Help Tool

help — Discovery Tool

The help tool returns the tool catalog and current usage guidance. It is the first tool to call at the start of a session: the response tells you exactly which tools are exposed by this server build, what parameters they accept, and which patterns are recommended right now.

Registered as a top-level tool since v4.5.x. Distinct from server_info(action:"help", topic:...) which returns the same kind of content but is routed through server_info.

---

help

Annotations:

Annotation	Value
readOnly	true
destructive	false
idempotent	true

Parameters: (none)

The implementation (registerHelpTool in tools_aliases.go) returns the same serverInstructions string that the server sends during the MCP initialize handshake. The content is regenerated on every call (no caching), so it always reflects the running binary.

---

Why call help first?

The 20-tool surface evolves between releases (minify_js was added in v4.5.7, git in v4.5.2, replace_range mode was added in v4.5.16, etc.). Calling help at session start gives the model:

• The exact list of tools this binary exposes
• The recommended workflow for this version
• The current best practices (REGLA 0 anti-rewrite, --auto-occ modes, step-through undo, etc.)

Avoid hard-coding the tool list in your system prompt — help is the canonical source and is cheaper than dragging the full docs into context.

---

Example: full catalog

help()
// → the full tool catalog with usage rules, workflows, and tips

The response is plain text. The exact contents depend on serverInstructions in the running binary — call once at the start of a session to see what your build returns.

---

Focused help topics

For a narrower, focused response, use the server_info tool with the topic parameter:

server_info({action: "help", topic: "recovery"})
// → Disaster recovery guide: undo_last, restore, restore_trash, purge_trash

Supported topics (from core/help_content.go / help_content.go):

Topic	Returns
overview (default)	Quick-start guide + tool summary
workflow	The 4-step efficient edit loop
tools	Full list of tools with one-line descriptions
read	Reading files (modes, base64, ranges)
write	Writing files (atomic, base64)
edit	Edit modes table + REGLA 0 anti-rewrite + OCC
search	search_files and list_directory patterns
batch	batch_operations + project_replace + pipelines
errors	Common errors and recoveries
examples	Practical call patterns
tips	Token-saving tips
recovery	Disaster recovery — step-through undo, restore_trash
all	Everything (comprehensive dump)

Prefer the focused topics when you only need a refresher on one area; use the default help (no topic) when onboarding to a new version.

---

Tool discovery patterns

Pattern 1 — call help once per session

// session_start
help()                  // full catalog → cache in your memory

// later
server_info({action:"help", topic:"recovery"})  // focused

Pattern 2 — use server_info({action:"stats"}) for runtime info

server_info({action: "stats"})
// → build commit + date + cache stats + worker pool + edit telemetry
// → backup directory + total backup count + latest backup
// → "UNDO last edit: backup(action:'undo_last')"

This is the right tool for “is the server healthy and where are my backups?” — not help.

Pattern 3 — use server_info({action:"artifact"}) for last-response replay

server_info({action: "artifact", sub_action: "capture", content: "..."})  // capture
server_info({action: "artifact", sub_action: "write", path: "out.txt"})  // write to file
server_info({action: "artifact", sub_action: "info"})                     // last artifact info

This is independent of help — it’s how the server exposes a tiny scratch buffer for content the model produced but did not have a destination for.

---

Related

• Core Tools — full tool matrix (20 tools)
• Claude instructions — when to call help and how to use the response
• Changelog — when each tool was added

---

Version: 4.5.29 Last updated: 2026-07-11