The version number checks out. The JetBrains Marketplace listing shows Version 1.18.0-261 — 18 Sep 2026, and its release notes match GitHub's list, including assisted approvals and message re-editing. The Marketplace date is four days before the changelog post. Some users may already have the plugin build that GitHub announced on the 22nd.
Copilot for JetBrains 1.18.0 hands low-risk approvals to the AI
Assisted approvals is the headline change. It is a preview, and it is the one to think about before turning it on. GitHub says the feature is "automatically approving low-risk tool calls while continuing to prompt for higher-risk actions." The goal is fewer interruptions for routine actions, with higher-risk decisions still left to the developer. The feature applies to Copilot agent sessions only.
Earlier versions gave you fairly blunt choices. JetBrains' documentation for its Copilot integration describes an Allow All switch: when it's on, GitHub Copilot proceeds without asking for approval each time it needs to use a tool, access a file path, or open a URL. The same page warns that it allows the agent to act without your approval, which may lead to data loss or security issues. Before that, GitHub's April 2026 update added Global Auto Approve. When enabled, it automatically approves all tool calls across all workspaces and overrides per category auto approve settings, including for potentially destructive actions such as file edits, terminal commands, and external tool calls.
In between those extremes were rules you wrote yourself. A March release let users configure auto-approve for MCP at both the server and tool level, and April added defaults for terminal commands and file edits not covered by existing rules. Assisted approvals takes a different route. Copilot decides which calls are low-risk, so you no longer have to list every safe command in advance.
The trade-off is that the risk decision now belongs to Copilot's own classifier. GitHub hasn't said how that classifier defines "low-risk," which tools qualify, or how assisted approvals interacts with existing rule-based auto-approve settings. Treat the preview label as a real warning. Watch what it approves on a sandbox project before trusting it with a production repository.
Enterprise administrators may already have some control. GitHub's August update on enterprise managed settings says administrators can set permissions.disableBypassPermissionsMode to disable to prevent the Copilot agent in JetBrains from using Bypass Approvals or Autopilot. GitHub hasn't said whether that policy also governs assisted approvals. Admins who rely on it should test before assuming it does.
Message re-editing in Copilot agent sessions rewinds files, not just chat
The second big change deals with a familiar agent problem: a bad early instruction that the rest of the session builds on. You can now re-edit a previous message in a Copilot agent session. Before your replacement message is sent, Copilot rewinds both the conversation and the file changes made after that point. The Marketplace notes describe the same thing: rewinding the conversation and file changes before sending the replacement message.
Before this, the only fix was to add another message such as "no, undo that and do X instead." The agent then had to reverse its own work while carrying the discarded plan in its context. Rewinding cuts that branch off completely. The conversation restarts from your corrected prompt, and the working tree goes back to how it was at that point.
Keep in mind that the rewind is destructive. GitHub hasn't said whether a discarded branch can be recovered, or how the rewind treats manual edits you made between agent turns. If a session has produced anything worth keeping, commit it or stash it in Git before re-editing an earlier message. That way your version control, rather than the plugin, holds the fallback copy.
Codex plan mode puts a checkpoint before implementation
The Codex agent inside Copilot for JetBrains now supports plan mode. You can review, refine or approve a plan before implementation starts. That gives you a chance to shape the approach before the agent changes anything, instead of judging the result afterwards.
This brings Codex in line with the rest of Copilot for JetBrains. Copilot's own planning agent has been around for a while: in March GitHub said custom agents, sub-agents, and plan agent are now generally available in GitHub Copilot for JetBrains IDEs. Developers who use Codex as their agent now get the same review step before code is written.
Plan mode pairs well with message re-editing. Approving a plan is a cheap checkpoint before any changes happen. Re-editing is the more expensive fix for when things have already gone wrong. Neither replaces reviewing the finished diff, but together they give you more places to step in along the way. The changelog doesn't say whether Codex plan mode is generally available or in preview, or what format the plans use.
Organization skills and managed instructions reach local and agent sessions
For teams, the governance change is that both local sessions and Copilot agent sessions now support organization and enterprise skills, plus organization-managed custom instructions. Shared skills and team guidance now work in both kinds of session, not just one.
In practice, a platform team's coding conventions, repository rules or reusable agent skills can follow a developer from one session type to the other. This builds on GitHub's steady expansion of Copilot customization in JetBrains. The March release, for example, added support for AGENTS.md and CLAUDE.md instruction files, including the ability to generate an initial AGENTS.md file.
The release note is thin on administration. It includes no setup steps, doesn't say which Copilot plans or admin roles are needed, and doesn't explain how organization instructions rank against a developer's own instructions or a repository's AGENTS.md file. Admins should roll this out on a pilot group and check what the agent actually receives before relying on it to enforce standards.
