Skip to content

Changelog ​

0.7.2 (2026-11-20) ​

Behavior changes ​

  • cursorAccount channel configurations now accept up to 50 repositories or projects (up from 20) across GitHub, Origin, GitLab, and Bitbucket, and reject GitHub and Origin repository names longer than 140 characters.
  • channels is now a reserved agent slug (alongside v1, playground, and docs) so multi-agent slugs do not collide with hosted channel routes.

New ​

  • The Per-user MCP connections guide explains how to run hosted cursorAccount MCP connectors as the signed-in user. It covers credentials: "individual" declarations, user sign-in, and turns without a signed-in user.

Fixes ​

  • agent-sdk channels teams doctor now accepts the public hosted Teams endpoint URL (https://api.cursor.com/v1/agent-sdk/<applicationId>/channels/<channelId>, and /<slug>/channels/<channelId> for multi-agent deployments) alongside the legacy /engine path.
  • When a direct tool call on an individual cursorAccount MCP connection fails because the user has not connected that server, the runtime shows the Connect prompt in the playground and Slack instead of returning only an authentication error.
  • In the playground, a tooltip opened by hover or focus now closes when the part of the screen it is on is hidden.

0.7.1 (2026-11-20) ​

Fixes ​

  • On Cursor hosting, scheduled runs record trace events under the platform session id so they appear in the Trace tab. Before, scheduled runs wrote traces under a local id and the Trace tab stayed empty.
  • GitHub and Origin cursorAccount connections ignore GitLab and Bitbucket repository entries in the shared event list instead of failing startup.
  • cursorAccount MCP server name matching trims surrounding whitespace on server names and identifiers and rejects blank names.

0.7.0 (2026-11-19) ​

Minimum versions ​

  • @grok/agent-sdk requires @cursor/sdk 1.4.2, so each model call's usage arrives as a model_call.completed event.

Behavior changes ​

  • The [cloud agent stopped] wake message now holds only the agent id and its status, and tells the model to call cloud__get to read the agent's result. It no longer holds the agent's details.
  • On Cursor hosting, a cloud agent that a cursor-cloud-agents conversation launched wakes that conversation once when its run stops. This includes a run that someone else cancelled. Nothing polls the agent. If no wake comes, ask the bot to check the agent. A stop in the middle of a turn does not wake the conversation yet.
  • Launches off Cursor hosting, from a mount that sets its own apiKey, or with envVars are not watched, and return watching: false. A launch that Cursor hosting cannot watch also returns watching: false. Before, the stop watch polled the agent every minute.
  • A wake that reached no session, such as a failed session restore or a busy session, is sent again every minute, up to 10 times.
  • With a skills callback, the skills index lists only the discovered skills available in the mounted tree, rechecked each turn. Skills whose availability check fails are left out.

New ​

  • teamsChannel() from @grok/agent-sdk/channels/teams connects an Azure Bot that you own to Microsoft Teams personal chats, group chats, and channel posts. It runs under agent-sdk dev and agent-sdk serve. The agent-sdk channels teams commands store the bot's credentials, build the Teams app package, and check the setup.
  • teamsChannel() reacts with 👀 (1f440_eyes) to each accepted inbound message when its turn starts. Set reply.ack to another Teams reaction id to change the reaction, or to false to turn it off for the channel. Return quiet: true from an onAppMention, onDirectMessage, or onChannelPost handler to start that message's turn without the reaction.
  • ctx.steer(message, { deliveryId }) on channel event and hook contexts (SessionContext) and on tool contexts (ToolContext) adds a message to the agent's own running turn through Run.steer. It never starts a turn, opens a session, or queues the message: it resolves to steered, no_turn, or not_delivered.
  • getCloudAgent from @grok/agent-sdk/extensions/cursor-cloud-agents and the cloud__get tool return result, the final answer of the agent's latest run, once the agent has stopped and that run finished. An errored or expired run has no result. Use it to read a cloud agent's report or JSON from your own code. For a stopped agent, they also read the latest run from the Cloud Agents API, and fail if that read fails.
  • ReminderHostApi.fireNow(id) fires an armed reminder as soon as the platform can, through its normal claim.
  • Extension.namespace is the namespace an agent mounted the extension under, for a tree that names its own <ns>__<name> contributions.
  • ReminderHostApi.createRequest(input) returns the request that create would send, validated the same way, without creating the reminder.
  • launchCloudAgent takes an optional agentId. Pass the same id again to retry a launch that timed out or failed: when the first launch created the agent, you get that agent and no second agent starts. A launch with cloud.envVars cannot take an agentId.

