NeuralStack

Divergent-thinking whiteboards for Claude. NeuralStack lets Claude open a shared thinking board on a hard, open-ended problem, hand it to you to brainstorm on, then read the board's structure back and converge to an answer it couldn't have produced from the bare prompt.

Most AI ideation fails the same way: it converges too early — the first plausible answer, polished. NeuralStack fights that twice. The engine fans your question out across many cognitive frames (first-principles, red-team, inversion, biological mechanism, extreme-scale, market lens, naive outsider…), scores and clusters the branches, and keeps the survivors. Then you add the angles, hunches, and domain knowledge a model never would. Claude converges from what you actually built.

The whole value is the delta: the board-informed answer minus the answer you'd have gotten cold.

The baton-pass

You ask Claude → Claude spawns a board → you get a live link
       → the board auto-diverges across frames
       → you brainstorm on the canvas (add, branch, prune)
       → you say "done"
       → Claude reads the board's structure and converges

Human and model never edit at the same time — it's a relay, not a shared mouse.

Install (one command — skill-only, no MCP, no per-client config)

Tell Claude (or any agent with a shell) to run:

npx -y neuralstack@latest install-skill

That installs the neuralstack skill into ~/.claude/skills/ and prints instructions that walk the agent through finishing setup with you, right there in chat:

1. API key — Claude points you to https://thinklm.vercel.app (sign in → Settings → API key → Generate); you paste the tlm_… key and it runs login for you.
2. Mode — Claude asks how you want to think:
   • Frontier — the agent itself reasons, on your own Claude/Codex quota (frontier quality, no separate bill)
   • Default — NeuralStack's free hosted models
3. Done. Say "use neuralstack to brainstorm X" and it spawns a board.

Prefer to set it up by hand? npx -y neuralstack login tlm_xxx then npx -y neuralstack mode hosted|byom.

How Claude drives it (the CLI)

| command | what it does | |---|---| | npx -y neuralstack status | { key_set, mode_set, mode, url } | | npx -y neuralstack login <tlm_key> | save your API key (once per machine) | | npx -y neuralstack mode [hosted\|byom] | read/set who reasons: hosted (NeuralStack) or byom (your agent) | | npx -y neuralstack create "<topic>" | spawn a board → prints { id, mode, preview_url } | | npx -y neuralstack add <id> --frame "X" --summary "…" --full "…" | push an agent-generated branch (byom) | | npx -y neuralstack score <id> | run the cloud critic over the branches | | npx -y neuralstack board <id> | print the board: tree + surviving insights + critic verdict | | npx -y neuralstack wait <id> | poll until you mark the board ready, then print it | | npx -y neuralstack answer <id> "<text>" | write the converged answer back onto the canvas |

NEURALSTACK_API_URL defaults to the cloud; set it to http://localhost:3100 for local development. The key resolves from --key → $NEURALSTACK_TOKEN → ~/.neuralstack/config.json.

What makes a board different from a chat

A prompt gives Claude words. A board gives it shape — which angle branched from which, what the critic clustered together, which leaves survived vs. got flagged as traps, the non-obvious-but-viable pick, the open provocation, and everything you added by hand. That structure is the raw material Claude converges from, and it's what a flat summary throws away.

Links

• Web: https://thinklm.vercel.app
• Source: https://github.com/UditAkhourii/neuralstack
• npm: https://www.npmjs.com/package/neuralstack

MIT licensed.