Claude Code has opened up Claude Mods, which let plugins change even its internal behavior. The first built-in mod, "You should know," is a side agent that watches along and points out what you or Claude missed. Two releases landed back to back, 2.1.286 (October 1, 04:10 KST) and 2.1.287 (October 2, 03:00 KST), and here are four changes that daily users will notice right away.
Claude Mods and the built-in You should know mod
With Claude Mods, plugins can change deeper parts of how Claude Code behaves. A mod is a plugin that uses function hooks, and it is installed and distributed the same way as any other plugin. Anthropic's design discussion describes how function hooks connect to internal behavior and the UI. You can see the source for the built-in mods in the mods directory of the Claude Code repository. The design discussion opened on September 4, and mods officially landed on October 1 with a "We're live!" announcement.
Boris Cherny, who created Claude Code, said on September 15 that Claude Mods were on the way, and that someone had already built a Tetris mod that runs inside Claude.
Source: X @bcherny
The first built-in mod, You should know, is a side agent that watches alongside you. It catches things you missed or that Claude missed and tells you. Turn it on with this command.
/plugin enable cc-plugin-you-should-know@builtinIt works only when telemetry is on in first-party sessions. What plugins can extend has gotten one step wider.
The verify skill runs automatically before a commit
If a project or user skill is named verify, Claude runs it right before a commit. Commits that change only docs and commits that change only tests are exempt. A project skill can live at .claude/skills/verify/SKILL.md, and a user skill at ~/.claude/skills/verify/SKILL.md. For how to write skills, see the Claude Code skills docs.
For example, a skill that bundles lint and tests could look like this.
---
name: verify
description: Run lint and tests to check before committing
---
Run npm run lint and npm test, and if either fails, do not commit and tell me why.Forgetting to verify before a commit becomes structurally less likely.
1M context by default on Bedrock, Vertex, and Foundry
Starting with Claude Code 2.1.287, Opus 4.7 and later and Fable use a 1M context window by default on Bedrock, Vertex, Foundry, and the Claude apps gateway, without the [1m] suffix. To go back to 200K, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1.
Starting with 2.1.286, if a model fallback shrinks the context from 1M to 200K, the notification shows that change. Teams that use Claude Code through a cloud account can work with a long codebase in one pass without extra configuration.
MCP servers can ask you to log in through a URL prompt
Servers that use the MCP 2025-11-25 protocol can show a URL prompt for things like login. If an MCP server fails to connect after you update to Claude Code 2.1.287, add bareElicitationCapability to that server's config entry. For example:
{
"mcpServers": {
"my-server": {
"type": "http",
"url": "https://example.com/mcp",
"bareElicitationCapability": true
}
}
}If MCP suddenly dropped after you updated, check this one line first.
Related reading
- Pi Coding Agent Adds MCP Support: Codemode and Jev Turn Tool Calls Into Code
- Claude Code Effort: When to Use Low and When to Use High or Max, From Thariq's 11 Experiments
- MCP
