Up Next
LLM Synthesis. Select a track or "All completed papers," choose a synthesis mode, and LitFlow will assemble your structured notes into a prompt and send it to your configured model. Scratch notes are excluded by default — they're your scrap paper, not your analysis. Your API key never leaves your browser.
⚠️ Screen sharing? Pause before running synthesis — your API key is visible in browser DevTools Network tab. Use a local model (Ollama) if you're in a shared environment.
✦ Synthesize Your Literature
Your synthesized output will appear here. Configure your API settings first (⚙ Settings), then select a source and mode above.
Per-paper synthesis: Open any individual paper card in the All Papers view to access per-paper AI actions — summarize the paper in one paragraph, generate a "what should I read next" recommendation, or ask a custom question about that paper's notes.
⚙ Project & AI Settings
Your Research Project
Personalizes the "Relevance to ___" note field across all paper cards.
Reading Tracks
Define up to 5 custom tracks for your literature. Each track gets a name and color.
LLM Provider
Do I need an API key?
Yes — and your Claude Pro or ChatGPT Plus subscription ($20/mo) does not include one. Those are chat-only plans. The API is a separate billing account:
· Anthropic (Claude): console.anthropic.com → create account → add credits. $5 minimum; a typical synthesis run costs ~$0.01–0.03, so $5 lasts months.
· OpenAI (GPT): platform.openai.com → same deal, separate from your ChatGPT Plus account.
Yes — and your Claude Pro or ChatGPT Plus subscription ($20/mo) does not include one. Those are chat-only plans. The API is a separate billing account:
· Anthropic (Claude): console.anthropic.com → create account → add credits. $5 minimum; a typical synthesis run costs ~$0.01–0.03, so $5 lasts months.
· OpenAI (GPT): platform.openai.com → same deal, separate from your ChatGPT Plus account.
Free alternative — Local models (Ollama): No API key, no cost, fully offline.
1. Install Ollama from ollama.com
2. In a terminal:
3. Select Local (Ollama) below — Base URL defaults to
4. Model name must match exactly what's running on that address — open localhost:11434/api/tags in a new tab and copy a
5. Model size vs. GPU memory: models much bigger than ~13–14B parameters often don't fully fit in a consumer GPU's VRAM. When that happens, Ollama offloads part of the model to CPU, which can make generation 10–50x slower — a large model can look "hung" when it's actually just grinding away for several minutes. Run
1. Install Ollama from ollama.com
2. In a terminal:
ollama pull llama3 (or mistral, phi3, etc.)3. Select Local (Ollama) below — Base URL defaults to
http://localhost:11434/v14. Model name must match exactly what's running on that address — open localhost:11434/api/tags in a new tab and copy a
"name" value from the list into the Model field. (Don't trust ollama list in a terminal for this — if you also have Ollama running inside WSL/Docker, the terminal command and your browser can reach different Ollama servers and show different model lists.)5. Model size vs. GPU memory: models much bigger than ~13–14B parameters often don't fully fit in a consumer GPU's VRAM. When that happens, Ollama offloads part of the model to CPU, which can make generation 10–50x slower — a large model can look "hung" when it's actually just grinding away for several minutes. Run
ollama ps while a request is in flight to see the GPU/CPU split, and prefer a smaller model for interactive use.
⚠️ Screen share / shared workstation notice: When you run a synthesis, your API key is sent directly from your browser to the provider (Anthropic, OpenAI, or your local endpoint) as a visible request header. Anyone watching your browser's Network tab in DevTools can see it. Do not run synthesis while screen-sharing in a lab meeting or on a shared machine. On local models, no key is transmitted.
🤖 Claude
Anthropic
Anthropic
💬 GPT
OpenAI
OpenAI
🏠 Local
Ollama / LM Studio
Ollama / LM Studio
Stored only in your browser's localStorage. Sent directly to your chosen provider — never to any LitFlow server (there isn't one).
Claude: claude-sonnet-4-6 · OpenAI: gpt-4o · Ollama: llama3, mistral, etc.
Used for Standard synthesis, Batch condense, and each Consensus agent call. Too low truncates the output mid-sentence. There's no single "correct" number — model limits change too often for LitFlow to prescribe one, so set this yourself: rule of thumb — estimate your expected output length in words, multiply by ~1.3 for a rough token count, then add 30–50% buffer (e.g. a 600-word literature review section ≈ 780 tokens → set 1,000–1,500). Check the "Max out" column in the reference table below for the model's real ceiling. Claude-specific note: LitFlow sends Claude requests as a single non-streaming call, so values above roughly 16,000 risk the request stalling or timing out rather than failing fast — if you genuinely need Claude to produce that much output, raise the Request timeout below substantially, or keep this under ~16,000.
Used only for the final merge step of Consensus/multi-agent synthesis. This step shares its token budget with the model's internal reasoning, so it needs more headroom than a single-shot call — too low can produce empty or cut-off output. Raising this may also require raising the Request timeout below. The same ~16,000-token non-streaming caveat for Claude (see above) applies here too.
📊 Model token-limit reference
| Model | Context | Max out (ceiling) | Consensus | Notes |
|---|
Sources: Anthropic model docs ·
OpenAI model docs ·
Ollama model library ·
Ollama API docs.
For Ollama, "context window" is the model's architectural maximum — the server's actual running context (num_ctx) often defaults smaller unless configured, which limits input size independently of the output-token settings above.
No "recommended Synthesis tokens" column is shown here on purpose — that number depends on your expected output length, not the model, and a hardcoded per-model guess would go stale as limits change. Use "Max out" as the ceiling and the rule of thumb above the Synthesis field to pick your own value.
If the model hasn't responded within this many seconds, LitFlow gives up and shows an error instead of waiting forever. Raise this for very large local models; lower it to fail fast while testing.
Data
Disk persistence — saves your library to a real file on your computer that survives clearing browser data. Requires Chrome or Edge 86+.