes. That self-evolution happens through reflection and scheduler scripts in the evolution module, not just prompt engineering.

Key highlights

• Vision-grounded loop: the agent grounds decisions in what it actually sees on screen rather than chat transcript dumps.
• Layered memory: working, global, SOP, and session-archive layers persist to ~/.photoagents/.
• Multi-provider LLM router with native Anthropic and OpenAI support plus a failover mixin.
• Broad client surface: Streamlit, PyQt, desktop companion, and chat-platform bots (Telegram, QQ, Feishu, WeCom, DingTalk).
• Optional observability via Langfuse and a cron-style scheduler.

Caveats

• The entire runtime is gated by a remote-validated API key checked against photo-agents.com; you cannot run it without an account and a live validation call.
• Status is beta and APIs may shift before 1.0.
• The README frames the package as “local” and privacy-preserving, but the mandatory license gate means it is not fully offline or independent of the vendor’s infrastructure.

Verdict

Worth a look if you want a local-ish agent framework with built-in browser automation, multi-provider routing, and an experimental memory model. Skip it if you need a fully offline, ungated open-source tool or if you are allergic to beta APIs that require vendor sign-in.