Vane
Facts
- Openness
- Open source
- Licence
- MIT
- Language
- TypeScript
- Pinned version
9e8d883(2026-04-09)- Topics checked
- 29 of 33
- Kind of agent
- General framework
What it does notably well
- Loop discipline
Vane picks the research loop's iteration bound from the request mode (speed 2, balanced 6, quality 25) and runs a bounded indexed for loop to that bound.
- Retrieval quality
In Vane's quality mode a web search skips the cosine stage: an LLM picker chooses at most 3 result URLs, each is scraped with Playwright and extracted with Readability, split…
Vane puts all research findings and widget outputs into the writer system prompt inside <search_results> and <widgets_result> tags, the writer prompt requires a [number] citation…
- Stopping
Vane's research loop ends in three ways: the iteration cap (loop bound by mode), a turn with no tool calls, or a done call; all three then return the findings gathered so far.
- Streaming
Vane emits newline-delimited JSON block frames (block, updateBlock, researchComplete, messageEnd) over a TransformStream, with same-process reconnect replay through POST…
- System prompt
Vane assembles its answer front-loaded: all search findings and widget outputs are gathered first and placed as <search_results> and <widgets_result> inside the writer's system…
See how Vane does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with Vane.