GitHub MCP Server toggle and per-tool MCP controls tighten tool management
The Model Context Protocol (MCP) is how Copilot's agents connect to outside tools and data sources. Version 1.18.0 changes two things here.
- A new setting turns the built-in GitHub MCP Server on or off without touching MCP servers you configured by hand. The built-in server stays enabled by default.
- Copilot agent sessions get persistent per-tool controls for MCP servers. You can manage individual tools and also decide whether the built-in server is enabled.
Before this, a developer who wanted a smaller set of tools, whether to shrink the agent's reach or to keep an unneeded server out of the way, had coarser options. Now the GitHub server can be turned off on its own, and individual tools can be switched off in a way that sticks between sessions. The interface also now labels the built-in GitHub MCP Server in the tool configuration screen and links straight to its settings, so it's easier to tell apart from servers you added yourself.
These controls sit below the enterprise policy layer GitHub added in August. There, administrators can use allowedMcpServers and deniedMcpServers to centrally control which MCP servers developers can connect to from GitHub Copilot for JetBrains. In that setup, the organization decides which servers are allowed, and the new 1.18.0 switches let each developer narrow things further within those limits. GitHub hasn't documented how the two layers interact in edge cases.
Inline chat disappears from JetBrains Gateway as the 2025.1 notice arrives
Two changes matter most for fleet planning. First, inline chat and its entry points are now hidden in JetBrains Gateway and remote development environments. The change covers inline chat only. GitHub doesn't say other Copilot features are removed from remote setups. Still, teams that work mainly through Gateway will lose an entry point they may be used to. In April GitHub turned inline chat into a launch point for agent work, noting that inline agent mode is now available in public preview, reached through inline chat with Shift+Ctrl+I (Windows) or Shift+Cmd+I (Mac). Remote developers who used that path will need to use the chat panel instead.
Second, anyone on a JetBrains IDE version 2025.1 will now see advance notice to upgrade to 2026.1 or later. GitHub says support is unchanged in this release, so this is a warning, not a cutoff. GitHub hasn't given an end-of-support date. Still, it's a clear sign that IT teams keeping IDE versions frozen should plan a move to 2026.1.
The quality fixes are mostly routine but worth knowing:
- Inline chat now keeps your edits when a request ends, and it respects the thinking-effort and context-window settings you chose.
- Codex sessions that failed to start have been fixed.
- Copilot behaves better when several project windows are open.
- Embedded editors and message re-editing work again on IntelliJ 2026.3 EAP builds, which matters for anyone testing the next IDE release.
The rest of the interface work is about making features easier to find. There are browsable usage tips above the chat input with shortcuts to commands, customizations and settings. The welcome screen is simpler and has a direct feedback link. Shortcuts for updating agent instructions and for viewing usage-based billing best practices are back. The /init tip has been reworded and grouped with customizations. There's also a new side-by-side chat panel switcher in the session toolbar, so you can keep a conversation open in the editor while browsing the session list in the tool window.
What this means for you
Most JetBrains users can update now. Enabling assisted approvals is the one choice that deserves a deliberate trial first, particularly on machines that can reach production credentials or infrastructure. Admins should look at three things: the new organization skills and instructions, the Gateway change, and any developers still on IDE 2025.1.
- Update to GitHub Copilot for JetBrains 1.18.0 from the JetBrains Marketplace. The build is dated September 18, 2026, and brings the reliability fixes along with the new controls.
- Try assisted approvals on a low-stakes project first. It's a public preview, and GitHub hasn't published how Copilot decides what counts as low-risk.
- Commit or stash your work before re-editing an earlier agent message, because Copilot rolls back file changes along with the conversation.
- Use Codex plan mode to approve an approach before the agent writes code, and turn off unused MCP tools or the built-in GitHub MCP Server if you want a smaller set of tools for the agent.
- Tell Gateway and remote-development users that inline chat is now hidden there, and point them to the chat panel.
- Start moving developers on JetBrains IDE 2025.1 to 2026.1 or later. Support continues for now, but the upgrade notice has started.
Version 1.18.0 continues a pattern visible across GitHub's JetBrains releases this year: agents do more of the work, and the approval model keeps getting finer controls. Early 2026 brought hand-written rules, April brought a global override, August brought enterprise lockdown policies, and assisted approvals lets Copilot make the routine risk calls itself. Whether that preview moves to general availability, and whether GitHub explains how the classifier works and how admin policies govern it, will decide how widely organizations let it run. Until then, plan mode, rewinds and per-tool MCP switches are the parts of this release teams can use with confidence today.