Medium’s new TK app brings AI assistance directly into the writing process

Medium released TK on Monday, a new writing app that embeds an AI assistant inside the editor so writers can summon help without leaving their draft. The tool is free for people with a Medium account and is available on the web at launch, with a desktop version listed on the roadmap. Users call the assistant by typing "@TK" followed by a request or prompt, and TK responds in the context of the piece being written.
The name borrows from newsroom shorthand. In publishing, "TK" marks a spot where material still needs to be added; it stands for "to come," with a K substituted for the C because the letter pair is rare in English and therefore easy for a proofreader to spot. Medium repurposes that placeholder convention as a helper for filling in the blanks rather than flagging them.
According to Medium, TK can answer questions a writer would otherwise open another browser tab to look up, such as the number of days in a given month, and it can pull in fragments the user previously saved to the app while developing ideas. It ships with specific skills for editorial feedback, headline feedback and matching a draft to the style guide of a Medium publication. Beyond in-editor help, the app helps writers recall and organize earlier work and suggests writing prompts.
TK can optionally connect to a person's own AI agent through an MCP server or a Claude skill. MCP, the Model Context Protocol, is an open standard for exposing tools and context to AI models; an agent connected this way could give TK a broader picture of a writer's ongoing projects, stated positions and stored ideas. A Claude skill, in the same vein, would let the app hook into Anthropic's Claude assistant. This design treats the writer's external agent memory as context the editor can draw on rather than as a separate tool.
There is also a human side. Medium says the app adds a social layer where writers can share early drafts and collect feedback from other people, not just the model.
On format and privacy, TK imports and exports Markdown but is not itself a Markdown app. Work in TK is not encrypted because the product is built for collaboration, though users can set access-control permissions on what they write.
Medium is addressing AI provenance directly. Writers can attach notes to a piece explaining how AI was used. In its launch blog post, the company described this as a beginning of what it believes publishing in the AI era requires: making the provenance of writing clear, transparent and inspectable, down to the prompting and source material. That is a notable design bet. Rather than trying to police AI use, TK tries to disclose it at the document level, which in practice means readers see a declaration of assistance instead of inferring it.
The launch lands in a contested climate. Authors have lost book deals over suspicions of AI use, and literary awards have faced scrutiny when winners were suspected of writing with AI assistance. Part of the difficulty is that AI detection tools are themselves unreliable, since models learn from human writing; distinguishing assisted from unassisted prose is not a settled technical problem. Medium's stated position is that these nuances will ultimately stop mattering. The company argues that artificial intelligence makes writing more important, not less, and that AI can amplify human intelligence with the written word as its foundation. At the same time, Medium acknowledges the risk of low-quality content and of blurring the line between human and AI writing, which is what the provenance notes are meant to counterbalance.
For developers, the MCP server and Claude skill are the most interesting surface. They suggest an editor that treats retrieval and context as pluggable, letting a user's existing agents act as a knowledge layer. The trade-offs are familiar: unencrypted collaborative storage and permission-based access mean teams should think about what drafts contain before inviting collaborators, and the value of agent integration depends on how much a person already stores in an agent's memory. The free tier tied to a Medium account also anchors TK to Medium's existing readership and publication ecosystem, including the style-guide matching feature, rather than positioning it as a standalone tool.
MCP is the Model Context Protocol, an open standard that lets AI models and assistants call external tools and read external context; for a writing app it is a way to reuse work already captured elsewhere. Style-guide matching draws on the conventions individual Medium publications publish, so the assistant's suggestions are shaped by house rules as well as general editing.
Why it matters: writers and publishers get an AI helper that lives in the draft and a disclosure mechanism for how it was used, while developers get an MCP server and Claude skill to connect their own agents to a writing environment. The likely effect, by inference, is that AI assistance becomes a documented, inspectable part of the publishing workflow rather than something authors hide, though whether readers and editors accept provenance notes as sufficient remains unproven. Product teams building writing tools should note the pattern: in-editor invocation, pluggable context and document-level transparency, all surfaced at launch without encryption in favor of collaboration.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Medium's new writing app embraces AI-assisted writing, despite growing controversy over the technology's role in publishing.