Tutorial reel
Long-form video tutorials for every part of lilbee, split into two tracks: drive it yourself in the TUI, or put it behind an agent over MCP. Each clip is the demo from the homepage, expanded with a longer caption. Videos are silent; play at your own pace.
Drive it yourself · full terminal app
Getting started
What is lilbee?
Index lilbee's own README, then ask what is lilbee in one sentence? and get a cited answer drawn straight from the source.
First run
The wizard pulls a chat model and an embedder in parallel; you can start using lilbee while they download.
TUI tour
A one-minute sweep through every TUI screen.
Command surface
Ctrl+P for the palette, ? for the cheat sheet, /help for the slash-command catalog.
Ask your library
Chat with cited answers
Ask the Crown Vic manual. Every answer cites a page; inline [N] markers open a source preview at the exact passage.
Leave, come back, keep going
Conversations save automatically. ctrl+o opens the drawer, a few letters filter, enter resumes a 240-message conversation. It no longer fits the model's context window, so lilbee folds the older turns into notes the model keeps. It marks the spot in the transcript where its view begins, then answers the next question with the page cited.
Add files
/add <path> ingests a file. The Task Center shows live progress; ask once it lands.
Crawl a URL
/crawl <url> fetches a page or a small site into your library, then it's chat with citations like anything else.
Crawl a whole site
Leave recursive crawling on and /crawl follows the links at depth 1 and indexes the whole site. Here it crawls hundreds of Wikipedia pages (fast-forwarded). Then one multipart question synthesizes across them, cited, on Qwen3-8B and a reranker.
Models & providers
Model catalog
Browse Hugging Face Hub inside lilbee. Tab between roles, toggle grid / list, search, open model info, pull live.
Use a model from Ollama
Already running Ollama? Point lilbee at it. The catalog labels the model "ollama" with no native models present; index the Crown Victoria manual on camera, then get a cited answer.
Use a model from LM Studio
Same flow with LM Studio: lms ls shows the model, the catalog labels it "lm studio", and lilbee answers from the manual with a citation.
Unsupported architectures, surfaced before the download
lilbee tags GGUFs the bundled llama.cpp can't load with an unsupported pill before you pull. Press f to hide them. The clip searches for gemma-4, an architecture upstream hasn't shipped yet.
Tune it
Settings
Tabbed editor for every knob: Models, Ingest, Generation, Retrieval, Display, Crawling, API-Keys, System.
Multi-GPU placement
Automatic placement
With more than one GPU, lilbee spreads the models across all of them on its own: the chat model is tensor-split across every card so it fits, and the embedder is copied to each. It then answers a grounded question while the per-card load bars move live.
Manual placement
Press ctrl+g for the placement drawer and pin each role to the cards you want. Each role shows what it is: a chat model split across cards, an embedder copied per card, a reranker on a single card. Preview the fit, apply, then chat against that exact fleet.
Put it behind an agent · MCP
It tunes itself
Agent fine-tunes lilbee mid-conversation
Turn 1 answers in outline. The user asks for source; the agent fine-tunes lilbee's retrieval, re-searches, and answers in full function bodies with file.py:L<start>-L<end> citations. Runs on a local 8B model.
Four agents on one local model
One model serves as many agents as you want to run. Four work at once here against Qwen3 Coder Next: 45GB across three RTX 4090s, at 138 tokens a second. Each works in its own clone of this repository through lilbee.
The same, on a thinking model
Gemma 4 26B reasons before it answers, and lilbee streams that reasoning as visible text. Every pane shows the model working through the problem -- reading files, second-guessing itself, checking the code -- before it commits to an answer.
Qwen3.6 27B, thinking out loud
The same four tasks on Qwen3.6 27B: each agent states what it expects first, then reads the code to check itself.
One agent on Devstral 2 123B
The biggest model in the set: 70GB across two A100 80GB cards. One agent gets the model's full 14 tokens a second -- a model this size trades pane count for depth.