Mistral Vibe: MCP Sampling, Monorepo Support, and Smarter Agent Behavior
Mistral VibeMistral Vibe 2.3.0 introduces MCP sampling support, enabling MCP servers to request LLM completions directly through the agent rather than requiring their own model connections. The release also adds monorepo-aware agent discovery, allowing Vibe to find agents, tools, and skills defined anywhere within a trusted directory tree — not just at the root. A fully rewritten system prompt built around an Orient/Plan/Execute three-phase workflow rounds out the release, improving how the agent approaches complex tasks.
Sources & Mentions
3 external resources covering this update
MCP Sampling, Monorepo Support, and Smarter Agent Behavior
Mistral Vibe 2.3.0 delivers several foundational improvements to how the agent integrates with MCP servers, discovers project configuration, and plans its work.
MCP Sampling Support
MCP servers can now request LLM completions directly through Vibe rather than needing their own model connections. When an MCP server issues a sampling request, Vibe routes it through its own model, returns the completion, and caches the result in the MCPRegistry. This dramatically simplifies MCP server development — servers no longer need API keys or model configuration of their own.
Monorepo Agent Discovery
Vibe now recursively searches subdirectories within a trusted directory tree to find agent definitions, tool configurations, and skill manifests. Previously, these had to live at the repository root. This is a significant quality-of-life improvement for monorepo setups where individual packages may define their own agents. The /resume command also gains interactive session selection, making it easier to pick up previous work.
Rewritten System Prompt
The agent now operates under a three-phase workflow: Orient (understand the codebase and context), Plan (outline the approach before acting), and Execute (carry out the plan). This structured approach reduces wasted tool calls and improves the agent's ability to tackle multi-step tasks. Brevity rules also encourage more concise responses.
Improved Permission Handling
Permission management gains allowlist and denylist globs, giving users fine-grained control over which files and directories the agent can access. Out-of-cwd detection warns when the agent attempts to operate on files outside the current working directory.
Quality-of-Life Improvements
The release includes terminal bell and title notifications (so you know when a long-running task finishes), subagent traces stored in an agents/ subfolder for debugging, and auto-scroll suspend to prevent the terminal from jumping during output.