Happy Friday friends

This week the assistant moved out of its own tab and into the tools where the work happens. Anthropic put a Claude teammate inside Slack channels. Epic wired Claude into the Unreal editor. A solo founder open-sourced a video editor an agent can run. And the team behind Semaphore raised to hand agents the keys to production. The thread through all of it: AI is most useful sitting inside the software you already open, not in a window you have to remember to visit.

For operators, the Claude Tag story is the one to try this week. For anyone building in 3D or video, the Unreal and Palmier releases turn your pro tools into something an agent can drive. The Semaphore founders' round points the same idea at infrastructure, and the tools at the bottom are the rest of the week's install-today list.

On June 23 Anthropic shipped Claude Tag, an AI teammate that lives inside Slack channels. You @-mention it like a colleague, it breaks a request into steps, reaches into connected tools like Drive, Gmail, and your codebase with permission, and posts the result back into the thread. An ambient mode follows the channel and jumps in on its own with a summary or a reminder. Anthropic says 65% of its product team's code now comes from its internal version of this.

The shift is where the assistant sits. A chatbot in a separate tab is a tool you have to remember to open; a teammate in the channel is part of the workflow you already run. One Claude per channel means the whole team shares its context and its memory, so it stops asking you to re-explain the project every morning. The unit of AI work moved from your private chat to a shared room.

It is in beta for Claude Team and Enterprise, and the old Slack app retires August 3, so admins get a 30-day migration window. Drop @Claude into one ops or research channel this week and hand it a recurring multi-step task, then judge whether ambient mode earns its place or just adds noise. Decide which channels you want it learning from before you grant it your Drive.

Air Design used to spend hours every day manually calling their 600 members to schedule seasonal tune-ups.

They turned on Podium's AI Membership Coordinator. It contacted 471 members, booked 187 jobs, and generated $24,000 in revenue.

Across home services, the story repeats.

Magnolia Plumbing cut invoice-to-payment time to 6 minutes and saved 60 hours of admin work every month.

This is what Podium's AI Operating System does: phones answered, jobs booked, invoices collected — automatically, without adding headcount.

If your schedule isn't packed this summer, it can be.

See How It Works

On June 17 at Unreal Fest Chicago, Epic shipped Unreal Engine 5.8 with an experimental Model Context Protocol plugin that connects Claude, Gemini, or Codex straight into the editor. The model does not paste suggestions into a side panel; it places assets, scripts behaviors, adjusts lighting, and generates whole scenes by reading and operating the project directly. Epic kept the choice of model open and the human in the chair. It is the final major UE5 release, with native AI arriving in UE6.

Same move as the Slack teammate, in a different room. Claude stopped being something you copy from and became an operator inside the tool, here a 3D engine instead of a chat thread. For a studio, that compresses the slowest part of game and film work, blocking out levels and dressing scenes, into a prompt. The skill that matters shifts from knowing the menus to directing the agent that drives them.

Anyone building in 3D or real-time can pull the 5.8 MCP plugin and let Claude Code dress one test level this week, then check what it gets wrong before trusting it on a real scene. The pattern runs past games: any pro tool that exposes an MCP server is now something an agent can run. Watch which of the tools in your own stack already do.

A solo team open-sourced Palmier Pro, a Mac video editor built from scratch in Swift that exposes an MCP server so Claude, Codex, or Cursor can operate the timeline: read the project, trim and reorder clips, and generate new footage in place. The base editor is free and GPLv3; only the generative models behind the timeline sit behind a login. It is shaped like Premiere, but its primary user is meant to be an agent.

This is the Unreal story without a billion-dollar engine behind it. One small team built a real editor whose main operator is a model, not a human dragging clips, and gave it away. When the interface a tool exposes is an MCP server, "who can use this software" stops meaning people and starts meaning any agent you point at it. The editor became an API with a timeline attached.

Make video and want to test this: install it and ask Claude to cut a rough first pass from your raw clips this week, then judge whether agent-led editing saves the boring 80% or just moves the work around. The read for builders is sharper than the tool itself: shipping an MCP server may now matter more than shipping a polished UI, because the agent is the one holding the mouse.

On June 24 Serbian startup SuperPlane raised €2.28 million in pre-seed led by Credo Ventures. The founders are Darko Fabijan and Marko Anastasov, the pair who built Semaphore, the CI/CD platform Confluent, Replit, and Superhuman ran on. Their new product is an open-source control plane that lets AI agents run production work, deployments, incident response, infrastructure operations, while engineers keep the final say. It already wires into more than 30 tools, including AWS, GCP, GitHub, Slack, PagerDuty, and Datadog.

It is the same pattern as the Slack and editor stories, one layer deeper. Those agents handle your messages and your scenes; SuperPlane hands an agent the systems that page you at 3am, then puts policy and guardrails between the agent and anything it could break. That is the harder and more valuable place for AI to live, and it is being built by Balkan founders with a real exit behind them, not a US lab. The bet is that the missing piece for production agents is not a smarter model, it is a control layer that keeps a human in the loop.

If you run infrastructure, the takeaway is the architecture, not the round: agents are moving from your editor into your deploy pipeline, and the teams that win wrap them in deterministic guardrails before handing over access. It is open-source, so you can read how they draw the human-in-the-loop line this week without a sales call. For regional founders, watch who wrote the cheque: a Czech fund backing Serbian infrastructure built by proven operators.

Four tools to install or test this week. Marketing and sales were quiet, so productivity gets two.

Productivity: Gemini in Sheets fixes a broken formula in one click. Google added a Fix button to formula errors in Sheets: hover the error, and Gemini reads the surrounding data, explains what broke, and drops in a corrected formula. It is live for Business Standard, Enterprise, and AI Pro and Ultra, with higher promo limits through July 15. Break a VLOOKUP on purpose and let it debug before you trust it on a real model.

Productivity: Perplexity moved Deep Research inside Computer. Perplexity now runs a research question and turns the result into a report, spreadsheet, deck, or dashboard in the same place, routing subtasks across 20+ models, with inline editing on any asset it builds. It is rolling across iOS, Android, Mac, and Windows this month. Point it at one recurring research-to-deck task you would normally hand off.

Design: Figma Motion puts a real animation timeline on the canvas. At Config 2026, Figma shipped Motion, with keyframes, presets, and MP4, GIF, and SVG export inside the same file as your components, in open beta on every plan from June 24. You can also ask the Figma agent to build the animation for you. Animate one prototype this week instead of bouncing out to After Effects.

Dev: Mistral OCR 4 reads documents without leaving your servers. Mistral released OCR 4, a document model covering 170 languages with paragraph-level bounding boxes, shippable as a single container so sensitive files never touch a third-party cloud. For fintech, legal, or health teams that is citation-ready extraction you can self-host. Run it against one messy PDF pipeline this week.

See you on the next edition,
Çelik