Claude Code Now Runs Your MCP Tools in the Background
Claude Code 2.1.319 moves long MCP tool calls to the background so your agent keeps working while they run, plus a deny-instead-of-prompt option.

Claude Code background MCP tasks arrive in 2.1.319
Claude Code shipped three releases in one week, and the one that matters most is 2.1.319, released November 6. It moves an MCP tool call to the background when the server runs it as an MCP task, then reports the result when it finishes. Your agent keeps working while the long call runs instead of sitting idle waiting for it.
That sounds minor until you have watched an agent stall for two minutes on a slow tool. MCP, the Model Context Protocol, is how coding agents talk to outside tools: databases, search indexes, CI systems, design files. Some of those tools are fast. Some are not. A semantic search across a large codebase, a test suite run, or a deployment check can take minutes, and until now the agent's whole loop waited on it. Background tasks break that coupling. The agent fires the call, moves on to other work, and picks up the result when it lands.
The same release train added a deny-instead-of-prompt option, which lets you configure tools to be denied outright rather than prompting you every time. That is a small quality-of-life change with a real purpose: if you know a class of tool call should never run in a given project, saying no once beats saying no fifty times.
The week in developer tools
Claude Code's releases landed in a busy week for agent tooling. GitHub shipped Copilot inside Xcode's built-in coding assistant in public preview, starting with Xcode 27.3 beta on macOS Tahoe 26.6 or later, and a day earlier made Copilot an agent inside JetBrains Air via the Agent Client Protocol. OpenAI released Codex CLI 0.167.0 with an Agent Command Center. VS Code 1.145 added chat view modes, remote browser preview, and cron expressions for local automations.
The pattern across all of them is the same: the agent is no longer a chatbot in a sidebar. It is becoming the environment. Copilot inside Xcode means Apple developers get AI assistance without leaving Apple's IDE. Copilot as an ACP agent in JetBrains Air means the same agent can run local tasks in an IDE or cloud tasks through Air Teams in Docker. Claude Code's background tasks mean the agent can juggle slow tools the way a developer juggles terminal windows.
Google's November Chrome DevTools release fits the same pattern from the other side: an upgraded DevTools MCP server that gives agents programmatic access to live browser runtimes, including performance profiles, memory analysis, and network traces. The browser is becoming a tool the agent can drive, not just a window the developer stares at.
Claude Code MCP tasks in practice
The practical effect is about parallelism, not speed. A single MCP call does not run faster in the background. What changes is that the agent can do useful work while it waits: read another file, draft a test, summarize what it already knows. For long-running workflows like running a full test suite or querying a slow internal API, that is the difference between an agent that feels stuck and one that feels productive.
What background tasks change in practice
The practical effect is about parallelism, not speed. A single MCP call does not run faster in the background. What changes is that the agent can do useful work while it waits: read another file, draft a test, summarize what it already knows. For long-running workflows like running a full test suite or querying a slow internal API, that is the difference between an agent that feels stuck and one that feels productive.
There is a catch worth knowing. Background tasks only work when the MCP server supports them. If your server runs a tool call synchronously and returns only when it is done, Claude Code cannot background it. The feature is an agreement between the client and the server, not a trick the client can pull alone. If you maintain an MCP server, supporting the task protocol is now a feature your users will notice.
If you are setting up Claude Code for a team, pair the background tasks change with the deny-instead-of-prompt option and a clear policy about which tools are allowed to run unattended. Agents that can work in parallel are more useful, and they are also more capable of doing the wrong thing in parallel. The guardrails matter more as the agents get faster.
The broader direction is clear from this week alone. GitHub Copilot is now inside Xcode and JetBrains Air, VS Code keeps absorbing agent features, and Gemini is landing inside GitHub Copilot too. The IDE wars are over. The agent platform wars have just started, and the winners will be the tools that waste the least of your time waiting.


