Distillate
Your research alchemist.
Conjure ML experiments that run themselves.
Distill insights from everything you read.
$ uv pip install distillate
$ cd desktop && npm run install:mac> uv pip install distillate
> cd desktop; npm run install:win
GitHub · Powered by Claude Code
Research should move at the speed of your ideas, not your keyboard. Distillate turns your research questions into autonomous agents. Describe an experiment, and Nicolas, your research alchemist, spawns an auto-research agent to test it. Sync the papers you read, let your highlights spark ideas, test them with autonomous agents, track their progress as they descend gradients and relentlessly improve your north star metric… Rinse. Commit. Repeat. Distillate: science at the speed of thought.
The Laboratory
Spawn research agents. Steer them. Distill results.
Describe a research question. Nicolas sets up a git repo, drafts a prompt, and spawns an autonomous research agent with a time budget.
Adjust goals, write steering instructions, or change direction based on early results.
Compare runs side-by-side, identify what moved the needle, and understand why certain approaches worked.
Distill an experiment's full session history into key findings. Zero API calls — works from local logs.
Quick overview of every experiment — new runs, breakthroughs, stuck agents, and what needs attention.
Turn paper insights into experiment ideas. Replicate results, test a variation, or apply a technique to your domain.
The Library
Read, highlight, synthesize across your collection.
Sync papers from Zotero, extract highlights from any reading device, generate AI summaries, and refresh your notes.
Browse trending research from HuggingFace, get personalized reading suggestions, and add papers to your queue.
Focus on a single paper — extract key methods, results, and implications from your highlights and margin notes.
How It Works
The research loop: read, experiment, learn, repeat.
Save papers to Zotero. Read and highlight on any device — reMarkable, iPad, desktop. Nicolas extracts everything and builds your knowledge base.
Describe a hypothesis or point at a paper. Nicolas sets up a git repo, drafts the prompt, and spawns an autonomous research agent with a time budget.
Every run is tracked with metrics, diffs, and decisions. Rich lab notebooks with research insights land in your Obsidian vault.
Connect paper insights to experiment results. What you read informs what you try. The cycle continues, and knowledge compounds.
Beyond Skills
Dashboards, charts, and email updates built into the workflow.
Track your reading velocity, top topics, engagement scores, and most-cited papers. Available in the CLI, desktop, and weekly email digest.
Every experiment gets a rich lab notebook with run cards, metric sparklines, and AI-generated research insights — saved to your Obsidian vault.
Export clean metric charts as PNG — designed for sharing on X/Twitter or dropping into a paper. One command: --chart
Experiment reports when a session finishes, daily paper suggestions from your queue, and a weekly reading digest. Set up with --email
Cloud & Email
Stay in the loop. Free account, no credit card.
Get experiment reports when a session completes, daily paper suggestions, and a weekly digest. Sign up with distillate --email — toggle each type independently.
Check what’s connected at a glance with --connectors. Configure Zotero, email, Obsidian, and reMarkable individually with --setup
Export your entire library to JSON with --export-state and restore it anywhere. Corrupt state? Auto-backed up as .bak — nothing is silently lost.
Open-source at the core
The CLI powers everything. The desktop app is a window into it.
$ uv pip install distillate
─── ⚗️ Nicolas ──────────────────────────────
Your research alchemist.
🧪 4 experiments · 12 runs · 1 running
📚 42 papers read · 7 in queue
──────────────────────────────────────────────
> /conjure tiny-matmul --duration 30m
🧪 Spawning research agent...
Created distillate-xp-tiny-matmul
Research agent spawned — 30 min budget, will report when done.
> /survey
🔬 Scanning 4 experiments...
✓ tiny-matmul: 8 runs, best loss 0.0023 (-42%)
⚗ sigma-classify: 3 runs, accuracy 94.2% (goal: 95%)
─ rlhf-reward: 1 run, paused
─ tokenizer-v2: no runs yet
Built with love and coffee by Romain Lacombe.