5. OpenClaw skill for the Edgible CLI¶
Teach the agent to run the hosting, not just talk about it.
5.0 Why¶
Ask OpenClaw to publish a port or list your apps right now and you will get a confident, invented command. It has a shell on the box, but nothing has told it that edgible exists, what its subcommands are, or that it is already logged into your organisation. It guesses, and a guess that runs is worse than a refusal.
Two obvious fixes are worse. Turning on host bash gives chat an unrestricted shell in order to reach one CLI, which is far more access than the job needs. Reading output on the guest and pasting it into chat yourself removes the point of asking from a phone. A skill is the narrow version: a described, named command surface the model can call, running as the user that is already authenticated, with the guide’s safety rules (never None on the Control UI port) in the skill repo rather than in your memory.
Prove it in the shell first, Control UI second. If Telegram’s /skill fails three chapters from now, you will already know the binary, the login and the model turn work, so the remaining suspect is the channel.
you, in chat Control UI (org door) · Telegram DM · WhatsApp
│
Ubuntu guest OpenClaw Gateway ──► edgible skill ──► edgible CLI (your org)
│
the same box that serves your Edgible apps
Where you run this: the install and the shell check on the Ubuntu guest; the chat check in Control UI: local dashboard on a guest desktop, or the phone on cellular from chapter 3.
5.1 The job¶
You install the edgible skill from openclaw-edgible (not this UX repo). /skill edgible <args> means run edgible <args> on the Gateway host. Prove it twice: shell first (CLI is healthy), then Control UI (the model pastes stdout into the bubble). Shell-first means a later Telegram /skill failure is the channel or the model, not a missing binary.
The skill does not sign you up, create an org, or register a device. Telegram and WhatsApp are chat clients; they are not this skill.
Done when
openclaw skills listshowsedgible(not the old threeedgible-app-*skills).- VM:
edgible whoamiprints org/account. - Control UI:
/skill edgible whoamirepeats that in the bubble (not only a tool card). - Control UI:
/whoamiis OpenClaw identity, not the Edgible block. /skill edgible doctoris a doctor report, not an app list.
Need first: 2. OpenClaw on the VM (loopback Gateway) (Gateway running, edgible auth login). Control UI (local openclaw dashboard or 3. OpenClaw Control UI through Edgible) for the chat test.
Not this chapter: Edgible signup, publishing apps (that is English helpers after the smoke test), or Telegram /skill (prove Control UI first).
5.2 Install¶
On the Gateway host (the Ubuntu VM):
openclaw skills install git:Edgible/openclaw-edgible
openclaw skills list
openclaw gateway restart
You want a row named edgible. Then in Control UI send /new so the session loads the skill.
--force if you are refreshing a git install (openclaw skills update --all only updates ClawHub, not git):
openclaw skills install git:Edgible/openclaw-edgible --force
openclaw gateway restart
If you previously copied edgible-app-create / edgible-app-list / edgible-app-delete into ~/.openclaw/workspace/skills/, remove those folders so OpenClaw is not choosing among four skills.
Copy-install (folder name must be edgible) is in the skill README. Prefer git: so you can --force later.
The helpers live next to SKILL.md (…/skills/edgible/scripts/), not ~/.openclaw/workspace/scripts/.
5.3 Shell: CLI is healthy¶
Same VM, before chat. This does not use the skill; it proves edgible on PATH is logged in.
which edgible
edgible --version
edgible whoami
You want a version, then a block with Profile, Environment, Account, Organization. If whoami fails, fix CLI login (edgible auth login). /skill cannot succeed until this does.
Optional: edgible doctor (diagnostics). That is a real top-level command, not app list.
Not this test: openclaw has no edgible whoami. OpenClaw’s /whoami in chat is a different command (5.4).
5.4 Control UI: skill pass-through¶
Smoke test. Open Control UI (3. OpenClaw Control UI through Edgible or local dashboard). New session (/new). Leave the model picker on Default. Send exactly:
/skill edgible whoami
Success: the bubble contains the same Profile / Environment / Account / Organization you saw in the shell. The model must exec edgible whoami and paste stdout. A tool card with no bubble text is not enough.
Also useful:
/skill edgible --version
/skill edgible doctor
doctor must run edgible doctor, not app list. If the reply invents an app named dcotr / whoami, the skill copy is stale. Run a --force install, restart, then /new.
Do not confuse these¶
| Where | Command | What you get |
|---|---|---|
| VM shell | edgible whoami |
Edgible session (org, account, environment) |
| Control UI | /skill edgible whoami |
Same CLI, via the skill (OpenClaw model turn) |
| Control UI or Telegram | /whoami or /id |
OpenClaw sender id (webchat:… / telegram:123…). Instant. Not Edgible. |
/skill is a full model turn. It can be slow on Gemini free tier (429). The CLI itself is milliseconds. If chat hangs, /stop, then retry in Control UI, not Telegram, for the first skill check. A paid or local primary is 9. Models beyond free Gemini.
Exec may wait for approval (approve-reads). Approve edgible whoami.
Verify¶
-
openclaw skills listshowsedgible(not the old threeedgible-app-*skills). - VM:
edgible whoamiprints org/account. - Control UI:
/skill edgible whoamirepeats that in the bubble (not only a tool card). - Control UI:
/whoamiis OpenClaw identity, not the Edgible block. -
/skill edgible doctoris a doctor report, not an app list.
5.5 After it works¶
English “list my apps on this machine” / “publish this port” / “take that URL down” use the Python helpers. CLI-shaped text stays pass-through (/skill edgible device list --type serving). Create flags and safety (never none on port 18789) are in the skill repo.
Telegram: same /skill edgible whoami in chapter 6 after pairing. Prove it in Control UI first.
Next¶
6. Telegram pocket client for OpenClaw. Other models: 9. Models beyond free Gemini. Series: README.