#mcp

A LocalLLaMA discussion around OpenCode shows why developers are experimenting with open, model-agnostic coding agents even when closed systems still lead on raw frontier performance.

Google says Gemini CLI now includes a read-only Plan mode that analyzes requests, codebases, and dependencies before any edits happen. The update also adds an ask_user tool and read-only MCP access so teams can clarify requirements and pull in outside context without risking accidental changes.

A fresh Hacker News thread pushed Chrome DevTools MCP back into view because it lets coding agents attach to an already running Chrome session instead of starting from a clean browser. Google’s December 11, 2025 update combines Chrome 144 remote-debugging controls with an --autoConnect flow in the MCP server, making it easier to hand an in-progress DevTools investigation to an agent.

GitHub said on March 10, 2026 that GitHub Copilot, VS Code, and Figma now form a continuous loop through the bidirectional Figma MCP server. GitHub’s March 6 changelog says users can pull design context into code and send rendered UI back to Figma as editable frames.

A March 14, 2026 Hacker News discussion highlighted a more nuanced MCP argument: local stdio MCP can be unnecessary overhead for bespoke tools, while remote HTTP MCP still solves auth, telemetry, and shared tooling at team scale.

Google introduced the Developer Knowledge API and an open-source MCP Server on February 4, 2026. The tools are meant to connect internal documentation, public URLs, and other team knowledge sources to Gemini Code Assist and AI-agent workflows with less custom plumbing.

A Show HN post for agent-browser-protocol argues that many browser-agent failures are harness failures caused by stale state. HN discussion centered on the project's freeze-after-action design, Chromium maintenance cost, and the reported 90.5% Online Mind2Web score with Opus 4.6.

GitHub used X on March 9, 2026 to resurface its guide to building reliable multi-agent systems. The company argues that most failures come from missing structure, and recommends typed schemas, action schemas, and Model Context Protocol as the core engineering controls.

A merged llama.cpp PR adds MCP server selection, tool calls, prompts, resources, and an agentic loop to the WebUI stack, moving local inference closer to full agent workflows.

Google's Chrome team has released an early preview of WebMCP, a new web standard enabling direct communication between websites and AI agents. Site owners can now explicitly define how AI agents interact with their services, replacing unreliable DOM scraping with structured APIs.

Developer Eric Holmes argues that MCP is already dying, claiming LLMs already excel at using CLI tools without a special protocol. He makes a strong case that CLIs compose better, debug easier, and work with existing auth systems.

A Hacker News thread highlighted Context Mode, an MCP server that reports reducing Claude Code tool-output context usage from 315 KB to 5.4 KB in tested workflows.