Claude Code 2.1.287 adds Claude Mods for deeper plugins

Anthropic released Claude Code 2.1.287 on October 1, 2026, and its headline feature lets plugins reach much deeper into the coding agent. The new Claude Mods let a plugin "modify deeper behavior," according to Anthropic's release notes on GitHub. That gives plugin authors far more power, and it gives teams more to review before they install one.

Claude Code is Anthropic's command-line coding agent. It reads a project, edits files and runs commands on a developer's behalf.

What Claude Mods change

Until now, plugins mostly added commands, tools or hooks. A hook is a small script that Claude Code runs at a set moment, such as before a tool runs. Mods go further. They let a plugin change how the agent itself behaves.

The tracking site ccleaks describes mods as "a documented path past shallow hooks." It also calls them a "privileges bump for shared images." Teams that run Claude Code from a shared setup should review which plugins they allow before rolling mods out widely, it advises.

The "You should know" side agent

The release ships one mod of its own. "You should know" is a side agent that watches the session and flags things you or Claude might miss, Anthropic's notes say.

It is off by default. You turn it on with one command:

/plugin enable cc-plugin-you-should-know@builtin

There is a catch. AI TL;DR and ccleaks both report that it works only in first-party sessions with telemetry turned on. That repeats a pattern from September, when Claude Code's AGENTS.md support also turned out to need telemetry.

Other changes in Claude Code 2.1.287

Anthropic lists more than 30 fixes alongside the new features. The main ones reported across the sources:

Area	Change	Reported by
Agents View	New n:<text> filter matches session names and tasks; Enter opens the first match	Anthropic
Telemetry	OpenTelemetry user_prompt events now include a prompt_text field	Anthropic
MCP	URL prompts support the 2025-11-25 protocol version for sign-in flows	Anthropic
Windows	New startup warning for Bash tool and PowerShell conflicts	Anthropic
Shell safety	Writes through symlinks or outside the working tree now need approval	AI TL;DR
Shell safety	Dangerous rm commands keep their safety prompt when output is redirected	claude-code-changelog
Remote sessions	Fast mode no longer stays off; Remote Control messages retry after 30 seconds	claude-code-changelog
Context	Opus 4.7 and later, and Fable, default to a 1M-token context on Bedrock, Vertex, Foundry and the Claude apps gateway	AI TL;DR

MCP, the Model Context Protocol, is the standard Claude Code uses to talk to outside tools and services. OpenTelemetry is an open format for sending usage and performance data to monitoring systems.

The package itself changed shape too. The claude-code-changelog tracker, which compares each npm release, counts 13 fewer bundled prompt files and 8,328 fewer prompt tokens, a 26.6% drop. The bundle grew by 854.6 KB, or 1.6%.

To upgrade, AI TL;DR gives the usual npm command: npm i -g @anthropic-ai/claude-code@2.1.287.

What this means for developers

Review plugins like dependencies. A mod can now change how the agent behaves, not just add a command. Before installing one, read its code or check who maintains it, and pin the version you approved.

Test mods in one session first. ccleaks suggests smoke-testing a single session before a fleet-wide rollout. On shared images, keep an allow list of approved plugins.

Check your telemetry pipeline. If you export Claude Code's OpenTelemetry events, the new prompt_text field is a copy of the existing prompt field, Anthropic's notes say. If you already drop or mask prompt, do the same for prompt_text, as the notes advise.

Decide on telemetry before trying "You should know." The side agent needs telemetry and a first-party session. Teams that keep telemetry off for privacy will not be able to use it.

Expect stricter shell prompts. Writes through symlinks or outside the working tree now ask for approval. Scripts that relied on those writes running silently may stall until someone approves them.