Claude Code CLI for OpenClaw: Nearly 4K Stars, But Read the Fine Print First
Almost 4,000 stars on SkillsMP with zero growth in the last seven days. That's the kind of number that makes you stop and look twice — not because the momentum is exciting right now, but because something clearly resonated with a lot of people at some point, and it's worth understanding why before you decide whether to install it.
I spent time going through the SKILL.md, the repo, and the actual setup flow. Here's my honest take.
What This Skill Actually Does
At its core, this skill teaches an OpenClaw agent how to install, authenticate, and invoke Claude Code CLI as a backend model. Instead of your agent making raw Anthropic API calls (which burn through tokens fast when you're feeding it file contents), it routes coding tasks through the claude CLI binary, which handles its own file exploration natively.
The practical result: your OpenClaw agent can reference claude-cli/sonnet-4.6 or claude-cli/opus-4.6 as model targets — either as a primary, a fallback, or a direct exec invocation — and Claude Code does the heavy lifting of reading files, understanding project structure, and making targeted edits.
The Problem It's Solving (And It's a Real One)
If you've ever tried to use an LLM for non-trivial coding tasks via raw API, you know the token math gets ugly fast. Dumping a 500-line file into context, plus system prompt, plus conversation history, plus the response — you're looking at 10k-50k tokens per task before you've even gotten useful output.
Claude Code sidesteps this by using tool calls to read only what it needs, when it needs it. The skill claims ~500 tokens per task versus 10k-50k with raw API. I can't independently verify that exact figure, but the directional claim is accurate — this is how Claude Code is designed to work, and it's one of the main reasons people use it.
The second piece is billing. Claude Code runs on Claude Max subscription (OAuth), not per-token API billing. If you're already paying for Claude Max and you're also running an OpenClaw setup with heavy coding workloads, routing through Claude Code instead of the API is a legitimate cost optimization. That's not marketing fluff — it's just math.
Key Capabilities Worth Noting
1. OAuth token flow is well-documented
The authentication section is probably the strongest part of this skill. Browser-based OAuth with a one-year token (sk-ant-oat01-*) is clearly explained, including the PTY terminal requirement (you need pty: true on your exec tool, which trips people up). The security warnings about not committing tokens are thorough — maybe overly thorough, but better than the alternative.
2. Config patch integration
The ~/.openclaw/config.patch approach for wiring in the CLI backend is clean. You get model aliases, fallback chains, and env var injection in one JSON block. The example showing claude-cli/opus-4.6 as a fallback for the tank agent is a practical, copy-pasteable pattern.
3. CLAUDE.md project context
This is underrated. The skill pushes hard on setting up a CLAUDE.md in every project root, which Claude Code reads automatically at session start. Think of it as a persistent system prompt scoped to the project — tech stack, directory structure, coding standards, critical rules. If you're running Claude Code across multiple projects through an agent, this is how you avoid the model making dumb assumptions about your stack on every new session.
4. MissionDeck integration There's a third mode mentioned — routing through MissionDeck.ai for session tracking via something called JARVIS integration. I have no way to evaluate this independently, and the skill doesn't explain it in much depth. Take it as a bonus if it works for your setup, not a reason to install.
5. Node/npm is the only hard binary dependency
The skill requires Node and npm to be present, which is a reasonable assumption for most developer environments. Beyond that, it's just the @anthropic-ai/claude-code npm package. No exotic toolchain.
Who Should Install This
Install it if:
- You're running OpenClaw >=2026.2 (this is a hard requirement)
- You already have a Claude Max subscription
- You're doing heavy coding workloads through your agents and the API token costs are real
- You want fallback model chains that include a CLI-backed option
- You're comfortable with OAuth flows and managing long-lived tokens in your environment
Skip it if:
- You don't have Claude Max — the skill explicitly states raw API keys (sk-ant-api03-*) will not work, full stop
- You're not on OpenClaw or you're running an older version
- You're in a locked-down environment where browser OAuth flows are impractical
- You're looking for a plug-and-play experience — this requires manual config patching and environment setup
Installation
Add the skill to your Claude skills directory:
# Global install
cp -r claude-code-cli-openclaw ~/.claude/skills/
# Or project-scoped
cp -r claude-code-cli-openclaw .claude/skills/
Then follow the setup sequence: install the npm package, run claude setup-token in a PTY terminal, store the OAuth token in your environment, and apply the config patch via gateway config.patch. The SKILL.md walks through each step with copy-pasteable commands.
Concerns and Limitations — Be Honest With Yourself Here
There's an attribution issue I can't ignore. The SkillsMP listing shows the author as openclaw, but the SKILL.md header says author: Matrix Zion (ProSkillsMD). The GitHub URL points to openclaw/skills, but the install links inside the skill point to ProSkillsMD/skill-claude-code-cli and proskills.md. These are different entities, and it's not clear which one actually owns or maintains this skill.
The repo README is also worth reading carefully. It says skills in the repository are backed up from clawdhub.com, explicitly notes that "there may be suspicious or malicious skills within this repo," and recommends treating it as a historical archive rather than a primary source. That's an unusual disclaimer for a repo hosting a nearly 4K-star skill. I'm not saying this skill is malicious — the SKILL.md content looks legitimate — but the provenance is murky enough that you should audit what you're installing before running it in a production agent environment.
The openclaw: ">=2026.2" requirement is also worth flagging. That's a future-dated version string. Either this skill was written ahead of a release, or the versioning scheme is non-standard. Either way, verify your OpenClaw version before expecting this to work.
Finally, the MissionDeck.ai and JARVIS references are opaque. There's no explanation of what data gets sent where, what the JARVIS integration does, or what the privacy implications are of routing sessions through a third-party dashboard. If you care about that — and you probably should — don't enable that mode without doing your own research first.
Verdict
The underlying concept is solid and the token efficiency argument is real. If you're running OpenClaw with heavy coding workloads and you're already on Claude Max, this skill gives you a legitimate way to route work through Claude Code CLI instead of burning API tokens on file-dumping. The CLAUDE.md pattern alone is worth adopting regardless of whether you use this skill.
But install it with eyes open. Audit the config patch before applying it. Keep your OAuth token out of version control (the skill tells you this repeatedly, and it's worth repeating). Understand that the authorship trail is unclear and the repo comes with its own disclaimer about potentially suspicious content.
I'd rate this a cautious install for the right use case — not because I have specific concerns about this skill's content, but because "cautious" is the right posture when provenance is unclear and you're wiring external CLI tools into your agent's model backend.
SkillsMP listing: claude-code-cli-for-openclaw