Cotext

Getting started

The five-minute version: install the extension, build your AI personality, react to a few responses, and watch the AI start respecting how you actually want it to behave.

If you live in the terminal or want CLI / MCP access, the developer setup is the right page. If you want to run your own copy of cotext.io, see self-hosting.

What Cotext is, in one paragraph

A browser extension that learns from how you react to AI responses (thumbs up, thumbs down, free-form tags), then writes you a personal preference prompt that gets prepended to every conversation on Claude, ChatGPT, Gemini, Grok, Perplexity, Mistral, and Copilot. Everything runs on your machine; nothing leaves unless you click Publish.

1. Install the extension

Coming soon to the Chrome Web Store. The listing is currently in review. Once it's live, install will be one click. Until then, follow the three commands below — it takes about a minute.

You need Node.js 20+ installed once. Then:

git clone https://github.com/cotext-io/cotext
cd cotext
npm install
npm run build

The build drops a dist/ folder. Load it into Chrome:

  1. Open chrome://extensions
  2. Toggle Developer mode (top-right)
  3. Click Load unpacked
  4. Pick the dist/ directory
  5. Pin the Cotext icon to your toolbar

That's it. The extension is installed.

Works on Chrome 120+, Edge, Arc, Brave, and other Chromium browsers.

2. Build your AI personality

Click the Cotext icon in the toolbar. You land on Create your AI personality. There are three ways to seed your preferences — use any or all:

Free-text description

Just type how you want AI to behave, in your own words.

Terse. No emoji. Push back when I'm wrong. Default to TypeScript unless I ask otherwise.

Trait chips

Twelve preset toggles — Terse, Direct, Warm, No emoji, Step by step, Examples please, Skip warnings, Challenge me, Python default, Formal, Bullets, Plain language, Cite sources. Click to add.

Starter packs

Twelve pre-curated full personalities, one click each. Each pack seeds 10-25 rules and gives you a working profile immediately — useful because reacting to ~20 responses to bootstrap is real friction.

Available packs: Code review, Email writing, Research / analysis, Patient explainer, Strategic thinking, Creative writing, Daily driver, Tutor mode, Brainstorm partner, Sprint planner, Data analyst, Support helper.

Mix and match. A starter pack + a few trait chips + a sentence of description is a great starting point.

When you're ready, click Build my personality →.

3. Pick where it runs

Next step: pick which AI interprets your reactions into rules. This is a one-time setup. Three options, in order of recommendation:

Chrome built-in AI

Uses Google's Gemini Nano, built into Chrome itself. Zero download, nothing to configure. Free and fast.

Requires Chrome 127+ with the experimental flag enabled. If the option doesn't appear in the popup, your browser doesn't support it yet — fall back to On-device AI below.

On-device AI (WebGPU)

Cotext downloads a small AI model into your browser cache and runs it locally on your GPU. Default is Hermes-3 3B (~1.7 GB, takes a few minutes the first time).

After the download, everything runs offline. Works on most modern laptops; if you have a discrete GPU, pick the Pro tier for better quality.

Ollama (advanced)

If you already run Ollama locally, point Cotext at it. Best quality of the three, but you have to install Ollama separately and keep it running.

Decide later

You can skip this step and come back to it. Signals still get captured; they just aren't turned into rules until you pick a provider.

4. Use Cotext on AI sites

Open Claude, ChatGPT, Gemini, or any other supported site. Two things happen automatically:

  1. A small overlay appears in the bottom-right corner of every AI response with 👍 / 👎 / 🏷 controls.
  2. Your synthesized prompt is prepended to every message you send. The AI sees your preferences before it answers.

React to a response

Hover an AI response. Click 👍 if it was good. Click 👎 if not. Optionally pick a tag (too-verbose, preachy, bad-format, etc.) or type a free-form note ("be more concise next time").

Cotext does the rest in the background:

  • Interprets your reaction with the local AI you picked
  • Adds a concrete rule to your profile
  • Re-synthesizes the prompt for the next message

You'll start seeing improvements within a few reactions.

Edit the prompt directly

Click the overlay to open the Cotext panel, then the Prompt tab. The full synthesized prompt is shown as editable markdown. Tweak it directly if the AI's interpretation got something wrong — your edits stick across re-syntheses.

You can also toggle injection per-site if you want certain sites to get the raw AI without your preferences.

5. (Optional) Sign in to cotext.io

If you want to:

  • Save your profile online so you can recover it across devices
  • Share your profile via a public URL
  • Compare versions or roll back to a previous state

…click the identity chip at the top of the popup and sign in with your email. You'll get a magic link; click it and you're done. Generate a username at cotext.io/settings, then click Push in the extension popup.

Your profile lands at https://cotext.io/@<your-username>/<slug>.

Sharing is opt-in and explicit — Cotext never publishes anything automatically.

Forking someone else's profile

Any public profile on cotext.io has a Fork button next to Star. Click it and the profile is cloned into your account — entries, templates, and context come with; the original publisher's user-facts (their name, default language, etc.) are stripped so nothing personal leaks. Your fork lands at @<your-username>/<slug> with a "Forked from @<them>/<slug>" badge linking back.

Not signed in? The same button reads Download and gives you the profile as a <slug>.cotext.json file you can drop into your folder (or pass to a future cotext profile import command). Same content, no account required.

Editing your own profile on the website

Your published profiles also have an Edit profile disclosure on the viewer (visible to you only). Open it and you can change:

  • Display name, visibility (public / private), and category — saved in place with no new version row.
  • Profile name, free-form context, and the rendered prompt markdown — saved as a new content-addressed version, exactly like a push from the extension.

The prompt textarea is the "manual edit" path — the same kind of hand-edit the extension docs point at when you want to override a synthesized rule by editing the markdown file in the folder.

Sync caveat: if the extension or CLI is also syncing this profile, the next push from there can overwrite a web edit. Pause sync (or fork to a web-only profile) if you want your hand-edits to stick.

6. (Optional) Multiple personalities

The default install gives you one profile called "General." Add more from the popup → profile selector → + New. Each profile has its own context tag — signals you record against the "Code review" profile only flow into that profile's rules.

Switch profiles from the selector; the active one is what gets injected.

What's next

  • Terminal user? Developer setup shows the CLI + MCP integration with Claude Code / Codex.
  • Want to self-host? Self-hosting covers running your own cotext.io deployment.
  • Curious how it works? Architecture explains the components, the .cotext folder contract, and the data flow.

Troubleshooting

"Download was interrupted" on WebGPU setup

A model shard fetch failed mid-download. Click Retry download first — the model resumes from the last completed shard. If that fails too, click Clear cache and re-download, which starts from scratch. If both fail, switch to Chrome AI or Ollama.

Signals captured but no rules appearing

Your provider isn't reachable. Open the popup → Advanced → click Test connection. The error message tells you what's wrong (no API key, Ollama offline, model not loaded, etc.).

Personality doesn't seem to reflect my reactions

Open the popup → Advanced → make sure Interpret signals into rules is toggled on. If it's off, reactions are captured but not processed.

Overlay doesn't appear on a supported site

Check chrome://extensions for errors on the Cotext card. Reload the extension card, then refresh the AI site's tab.

For deeper troubleshooting, see the full architecture reference.