Notion is making a claim on where your AI agents do their work, rather than only on where you keep your notes. On 13 May the company launched a developer platform that lets teams write custom code, pull in data from outside systems and connect agents directly into the workspace.

The three pieces

The platform has three main parts. Workers is a cloud sandbox where you deploy your own code; from 11 August, running it draws on Notion credits. Database Sync pulls records in from any system with an API, naming Salesforce, Zendesk and Postgres as examples. And External Agents lets you bring in agents you did not build, with Notion saying Claude, Codex and Decagon work out of the box.

Why it is a smart move

Notion already holds a lot of a company's working context — docs, tasks, wikis, project notes. That context is exactly what an AI agent needs to be useful. By opening the workspace to outside agents instead of betting only on its own, Notion is trying to become the place agents plug into rather than one more tool they have to learn. More than a million of its own custom agents have been created since that feature launched in February, so the demand is there.

The competition it implies

Every productivity suite is making the same pitch now, from Microsoft to Google to a dozen startups. The winners will be the platforms that hold the most useful context and make connecting an agent easy. Notion is brisk and developer-friendly, which suits it here. The risk is the one Notion always faces — that the giants will bundle a competing feature, forcing Notion to be clearly better at its chosen function.

The interop bet paid off, and then stopped being a differentiator

The judgement here was that opening the workspace to agents Notion did not build was the smart move, because the platform agents plug into beats the platform with its own agent. Three months on, the first half is confirmed and the second half has a complication.

Bringing in Claude, Codex and Decagon out of the box was ahead of where the market was in May. It is now roughly where the market is. Agent Plugins 1.0.0 shipped in August, packaging Agent Skills and MCP servers into a single portable directory that five competing vendors agreed on. Portability across agent hosts is becoming a specification rather than an integration each platform negotiates.

For anyone building agents that is straightforwardly good news. For anyone selling the socket it removes a reason to choose one socket over another. External Agents was a reason to choose Notion in May. When the folder layout is standard, supporting outside agents is table stakes, and the question moves back to the thing the piece already identified as the real asset: whose context is worth plugging into.

MCP turned out to be small enough to be commoditised

One assumption behind a developer platform is that connecting an agent is hard enough to be worth outsourcing. That assumption has weakened considerably.

The protocol underneath most of these integrations is JSON-RPC over a pipe, and a working server can be written in about sixty lines without a framework or any dependencies. Giving an agent search over a private document set takes roughly seventy-eight lines, joining a local embedding model, a vector store and MCP.

This does not make Notion's platform redundant. Features like Workers, Database Sync and version history are valuable operational conveniences, and a reliable sandbox is worth paying for. It does, however, change what is being sold. The pitch is no longer that connecting an agent is difficult; it is that Notion will run and maintain the connection, and that the workspace on the other end of it already holds the context.

The credits detail became the more interesting part

Workers drawing on Notion credits from 11 August was a single line in the original announcement. It is the part that aged into significance.

Metered consumption has become the default shape of agent billing across the sector, and the layer that meters it is being consolidated: Stripe agreed in August to buy OpenRouter, a gateway that routes work across roughly 400 models and meters the token spend. A workspace that hosts other people's agents in its own sandbox is taking on the cost of running work it does not control the efficiency of, which is a different exposure from selling seats.

The bundling risk is unchanged and better documented

The risk named here was the perennial one: the giants bundle, and Notion has to be clearly better at the thing it picks. Nothing since has softened that, and the market data makes the position more specific.

An audit of 1,347 researched AI product entries found that the five categories everyone reasons with — chat assistants, copilots, command-line agents, autonomous runtimes, model providers — describe only 46 per cent of the market, and that the largest shape in it is the single-purpose application. A workspace positioned as a universal host for agents competes not only with bundled suites but also with a long tail of single-purpose tools that need no host at all.

The million custom agents created since February remains the strongest number in the piece, because it measures the context advantage rather than the integration surface. That is the asset the specification work cannot commoditise.