CONFIRMED • OFFICIAL

Copilot CLI 1.0.103-0 improves MCP and sandbox
The preview release of Copilot CLI 1.0.103-0 improves MCP in worktrees and makes ACP sessions resume with saved sandbox permissions.
Marcelo Santiago
Criador do bridgecode
Get AI articles in your inbox
Technical curation on models, agents and coding tools. No spam, unsubscribe anytime.
On October 22, 2026, at 00:40 Brasília time, GitHub published Copilot CLI 1.0.103-0 as a preview release. The update changes the use of MCP tools, ACP session resume, and sandbox permissions.
The release does not announce a new model. It updates CLI behavior and is marked as preview on the official page.
In this article you will see:
- copying the original code in completed responses
- using MCP tools across worktrees
- sandbox rules preserved when resuming ACP sessions
Credit: GitHub.
Copilot CLI copies the original code from responses
The release adds an action to copy the original code from a completed response block. The action works when this behavior is enabled in the CLI settings.
MCP reuses snapshots between Copilot CLI worktrees
MCP, short for Model Context Protocol, is a standard for connecting AI applications to tools. Version 1.0.103-0 allows reusing local snapshots of MCP tools between Git worktrees registered in the same repository.
A Git worktree is an additional working folder linked to the same repository. This adjustment brings the CLI closer to agentic flows with more than one workspace, a topic also explored in the analysis on agentic development environments.
ACP sessions preserve sandbox settings
ACP is the Agent Client Protocol, a communication protocol between clients and programming agents. The release states that sessions now respect saved sandbox settings and the options --sandbox and --no-sandbox, including when they are resumed.
The notes describe the resumption of saved rules, but do not report changes to system isolation limits.
Copilot CLI fixes connections, Gemini calls and permissions
The notes also fix cloud responses that could lose text after a disconnection. The session browser now shows the registered title when it exists, instead of a generic name.
Gemini tool calls no longer fail with a 400 validation error on nullable fields or enum. The version also adjusts sandbox permissions for .NET commands, NuGet and the npm.cmd shim on Windows.
The CLI applies Responses verbosity defaults for GPT-5 models when the session does not define another value. It also updates mention suggestions when the GitHub client is ready.
How to test Copilot CLI 1.0.103-0
The official page offers installation files for Windows, macOS and Linux. As this version is preview, test it in a separate environment before using it in production tasks.
To compare another CLI flow and model distribution, see how OpenAI brought Codex flows to Google Cloud.
Frequently asked questions about the Copilot CLI 1.0.103-0
Is Copilot CLI 1.0.103-0 stable yet?
No. GitHub marks the release as preview, so it is not a stable release yet.
What changed in Copilot CLI 1.0.103-0?
The release adds copying of original code in completed response blocks, improves MCP snapshot reuse, and fixes ACP session resume with sandbox.
How does Copilot CLI 1.0.103-0 use MCP across worktrees?
It allows reusing local MCP tool snapshots between Git worktrees registered in the same repository.
What happens to the sandbox when resuming an ACP session?
The session now respects saved settings and the --sandbox and --no-sandbox options during resume as well.
Does Copilot CLI 1.0.103-0 include a new model?
No. The release covers changes to the CLI and its integration with tools and sessions. The notes also fix the Responses verbosity default to GPT-5 when the session does not set another value.
Copilot CLI 1.0.103-0 focuses on continuity and control
The gain of this preview release is in daily work with tools and sessions: the CLI reuses MCP data between worktrees and recovers sandbox permissions when resuming a session. The publication does not announce a new model version.
Stop switching between AI tools. Coordinate your agents in a single flow on bridgecode.
Source: GitHub
Related articles

Karpathy builds a video platform over a weekend with Claude
Andrej Karpathy built HackerTube in a weekend with Claude. What it says about development speed with coding agents.

Rich Sutton reinforces that compute alone does not create intelligence
Turing Award winner Rich Sutton reinforces that raw compute is not enough for intelligence and calls for new methods, continuous learning and world models.

AI auditors detect misbehavior but generate false alarms
AI models catch a good portion of other agents' misbehavior, but they generate a lot of false alarms. See the numbers from Andy Hall's study.
