The AI coworker is a weekend project now
Viktor and Martoshi make the AI-employee-in-Slack look like a company. We built the same shape on Vercel's Eve in one session.
By Ilko Kacharov, CTO & Co-founder, Juma Labs
- Published
- Words
- 779
- Reading
- 4 min
- Sections
- 04
Viktor calls itself an AI employee. Martoshi is an AI coworker you tag in a channel. has its own Slack agent now too. Every one of them reads like a company you would need a year and a Series A to build. We built the same shape of thing this week, in one working session, as a directory of files. The category is thinner than the pricing pages make it look.
The category looks like a moat, and that is fair
Look at what they sell. gives each user a computer in the cloud that writes and runs code, wired to 3,200 tools, in the footer. tags into a channel and drops back a PDF, a dashboard, an ad audit, over a thousand integrations behind it. It reads like platform. Isolated compute, persistent memory, an integrations team, a compliance page. From the outside the hard part looks like everything.
It is a fair read. That is genuinely what an AI coworker needs. The mistake is assuming you have to build all of it.
Most of that moat is now framework primitives
Eve is the tell. shipped it in mid-2026, a filesystem-first framework for durable agents, and still in beta. You do not wire an agent. You lay out a directory. agent/instructions.md is the system prompt. agent/channels/slack.ts is the Slack entry point. agent/tools/*.ts is one typed tool per file. agent/connections/*.ts is your integrations. discovers the files at build time and compiles them into an app that runs on . No registration code, no glue.
Sessions stream token by token, and they resume after a or a deploy and pick up where they left off. That "his own computer in the cloud" bullet everyone markets? It is just the platform. It is the default.
This is the part that matched our stack so cleanly it felt like cheating. We already ship TypeScript on Vercel and talk all day. Eve is native to all three. There was nothing to bridge.
The integration is one file
The scary number on every one of these pages is the integration count. Here is ours.
agent/connections/juma.ts is a defineMcpClientConnection pointed at our own MCP server, authed per user through Vercel Connect so every teammate links their own account. Connect is the quiet hero here. The dance, the per-user token exchange, the refresh, the callback route you always dread writing, it owns all of it and hands you a token when you ask for one. I never stood up an auth server. That one file gives the agent the whole surface of : brand profiles, knowledge, projects, and the tools that generate on-brand decks, reports, and images. That is Martoshi's entire deliverable list, sourced from the product we already run.
This session we added a second file. agent/connections/linkedin.ts, a defineOpenAPIConnection with an inline spec and the w_member_social scope, so the agent can draft a post in Juma and publish it to with a human approval gate in front of the publish call. Wiring it was a file, not a project, because with Connect anything that speaks OAuth is one connector away. "3,200 integrations" is a number you reach one file at a time, and half of them are MCP servers already waiting to be pointed at.
The integration flipped
An integration used to be a project. On Eve it is a file in a directory. The moat was the glue, and the glue is gone.
What is actually hard is the part you already own
Strip away the Slack bot and the connectors and you are left with the only thing that was ever the product: what the agent knows. Martoshi's deliverables land because of the brand and campaign context behind them, not the channel they post in. Ours land because Juma holds our brand, our knowledge, our source of truth. The Slack agent is a thin shell around that context, and context is the new CMS.
The shell took an afternoon. The context took years, and it is the only part a competitor cannot clone by reading this post.
So the math flipped again, the same way it did when we deleted Jira. When the build cost of the surface collapses, the value moves entirely to what sits behind it. As of mid-2026, the surface is a weekend.
You do not hire the AI coworker anymore. If you own the context, you compile one, and you keep the keys.
Glossary
Terms
- Apache-2.0
- A permissive open-source license. Anyone may use, change and redistribute the software, including commercially, as long as they keep the notices and the license text.
- cold start
- The delay when a program or serverless function starts fresh, with nothing already running or loaded. For an agent it also means starting with no memory of earlier work.
- SOC 2
- An audit report in which an independent firm checks that a company handles customer data securely. Buyers often ask for it before they will trust a vendor.
Tools
- Claude
- Anthropic's family of large language models and the assistant built on them, available in apps and through an API.
- Eve
- Vercel's framework for building agents, where behavior is defined in Markdown and TypeScript files and deployed to channels such as Slack and the web.
- Juma
- A collaborative AI workspace for marketing teams. It holds brand profiles, knowledge and projects, and produces on-brand decks, reports and images from them.
- A professional social network where people and companies keep profiles, connect with each other and publish posts.
- Martoshi
- An AI marketing assistant that works inside Slack and Teams and returns finished outputs such as PDFs, dashboards and audits instead of plain text replies.
- MCP
- Model Context Protocol, an open standard for connecting AI applications to external data sources, tools and workflows.
- OAuth
- The industry-standard protocol for authorization. It lets an app get limited access to a user's account on another service without handling the user's password.
- Slack
- A messaging platform for work that brings team conversations, apps and customers together in one place.
- Vercel
- A cloud platform for deploying web apps, APIs and AI agents. It builds from a Git repo or the CLI and runs server-side code as functions.
- Vercel Functions
- Vercel's server-side code runtime for API routes, webhooks and agent request handlers. It scales automatically with traffic.
- Viktor
- An AI employee product that works inside Slack and Microsoft Teams, connects to external tools and carries out tasks such as reports, dashboards and code.
