Skip to main content
A connected agent can call every tool in the catalogue. A skill tells it how to use them well: read before writing, quote excerpts from the read, keep human edits, evaluate a comment before acting on it, and never mistake “threads resolved” for “approved”. Sidecars publishes four skills in the Agent Skills format, hosted on this site.

sidecars-documents

Find or create a document, read it with its version, make the smallest edit, and run a resumable review loop over its comment threads.

sidecars-plans

Write a plan as a Sidecars document: where it lives, a template sized to the task, review rounds, version-pinned approval, hand-off.

sidecars-themes

Give pages the workspace’s look: read the theme first, write with its variables, change the look by editing the theme, offer to re-theme.

sidecars-automations

Create, revise and operate automations: learn the vocabulary, write the plan JSON, draft until it validates, read it back, publish or pause.

​
Install

One command installs all four into the agent you pick, at project scope, using the Skills CLI:
Add -g for a global install, -s sidecars-plans to pick one skill, or -a claude-code -a cursor -a codex to choose the agents without a prompt. npx skills list shows what is installed and npx skills update picks up new versions. Or hand the whole setup — connection and skills — to the agent you already have open. It is the same prompt as on Connect your agent; the install step lives in sidecars.ai/agent-setup.txt:

Get your agent set up

Installing skills is separate from connecting the MCP server. The skills tell the agent how to use the Sidecars tools; the tools only exist once the agent is connected and signed in.

​
What a review round looks like

Someone leaves three comments on a plan an agent wrote: one anchored on a paragraph asking for a different default, one on the whole document asking for an acceptance-criteria section, and one asking “should this be a separate service?”. Then they tell the agent to address the review.
  1. read_comments returns the three open threads; read_document returns the current text and its version.
  2. The agent classifies them: two change requests, one decision only a person can make.
  3. For the anchored thread it calls edit_document with mode: "replace_section" and the thread’s quoted passage as section_excerpt, reads the document back, replies in the thread with what changed, and resolves it.
  4. For the whole-document thread it appends the new section, reads back, replies, resolves.
  5. For the question it replies with the trade-off it can see and leaves the thread open.
  6. It reports: two threads resolved, one open for a decision, version N.
Nothing about the document’s approval state changed. That happens when a person says so, and the plan skill records it in the Status line with the version it applies to.

​
For hosts without the Skills CLI

Every skill is a plain SKILL.md file. Save it under your agent’s skills directory (.claude/skills/<name>/SKILL.md for Claude Code, .cursor/skills/<name>/SKILL.md for Cursor, .agents/skills/<name>/SKILL.md for Codex and other hosts that read the shared layout): The machine-readable index is at /.well-known/agent-skills/index.json.