Medium has moved its AI ambitions from the publishing page into the drafting stage. The platform has launched TK, a free notes and writing app that embeds an AI assistant directly inside the editor, letting writers summon help by typing @TK followed by a request. The app went live on November 9, 2026, and is available to anyone with a Medium account.
- AI at the start, not the end: TK places assistance in the first draft rather than in final proofreading, answering quick questions and retrieving previously saved fragments without leaving the page.
- Built-in editorial skills: The assistant offers feedback on drafts and headlines, matches text to the style guides of Medium publications, and suggests writing prompts drawn from older notes.
- Transparency layer: Authors can publish notes explaining how AI was used, an attempt to make the provenance of a piece clear and inspectable.
- Free, web-first: TK runs on the web with a desktop app on the roadmap, imports and exports Markdown, and supports optional connections to a user’s own AI agent.
What the @TK Assistant Actually Does
The name borrows from newsroom shorthand: TK is the placeholder reporters drop into a draft to mark information still to come, spelled with a K so it stands out to proofreaders. In Medium’s app, the assistant is meant to fill those gaps on the spot.
According to the company, TK can handle small lookups that would normally pull a writer into another browser tab, such as how many days a given month contains. It can also pull back paragraphs saved earlier while an idea was still taking shape, and it ships with specific skills for editorial feedback, headline feedback and style-guide alignment. Beyond that, it helps organize existing material and generates prompts to get a stalled draft moving.
For writers already working with agentic tools, TK can optionally connect to their own AI agent through an MCP server or a Claude skill, giving the assistant access to project context, positions and ideas stored in that agent’s memory. The app also imports everything a user has published on Medium, along with vaults from apps such as Bear and Obsidian, and can export single notes or an entire archive as Markdown — though it is not itself a Markdown-native app.
Human Feedback and the Social Layer
TK is not purely a machine-in-the-loop product. Medium has built what it calls social libraries, which function like small private blogs inside the app. Three are open at launch: a public journal for building a daily writing habit, a space for feedback on drafts, and a group for people working toward self-publishing a book. Writers can share early drafts and collect responses from other people, not just the model.
The trade-offs are deliberate. Notes are cloud-based and not end-to-end encrypted, because the app is designed for collaboration and for agents to read; users can set access-control permissions on their work. Medium says no AI company trains on a user’s writing without permission.
The Transparency Bet
The launch lands in the middle of an industry argument. Authors have lost book deals over suspicions of AI use, and literary awards have faced scrutiny after winners were suspected of writing with machine assistance. Medium concedes that AI raises the risk of low-quality content and blurs the line between human and machine authorship. Its answer is disclosure: authors can attach notes explaining how AI contributed, which the company frames as a first step toward making provenance traceable down to the prompting and source material.
Artificial intelligence makes writing more important, not less.
That slogan is the company’s framing of AI as an amplifier of human intelligence rather than a replacement. The more consequential design choice is the audit trail, not the tagline: in an internet increasingly filled with unlabeled synthetic text, documenting how something was written starts to matter as much as what it says.
The risk runs the other way too. An assistant present from the first line can help a writer start, but it can also smooth a distinctive voice into machine-shaped suggestions before the writer has found their own argument. TK’s value will be decided less by its feature list than by whether it produces more deliberate writing or merely more writing — and that depends, as ever, on the person holding the cursor.