Skip to main content

goclaw

Manage GoClaw AI agent gateway servers from Claude Code. Use this skill when the user mentions "goclaw", "gateway server", "AI agent platform", or wants to execute shell commands remotely on a server, manage AI agents, approve agent actions, inspect chat sessions, or administer multi-tenant AI infrastructure. Wraps the `goclaw` CLI to call REST + WebSocket APIs of GoClaw Gateway.

Source facts

Repository
nextlevelbuilder/goclaw-cli
Last source activity
April 17, 2026 at 09:42
Detected SKILL.md language
English
Stars
5
Forks
5

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

File Explorer
22 files

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
goclaw
description
Manage GoClaw AI agent gateway servers from Claude Code. Use this skill when the user mentions "goclaw", "gateway server", "AI agent platform", or wants to execute shell commands remotely on a server, manage AI agents, approve agent actions, inspect chat sessions, or administer multi-tenant AI infrastructure. Wraps the `goclaw` CLI to call REST + WebSocket APIs of GoClaw Gateway.
when_to_use
goclaw CLI operations, remote shell execution via exec tool, AI agent lifecycle, chat session inspection, tenant administration, skill/provider/MCP management, execution approvals
allowed-tools
Bash(goclaw:*)
disable-model-invocation
false
user-invocable
true
argument-hint
<natural-language-intent>
# GoClaw CLI Skill Lets Claude invoke the `goclaw` binary via Bash to interact with a GoClaw Gateway server. Single source of truth = the CLI; this skill teaches Claude the ergonomics. ## Conventions (always apply) 1. **Always append `--output json`** to every goclaw call. Rely on JSON parsing, not table output. 2. **Read auth from `~/.goclaw/config.yaml`** — do NOT accept token via prompt; user must have run `goclaw auth login` beforehand. 3. **Destructive ops require user confirm.** Any command with `--yes` flag or verbs `delete`, `reset`, `revoke`, `unpublish`, `clear`, `rotate`, `deny`, `execute-merge` → ask user to confirm *before* adding `--yes`. 4. **Streaming commands are NOT Bash-friendly.** Refuse to run: `chat` interactive mode, `logs tail`, `auth pair`, `approvals watch`. Suggest polling or user-run-manually alternative. 5. **Exit code 1 after auth error** → suggest `goclaw auth login`, do not retry-loop. 6. **Never hardcode** server URL, tenant ID, agent ID, user ID — use placeholders in examples, actual values from user context. ## Hero use case — execute shell on server Tool `exec` registered as builtin on GoClaw Gateway. Claude invokes via: ```bash goclaw tools invoke exec --param command="ls -la /workspace" --output json # with working directory: goclaw tools invoke exec --param command="npm test" --param working_dir="/workspace/app" --output json ``` Approval flow when server deems command sensitive (package installs, deny-patterns): call hits `approvals` queue. Handle via `references/exec-workflow.md`. ## Navigation — load relevant reference on demand Claude: when user's intent matches, `Read` the listed reference file before constructing commands. | Intent signal | Reference | |---------------|-----------| | exec / run shell / remote command / approvals | [references/exec-workflow.md](references/exec-workflow.md) | | login / token / profile / tenant switch / credentials / api-keys | [references/auth-and-config.md](references/auth-and-config.md) | | agent list/get/create/delete, agent files, instances, wake | [references/agents-core.md](references/agents-core.md) | | agent share, link, delegate, regenerate | [references/agents-advanced.md](references/agents-advanced.md) | | chat with agent, session list/preview/delete | [references/chat-sessions.md](references/chat-sessions.md) | | health, status, logs, traces, usage, metrics | [references/monitoring-ops.md](references/monitoring-ops.md) | | knowledge graph, entity dedup, memory | [references/knowledge-memory.md](references/knowledge-memory.md) | | teams, members, team tasks, workspace files | [references/teams-collaboration.md](references/teams-collaboration.md) | | channels, contacts, pending messages, writers | [references/channels-messaging.md](references/channels-messaging.md) | | export / import / storage (workspace files) | [references/data-movement.md](references/data-movement.md) | | providers, skills, built-in tools list/config, packages | [references/providers-skills-tools.md](references/providers-skills-tools.md) | | cron, heartbeat, device pairing | [references/automation-scheduling.md](references/automation-scheduling.md) | | MCP servers, grants, requests | [references/mcp-integration.md](references/mcp-integration.md) | | tenants, system-config, audit activity, TTS | [references/admin-system.md](references/admin-system.md) | | media upload/download | [references/media.md](references/media.md) | | API documentation browsing | [references/docs-api.md](references/docs-api.md) | ## Compatibility Tested against `goclaw` CLI ≥ 0.3.0. Run `goclaw version` to check. Schema drift caught by `check-drift.sh` in CI. ## Prerequisites - `goclaw` binary in PATH (download from https://github.com/nextlevelbuilder/goclaw-cli/releases) - Authenticated: `goclaw auth login` or paired via `goclaw auth pair` - Permissions granted in `~/.claude/settings.json` (see install.sh) ## Troubleshooting - Command returns 401 → token expired, run `goclaw auth login` - Command returns 403 → user lacks role for resource; check `goclaw whoami --output json` - Command hangs → likely streaming op — Claude should refuse, see convention #4
View on GitHub