Nothing to install, just open it
No installation required; a few well-chosen tools you can master at once.
Bi
Features
Bi compresses the tools, boundaries and reversibility an agent needs into one minimal local kernel. Everything below ships in the current version.
Minimal core
Built for your context budget
A per-turn prompt budget; stale tool results collapse into placeholders; history is summarized deterministically by turns.
Symbol-level code navigation
Parse-based indexing of functions, classes and types — more precise than grep, and cheaper.
Documents & artifacts
Upload and preview in place
Images, audio, PDF and Office files preview right in the workbench, without switching windows.
Artifacts in one panel
Generated images, code and reports gather in the artifacts panel, kept in one place.
Web search
The smart-search toggle drives a websearch plugin for live information.
Models & context
Multi-provider routing
Pick a provider per run, with automatic failover when one is unavailable.
Local model support
Point at Ollama, LM Studio and more so sensitive content stays on your machine.
Context compaction
Long sessions summarize deterministically by turns — no LLM calls, stable and cheap.
Prefix caching
Keeps the request prefix byte-stable so provider prompt caches keep hitting.
Trust & rollback
Tool permission approval
Reads, writes and commands are authorized separately; every call requires confirmation by default, with configurable auto-approval.
Intercept rules
A hard gate on specific command patterns; matches reject outright, or can require approval.
Checkpoint rollback
A snapshot is kept each turn, letting you roll back to any step and continue.
The agent loop
Bi is built around an explicit loop: the model reads context, picks a tool, executes, and the result feeds back in. Permissions and checkpoints attach to every step of it.
Interface
