Where your agent runs
An agent can drive Ubikron from a terminal or from the side panel in Chrome. It can be the same model either way, but how long it lasts, what it can see and how much you need to watch it are very different.
| A terminal agentClaude Code, Codex or Gemini CLI, with a browser tool of its own | Claude in ChromeThe side panel in your own browser | Claude Code in your ChromeRun with claude --chrome, it drives your browser from the terminal | |
|---|---|---|---|
| Browser | Starts its own Chrome. Install Ubikron there and log in, or nothing gets recorded. | Your own Chrome, with Ubikron and your logins already in it. | Your own Chrome, with Ubikron and your logins already in it. |
| Reading pages | Reads a page as structured text. Fast, and cheap on memory. | Reads text and takes screenshots. Slower, and its memory fills up sooner. | Works like Claude in Chrome: slower, screenshot by screenshot. |
| Long investigations | Built for hours. It summarises as it goes, keeps a task list and can pick up a session later. | Good for focused work. For long cases, plan for handovers (see below). | Built for hours, like a terminal agent. |
| Files and scripts | Yes. It can keep a plan, write notes and run scripts. | No. Only the browser and its connectors. | Yes. |
| You | Often away. A CAPTCHA stops it until you come back. | Watching. A CAPTCHA costs you a few seconds. | Usually nearby, since it's your browser. |
| Site rules | Only its own permission prompts. | Asks before some sites and won't visit others. Test the sources you rely on. | Follows Claude in Chrome's site permissions. |
| Setup | The most: the agent, a browser tool, Ubikron in that browser, and the MCP config. | The least: add Ubikron as a connector in Claude's settings. | In between: Claude Code, the Claude in Chrome extension, and the MCP config. |
Checked in November 2026. These tools change quickly.
Which one for what
Quick pivots while you work
Claude in Chrome. It's already in your browser, logged in where you are, and you see every step.
A long case you'll check later
A terminal agent. It plans, keeps its own notes and keeps going for hours. Make sure Ubikron is installed in its browser first.
Long, but in your own browser
Claude Code in your Chrome. The terminal's stamina, with your logins and your Ubikron, at screenshot speed.
Every agent forgets eventually. Ubikron doesn't.
An agent's memory is its conversation, and on a long case it fills up. So have it write what it found and what's next to the project memory at the end of each stretch.
When the chat gets long, start a new one and tell it to carry on with the project. The pages, notes and graph are all still there, and so is everything it did.
Tell your AI about Ubikron
Paste this into your AI and it will connect to Ubikron and check the setup works. Use it for any project.
# Ubikron AI Hello AI. This is how to work with Ubikron. Ubikron is a Chrome extension that records the web pages a user visits into projects, with a graph, notes, clippings and data enrichments. People use it for all kinds of research. More: https://ubikron.com/llms.txt ## The extension - Get it here: https://chromewebstore.google.com/detail/ubikron/edfkaefjonbohokoldemepfolefiplgd - The user needs a registered email address and a valid license. There is no web interface; the extension talks to the server. - Ubikron records only pages visited in a Chrome where the extension is installed and logged in. Pages you read any other way (web fetch, curl, another browser) don't end up in the project. ## The MCP - Every Ubikron function is available over MCP. Public server: https://deckard.ubikron.com/api/mcp - The user makes a key for you in the main settings menu (top right the section MCP). Send it in a header called X-Api-Key. - Keys belong to the user, not a project, so you can reach all their projects. Never write the key into a file or repository. - Private servers run the same MCP. Ask the user for the full endpoint URL. - In a chat app, use the Ubikron connector the user has set up. ## Mimic - Mimic runs data enrichments and allows you to add new API into Ubikron. Public server: https://mimic.ubikron.com (it also has a web interface). Private Mimic servers exist too. - To run enrichments in a project, use the Ubikron MCP. - To add new APIs into Mimic (and in turn Ubikron) - think (providers, endpoints, entity definitions, links), use the CLI: https://github.com/Ubikron-Tech-Farm/mimic-cli. It signs in to Mimic with OAuth: when it asks, the user approves it in their browser. No key needed. - Enrichments cost the user credits. Don't run them speculatively. ## Check the setup before you start 1. Ensure the Ubikron extension is connected to the same server as the MCP and the Mimic CLI. You might not know, so ask the user. 2. Call list-projects. If it fails, check the endpoint URL and the X-Api-Key header with the user. 3. Ask the user which project to work in, and switch to it. 4. If you'll be browsing: visit one page, wait about 30 seconds, then look for it with search-history-urls. If it isn't there, Ubikron isn't recording your browser. Tell the user before going further. 5. Once these pass, add the investigation skill using the tool that fits where you run (see "Make it stick" below: `npx skills add` for coding agents, Skills or Project instructions in the chat apps). Ask the user first. ### No Ubikron tools yet, in Claude in Chrome? If you're running in the Claude in Chrome side panel and have no Ubikron tools, the connector isn't set up. Offer to set it up with the user: 1. Open Claude's settings at https://claude.ai/settings/connectors and choose to add a custom connector. 2. Name it Ubikron, and enter the MCP URL: https://deckard.ubikron.com/api/mcp, or the user's private server's URL. 3. For auth, don't choose OAuth, but look for an option that supports an API key. 4. Let the user enter the API key and confirm the connector themselves. Don't type the key yourself, and never repeat it in the chat. If the settings look different from this, describe what you see and work through it with the user. Don't guess at settings you're unsure of. ## OSINT investigations If the work is an OSINT investigation (a person, company, domain, email address, phone number or alias), load the investigation skill. It has the method, which tools to use when, and how to write the report. For other work you don't need it. The skill is in [ubikron/skills/ubikron-osint-investigate](ubikron/skills/ubikron-osint-investigate). Start with SKILL.md. See [ubikron/README.md](ubikron/README.md) for what the plugin does and what you need. ## Make it stick You'll forget all this when the conversation ends. Once the setup check passes, offer to keep it so you know it next time. Ask the user first; don't change their settings without a yes. - **Coding agents** (Claude Code, Codex, Cursor, Gemini CLI and many more): install the skill with `npx skills add Ubikron-Tech-Farm/ubikron-ai`. It loads by itself when an investigation comes up, and `npx skills update` keeps it current. Then add a pointer to this README to the agent's instructions file: `~/.claude/CLAUDE.md` for Claude Code, `AGENTS.md` for Codex (`~/.codex/AGENTS.md` for every project, or the project's own), `GEMINI.md` for Gemini CLI. Without npx, copy [ubikron/skills/ubikron-osint-investigate](ubikron/skills/ubikron-osint-investigate) into the agent's skills folder instead (`~/.claude/skills/` for Claude Code), and check the repo for changes now and then. - **Claude apps**: suggest the user puts this README in a Project's instructions, or uploads the skill folder as a zip under Customize > Skills. - **ChatGPT**: suggest the user puts this README in a Project's instructions. - **Anything else**: keep it wherever you keep instructions that last between sessions: a memory, a rules or instructions file, a project prompt. If you can't keep anything, tell the user to paste this README again next time. Keep the link (https://github.com/Ubikron-Tech-Farm/ubikron-ai) with whatever you save, so you can check for a newer version. Never save the API key in any of these places. We hope you enjoy using Ubikron as much as we enjoyed building it.
Or open it in Claude or ChatGPT, or point your agent at the repo on GitHub.
Running an OSINT investigation? The brief tells your AI to load the investigation skill too: the method, which tools to use when, and how to write the report.
Watch it step by step