Cursor Calls Its Agent SDK Generally Available, Under a New Package Name and With No Price Listed

Cursor's October 20 Agent SDK post shows a package called @cursor/agent-sdk, an agent-sdk deploy command and a defineAgent call. The SDK Cursor shipped on April 29, 2026 was @cursor/sdk, and none of the Cursor pages read for this story explain how the two relate.

The October post, written by Rustam Lalkaka on cursor.com, calls the Agent SDK generally available. It does not state a price, a rate limit or a licence.

What the Agent SDK post describes

The post says the SDK lets developers "define agents in code and manage them like any other software project." Agent instructions are Markdown files and tools are TypeScript functions. Each agent runs as a single Node process.

Deployment has two routes: agent-sdk deploy to Cursor's infrastructure, or a container on the developer's own. The one code sample sets a name, a model ID with parameters such as effort and fast mode, and a local working directory. The post also mentions evals, event replay, subagent handoffs and pull request events from GitHub, GitLab and Bitbucket.

Cursor says it builds its own managed agents on the SDK, including Bugbot, Security Reviewer and Rollouts. An internal agent called kb-reviewer was built in about a day, according to the post.

The one outside voice is Michael Gyarmathy, Principal Engineer at PGA America: "The Agent SDK has enabled PGA to build every component of our software factory on Cursor's infrastructure." The post gives no figures on what that factory runs.

What shipped in April, and what was said then

The April 29 changelog described the Cursor SDK as agents built on "the same runtime, harness, and models" behind Cursor's desktop app, CLI and web app. It ran locally or on a dedicated cloud VM, and it was billed on token-based consumption. Its status was public beta, open to all users. The changelog example used the model composer-2.5.

Detail	April 29 Cursor SDK	October 20 Agent SDK
Package	@cursor/sdk	@cursor/agent-sdk
Status given	Public beta	Generally available
Pricing given	Token-based consumption	None given
Deploy route	Local or Cursor cloud VM	agent-sdk deploy or own container

The columns come from two different pages written six months apart, so the table shows what each page says, not a verified difference in the products.

An outside reader's caveats

SpotDev, a UK consultancy that sells Cursor and AI services, published a guide on October 19 and updated it October 21. Its author, John Kelleher, wrote that SDK runs bill under the same pricing, request pools and Privacy Mode rules as the editor and cloud agents. Service accounts are an Enterprise feature.

The guide lists limits. Local agents run with the sandbox off by default and approve tool calls automatically. A cloud agent allows one active run at a time, and a second run gets a busy error. It also says products sold to customers need a commercial position agreed with Cursor, because the public terms do not cover embedding. Its Node.js minimum is 22.13.

That guide never mentions agent-sdk. It says the current docs carry no beta label, so readers should check status themselves. SpotDev has a commercial interest in Cursor work, and its claims are single-source.

What to check before building on it

Anyone who started on @cursor/sdk this summer should read the migration notes, if Cursor has published any, before assuming the new package is a rename. The October post does not link one. For a team picking between running agents on Cursor's infrastructure and in its own container, the missing price is the open item: the April billing model was tokens, and the October post is silent.

Agent toolkits carry their own data questions: Hallmark and Impeccable drew 128,000 stars while only one README documented its telemetry. Hosted agent backends have also been a target: OpenCode and Warp both lost server-side data within 10 days.

Cursor's documentation at cursor.com/docs is where the pricing and package details should appear next.