Fixes ​

  • On Cursor hosting, a retried delivery continues on the cloud agent that its first attempt created, even when that attempt stopped before it recorded the agent. Before, the retry launched a second cloud agent that did the same work. A retry of a cancelled cloud run stays cancelled. An agent that sets cloud.envVars can still start a second cloud agent on a retry.
  • On Cursor hosting, host.reminders.dispatch() runs the reminder's handler once. Before, it failed with HTTP 405, and the handler ran again at the reminder's fire time.
  • The CLI writes stored credentials, MCP OAuth tokens, and .env.local so that only your user can read them. It narrows files that earlier versions wrote the next time it reads them.
  • In Slack, images in a reply that arrives while a turn runs reach the follow-up turn. Before, only the reply's text did.
  • In Slack, a file that the agent can't attach is named in the prompt with the reason, instead of dropped without a word. Files shared from another workspace now download, and a video sends its poster frame as an image.
  • In Slack, channel, user, and user group mentions that arrive without a label reach the model as names. Before, an unlabeled channel mention read as #channel. Private channels and DMs show their id.
  • On an HTTP 403, agent-sdk deploy prints the server's reason instead of guessing a cause.

0.6.0 (2026-11-03) ​

Minimum versions ​

  • @grok/agent-sdk requires @cursor/sdk 1.3.0.

Behavior changes ​

  • On Cursor-managed hosting, when the team turns on its Cloud Agent Allowlist, a turn checks the url: and command: connections it opens against that list. It also checks their network hosts against the team's network allowlist for Cloud Agents. Before, it used the team's MCP Allowlist. Agents that you run with agent-sdk serve keep the MCP Allowlist.
  • On Cursor-managed hosting, when the team turns on its Cloud Agent Allowlist, the Cursor backend checks each cursorAccount connection against that list. The turn's local MCP checks no longer refuse these connections.
  • An agent with cursorAccount connections starts its shared workspace after those connections resolve. Other agents still start it when the server starts.

0.5.0 (2026-10-28) ​

Breaking changes ​

  • No change to which file tools an fs agent has or what they do. Their results now use Cursor's native tool-call shape instead of Agent SDK's earlier MCP-style text, in tool_call events, transcripts, and direct tool calls. The same holds without fs. Update code that parses these results. For example, a direct read call returns { status: "success", value: { content } } instead of numbered lines.
  • With fs, a tools list gets only the file tools it names. Add "read", "ls", "glob", "grep", "edit", and "delete" to keep them.
  • @grok/agent-sdk/executors is removed. Remove that import. The npm package has no replacement.
  • synthesizeFileSystemTools is removed. Pass your tree as fs.
  • nearestNeighborMissHint is removed. Read suggests nearby names when a file is missing.
  • RunnerTurnRequest carries fs, reportedWorkspacePath, and settingSources instead of root and subagentInherit.

Minimum versions ​

  • @grok/agent-sdk requires @cursor/sdk 1.2.0.

Behavior changes ​

  • With fs, AGENTS.md and .cursor/rules/ in the workspace load as project rules, as they do on the host. Don't make a checkout you don't trust the workspace.
  • With fs, omitting tools gives every tool the tree can back, including subagents, web search, web fetch, todos, and MCP.
  • With fs, a tools list that names a tool the tree can't run, such as readLints, leaves it out with a warning. Before, readLints with fs failed validation.
  • With fs, LS, Grep, and Glob search the workspace when a call names no path.
  • With fs, each turn gets an empty /tmp. Large web results are written there.
  • With fs, subagents work on the agent's tree, and their writes land in it as the agent's own do. Before, each subagent wrote to its own copy.
  • local.sandbox works with fs. A tool call that needs approval, such as an external MCP call, fails closed. File access doesn't change.
  • With local.sandbox and no fs, an agent with shell in tools, or with tools omitted, can read outside the session workspace with the file tools, as its shell already could. Writes and deletes outside it are still refused.
  • With local.sandbox and no fs, an agent whose tools list file tools and no shell runs on its session workspace, mounted at /workspace, and its reads stay inside it. Cursor hooks don't run for it, so a Slack follow-up waits for the turn to end. .cursor/mcp.json from its working directory isn't loaded, host git status doesn't reach the model, and its skills are at /agent/skills.
  • An agent tool can be named Read, Grep, or another file tool's name.

New ​

  • workspace sets the folder inside fs that the agent works in. For a sandboxed agent without shell, it sets where the session workspace is mounted.

Fixes ​

  • An agent that exports OpenTelemetry traces no longer logs a duplicate propagator registration error on its first HTTP request.

0.4.0 (2026-10-27) ​

Breaking changes ​

  • The FileSystem type, its helpers, and its errors match the file tools, and the package root exports readOnlyFs.

Fixes ​

  • A reply in a Slack thread that a schedule started continues that session.
  • A hosted Slack agent with botMentions.allowFrom starts a session when an allowed bot mentions it.

0.3.0 (2026-10-25) ​

  • GA Release
  • persistent option is removed from defineAgent

0.2.29 (2026-10-23) ​

  • Bug Fixes