Prism

One prompt refracted into three answers, side by side. A single web page fans your prompt out to Claude, Gemini, and ChatGPT by shelling out to each vendor's official CLI, so it runs on your existing subscriptions instead of per-token API billing.

Built on CLIs, not APIs. Prism does not call any model API and stores no API keys. It spawns the official command-line tools (claude, agy, codex), which authenticate with your existing subscription logins (Claude Pro/Max, Google account, ChatGPT). That means zero per-token cost, but also: the three CLIs must be installed and logged in on the machine running Prism, and the responders are coding agents that can use tools (web search, file reading), not bare chat endpoints.

Browser (3 panes + 1 prompt)
   |  POST /ask  -> streamed NDJSON back
Node/Express
   |- spawn  claude -p ... --output-format stream-json   -> pane 1 (live tokens)
   |- spawn  agy --print ... --model "Gemini …"          -> pane 2 (Antigravity)
   |- spawn  codex exec --json ...                       -> pane 3

Features

Asking

• One prompt fans out to all three models; answers stream side by side.
• @Claude / @Gemini / @Codex (or @ChatGPT) targets only those models; no mention uses your default models (Settings).
• Drag & drop (or paperclip) attaches text, code, images, and PDFs.
• Markdown and LaTeX (MathJax) rendering; Enter sends, Shift+Enter newlines.

Two modes (chosen per chat)

• + New chat -> Standard: the classic three side-by-side panes, full answers, one row per prompt. ✓ on an answer makes the next prompt build on it.
• + New tree chat -> Tree of thoughts: a branching node graph (below).
• A chat keeps the mode it was created in; a single chat is never both.

Consensus mode (Standard chats, checkbox under the prompt box)

• All models still answer, then one model you pick (Consensus by …) reads their answers and writes a single combined response, shown on top.
• The individual answers are hidden behind a Show the N individual answers toggle, so you read one coherent answer unless you want the originals.
• The judge can be Claude, Gemini, or ChatGPT; the choice is remembered. The combined answer is saved with the chat like syntheses and critiques.

Tree of thoughts (tree-mode chats)

• The chat is drawn as a node graph: a dark prompt node is connected by lines to its answer cards, and a selected answer is connected to the follow-up it spawned. The edges show which answer each branch grew from. Node positions are laid out automatically; children sit under their parent.
• Each answer is a compact card (tinted) with a one-sentence summary the model writes itself.
• Click an answer to open it (full text, white) and select it; click again to close it. With an answer selected, the next prompt makes all models build on that one answer, converging into three new cards under it.
• Click a prompt node to select it (and open its answers); the next prompt then has each model continue its own answer in that group, giving three parallel parent-to-child lines.
• With nothing selected, a new prompt continues from the last answers the same way (three parallel lines, each model from its own).
• Other answers are kept. Select any answer or prompt later and ask again to explore a different branch, so the conversation grows into a tree.
• A node or prompt that already has follow-ups can't be branched again; delete its descendants first (× on the follow-up prompt) to ask a different question there.
• ⤢ on a card opens/closes it without selecting; Expand all / Collapse all open or close every card. New answers open automatically while they stream. The layout re-flows around the larger nodes.

Comparing the models (Compare answers / Criticize)

• Compare: a semantic synthesis of a prompt's answers (consensus, differences, unique points). Criticize: the other models critique every answer, so each answer shows the others' critiques of it, each labeled by the model that wrote it. The ! icon does the same for a single answer.
• Both open in a popup. In Standard mode, closing the popup also shows the result inline (synthesis below the answers, the others' critiques under each card); Tree mode keeps it popup-only.
• Syntheses and critiques are saved with the chat.
• Revise button (next to Send): hidden until an answer has critiques. Click it and each critiqued model rewrites its own answer using the other models' critiques, with no prompt from you, appended as a new turn. Each model's previous answer and the critiques of it are passed in its context; otherwise critiques stay display-only.

Per-card tools (expanded card header, icons)

• Copy the answer as markdown; criticize just this answer.
• Regenerate: re-run that one answer in place, with the same prompt and context, replacing it. Blocked if the answer already has follow-ups branched from it (delete those first).
• Humanize: rewrite an answer to remove the AI tone (Korean text); a Settings toggle applies it to every response automatically.

Chats

• Auto-saved history in the sidebar: rename (double-click), delete, deep-link via ?conv=<id>.
• Temporary chat mode: nothing is saved.
• Deleting a prompt (× on hover) removes it, its answers, and the whole subtree branched from them.

Settings & Skills

• Default models, base model per service (opus/sonnet/haiku or custom), font family/size, and theme (light / dark / system, follows the OS).
• Personalization: standing instructions (like custom instructions / CLAUDE.md) sent to every model with every prompt.
• Enable/disable each CLI's skills from the Skills panel: Claude and Codex skills, plus the Gemini pane's Antigravity (agy) plugins.
• Update banner when the GitHub repo is ahead of your running version; the server auto-restarts when a CLI binary updates.

Prerequisites

The three CLIs must be installed and logged in (each uses its own subscription auth):

Pane	CLI	Install	Log in
Claude	claude	Claude Code	already signed in
Gemini	agy	Antigravity (antigravity.google)	via Antigravity
ChatGPT	codex	brew install codex / npm	codex login

Check: claude --version, agy --version, codex --version should all print.

Note: Google ended the free Gemini CLI tier for individuals, so the Gemini pane now runs through Antigravity (agy), which serves Gemini models. Pick the exact model in Settings → base model (default: Gemini 3.1 Pro High).

If agy isn't signed in, the Gemini pane returns a short "not signed in" note instead of hanging on its interactive sign-in prompt. Run agy in a terminal (no arguments) to sign in, then retry.

Run

npm install
node server.js        # http://localhost:3000

Type a prompt, press Enter (Shift+Enter for a newline). Each box streams its model's answer. Past chats are saved automatically and listed in the left sidebar.

Run at startup (macOS)

A launchd LaunchAgent starts the server at login and restarts it if it crashes, so http://localhost:3000 is always available. The agent sets PATH/HOME explicitly because launchd's default environment is too minimal to find the three CLIs.

# install / reload
launchctl load -w ~/Library/LaunchAgents/com.dongwookim.aggregationai.plist
# status (pid, last exit code)
launchctl list | grep aggregationai
# stop + remove from startup
launchctl unload -w ~/Library/LaunchAgents/com.dongwookim.aggregationai.plist
# logs
tail -f server.log server.err.log

After editing server.js, restart it with: launchctl kickstart -k gui/$(id -u)/com.dongwookim.aggregationai

How it works

• server.js spawns the three CLIs in parallel, each with its working directory set to a throwaway scratch dir ($TMPDIR/aggai-scratch) so the agentic CLIs never touch real projects. Each model's parse() pulls plain text out of one stdout line:
  • Claude: stream-json deltas (stream_event -> content_block_delta -> text_delta)
  • Codex: --json events (item.completed where item.type === "agent_message")
  • Gemini: agy --print (Antigravity) plain stdout, forwarded as-is
• The browser streams the response as newline-delimited JSON and appends each chunk to its pane.

Notes

• These are coding agents, not chat endpoints. For a plain question they just answer, but they can use tools. Codex runs --sandbox read-only; Claude/Gemini run in the scratch dir.
• To add or reorder models, edit the MODELS map in server.js. The UI builds its panes from /models.