COPILOT CLI 1.0.21 SHIPS MCP SUPPORT; SAFER AGENT LIMITS LAND IN 1.0.22-0 PRE-RELEASE, WHILE COPILOT UPDATES DATA-TRAINING POLICY FOR INDIVIDUALS
GitHub Copilot CLI now manages MCP servers, adds agent safety limits in pre-release, and GitHub updated Copilot’s data training policy for individual accounts. ...
GitHub Copilot CLI now manages MCP servers, adds agent safety limits in pre-release, and GitHub updated Copilot’s data training policy for individual accounts.
GitHub released Copilot CLI 1.0.21 with a new copilot mcp command for managing Model Context Protocol servers, plus stability and memory fixes releases. A pre-release 1.0.22-0 adds sub-agent depth and concurrency limits to curb runaway agent spawning, warns on session conflicts, and fixes a V8 crash.
An experimental “Rubber Duck” reviewer reportedly uses a second model to critique plans and edits, accessible via /experimental when using Claude, with GPT-5.4 behind the scenes Neowin. Separately, GitHub’s privacy update says Copilot interaction data for Free/Pro/Pro+ may train models starting April 24, with opt‑out available; Business/Enterprise are excluded GitHub Community FAQ.
MCP support lets teams expose vetted internal tools to agents in a controlled way.
Agent depth/concurrency limits reduce the risk of runaway processes and surprise token bills.
-
terminal
Connect an internal MCP server via
copilot mcp, then run a real workflow and measure success rate, latency, and token spend. -
terminal
In a sandbox, try 1.0.22-0 with high-complexity tasks; tune sub-agent depth and concurrency, and compare token usage and stability with/without experimental Rubber Duck.
Legacy codebase integration strategies...
- 01.
Wrap existing CLIs and data services behind MCP to grant least-privilege tool access to Copilot without broad credential exposure.
- 02.
Ensure developers use Enterprise seats for company code to avoid individual-account data training; document opt-out for any personal accounts used.
Fresh architecture paradigms...
- 01.
Design new automation around MCP-first tools with clear, idempotent commands and machine-friendly outputs.
- 02.
Add quality gates and acceptance criteria to agent runs; agents still miss non-functional requirements analysis.