Lumen

Category: OSS Agent

License: MIT (open source, free)

Official site: https://lumen.omlabs.xyz

GitHub: https://github.com/omxyz/lumen

npm: @omxyz/lumen (current version: 0.2.0)

What it is

Lumen is a vision-first browser agent with self-healing deterministic replay. Developed by Jina AI as their underlying browser agent, it operates by taking screenshots of the browser state and feeding them into a multimodal model rather than using DOM manipulation or CSS selectors.

Key capabilities

• Vision-only perception loop: Screenshots feed into the model on every step; no DOM traversal or CSS selectors required.
• Multi-provider support: Works with Anthropic, Google, OpenAI, and any OpenAI-compatible endpoint.
• Deterministic replay: Caches and replays action sequences without re-invoking the model; self-heals when page structure changes.
• History compression: Drops old screenshots and uses LLM summarization to stay within token budgets on long tasks.
• Streaming: agent.stream() yields typed StreamEvent objects for real-time UI integration.
• Safety gates: Domain allowlist/blocklist policies, confidence gates for uncertain steps, and CDP-based post-action verification.
• Browserbase integration: Supports remote cloud browsers via the type: "browserbase" browser option in addition to local Chrome.

Benchmark

On a subset of 25 WebVoyager tasks across 15 sites (scored by Gemini 2.5 Flash, 3 trials per task):

Metric	Lumen	browser-use	Stagehand
Success Rate	25/25 (100%)	25/25 (100%)	19/25 (76%)
Avg Steps	14.4	8.8	23.1
Avg Time	77.8s	109.8s	207.8s
Avg Tokens	104K	N/A	200K

Pricing and safety

Lumen itself is MIT-licensed and free. You supply your own LLM API key (Anthropic, Google, or OpenAI). No ToS caveats specific to Lumen; standard provider usage policies apply.