You think in ChatGPT or Claude. You ship in Cursor, Claude Desktop, Claude Code, or Windsurf. Threadport captures the thread and compresses it into a structured capsule — intent, decisions, assumptions, weak spots — then serves it to your IDE-side AI over MCP. No copy-paste.
npm i -g @openai/codex ai-capsules tool to load my latest capsule and continue.
get_latest_capsule() approved · 12ms npm i -g @openai/codex…
Every AI tool you use is its own island. Moving work between them costs you the only thing you can’t buy back — context.
Selecting and re-pasting a 40-message thread breaks every time. You lose half the nuance and your wrist.
The next AI doesn’t know what you decided, what you rejected, or which assumptions the prior model relied on.
Whatever the first AI got wrong rides into the next session unchallenged, because nobody flagged it.
Pasting raw chats burns tokens. A structured capsule compresses 85–95% while keeping every decision and constraint.
Threadport ships as a Chrome extension plus a tiny local MCP server.
Both run on your machine. Capsules live at ~/.ai-capsules/capsules.json.
Open ChatGPT.com or Claude.ai, click the Threadport button. The extension reads the visible thread — never your tabs in the background, never your account.
A local LLM call extracts intent, context, decisions, assumptions, weak spots, open questions, and next actions into a compact capsule (~600–1,500 tokens).
Open Cursor, Claude Desktop, Claude Code, or Windsurf. Ask: “Use the ai-capsules tool to load my latest capsule.” Your IDE-side AI pulls it natively over MCP.
A capsule isn’t a summary. It’s a structured judgment object the next AI can reason against — not just read.
1{
2 "title": "Codex Computer Use macOS 14.4.1",
3 "user_intent": {
4 "primary_goal": "Use Codex on Mac without OS upgrade."
5 },
6 "decisions": [
7 { "decision": "Drop OS-upgrade path" },
8 { "decision": "Use Codex CLI w/o Computer Use" }
9 ],
10 "weak_points": [
11 { "issue": "SIP framed as absolute blocker", "should_challenge": true },
12 { "issue": "Storage constraint not pressure-tested" }
13 ],
14 "next_actions": [
15 { "action": "npm i -g @openai/codex" }
16 ],
17 "recommended_handoff_target": "claude"
18}MCP (Model Context Protocol) is the standard way AI assistants in Cursor, Claude Desktop, Claude Code, and Windsurf pull live data from your machine. Threadport ships a tiny local MCP server that exposes your capsules as tools the assistant can call.
{
"mcpServers": {
"ai-capsules": {
"command": "npx",
"args": ["-y", "@ai-capsules/mcp-server"]
}
}
} You: Use the ai-capsules tool to load my latest
capsule, then continue from the next action.
Claude: Reading capsule "Codex Computer Use macOS 14.4.1"…
Decisions: ✓ Drop the OS upgrade path
✓ Use Codex CLI w/o Computer Use
Next action: `npm i -g @openai/codex` and
wire into Cursor session.
Weak spots: SIP-blocks-all-patching framed
absolutely; not actually true. … No accounts. No data resale. Capsules never leave your machine unless you opt into Pro hosted generation.
Bring your own OpenRouter or OpenAI key. Local-only forever.
We run the LLM. Zero API key setup. Premium extraction prompts.
Hosted Pro launches alongside the Chrome Web Store release.
Pro forever, one payment. First 200 founders only.
Reservations open. No charge until Pro launches.
One short email when the Chrome Web Store build is live, and again when hosted Pro opens. No drip campaigns. Unsubscribe in one click.
Store listing is in submission — one-click install when it ships. Until then, clone the repo, run the local server, load the extension, and paste one MCP block into Cursor or Claude Desktop.
{
"mcpServers": {
"ai-capsules": {
"command": "node",
"args": ["/ABSOLUTE/PATH/TO/ai-capsules/mcp-server/dist/index.js"]
}
}
}~/.ai-capsules/capsules.json. The
MCP server reads from that file. Nothing is uploaded unless you
turn on Pro hosted generation, which only sends the chat content
we need to extract a capsule and never persists it.