Put this project on 0G

Metadata

• Category: private-computer
• SDK: Not applicable. This skill configures Claude Code to reach the 0G router; it uses no 0G SDK, no wallet and no .env.
• Activation Triggers: "接入 0G", "配置 0G", "把 Claude Code 接到 0G", "用 0G 的模型", "0G 怎么配", "set up 0G PC"
• Supported Host: Claude Code only. It configures a project's .claude/ directory, which Cursor and GitHub Copilot do not read.
• Source: skills/0g-pc-setup/SKILL.md in the 0gfoundation/0g-pc-skills repository, where it is tested and released. Changes are made there first.

One job: get <project>/.claude/ configured so Claude Code talks to the 0G router.

The installer does the writing. What this skill adds is the decision in front of it — which model, and on what basis — plus the two things people get wrong afterwards.

Hard rules

1. Never write ~/.claude/settings.json. The user's global file holds their hooks, plugins, status line and /model choice. Read it if you need to; everything installed here goes inside <project>/.claude/.
2. The key never enters the conversation. You do not ask for it, receive it, read it out of a file, or put it on a command line you run. Hand the user the command and let them paste their own key into their own terminal. If they paste a key into the chat anyway, tell them to rotate it at pc.0g.ai — it is now in a transcript.
3. You do not run the installer. It takes the key as an argument, so running it is the user's step, not yours. Hard rule 2 is why.
4. Do not touch the permission gate. The shipped config points modelOverrides["claude-sonnet-5"] at 0gm-1.0-35b-a3b. In auto mode the safety classifier resolves through the Sonnet tier, not Haiku; put a reasoning model there and every Bash, git and network call times out with "temporarily unavailable" while chat keeps working — which reads as "model fine, tools broken".
5. Every place the user has to decide is a tool call, not a sentence. This file is loaded into whatever model drives the session. "Let the user choose", written as prose, reads as narration and gets walked past. Use AskUserQuestion and let it block. That tool is Claude Code's; in a host without an equivalent blocking prompt, stop and hand the choice back to the user rather than deciding for them.

0 — The one precondition

Run this from the host project root — the directory you want on 0G. These paths resolve against the working directory, so running it inside .0g-skills/ or any other checkout of this repository would configure the wrong project.

python3 - <<'PY'
import json, pathlib

# "在不在 0G 上"看的是**生效配置里的接入点**，不是某个文件在不在。
# 文件可能因为别的原因存在（团队自己的 settings.json），也可能因为新布局而不存在；
# 两种情况下"文件在不在"都答错。Claude Code 后加载 local，所以 local 压过 project。
ROUTER_URL = "https://router-api.0g.ai"

def load(p):
    try:
        d = json.loads(pathlib.Path(p).read_text())
        return d if isinstance(d, dict) else {}
    except Exception:
        return {}

proj = load(".claude/settings.json")
local = load(".claude/settings.local.json")
env = {**(proj.get("env") or {}), **(local.get("env") or {})}

if env.get("ANTHROPIC_BASE_URL") != ROUTER_URL:
    print("absent")
elif (local.get("env") or {}).get("ANTHROPIC_BASE_URL") == ROUTER_URL:
    print("present")
else:
    print("present-old-layout")
PY

present means this project is already configured. Do not reinstall over it: say so, and point at /0g-pc-switch-model (skills/private-computer/0g-pc-switch-model/SKILL.md here) if what they actually want is a different model, or /0g-pc-uninstall (skills/private-computer/0g-pc-uninstall/SKILL.md here) if they want out. Reinstalling to change a model throws away a key for no reason.

present-old-layout means the 0G config sits in .claude/settings.json, which is where an earlier version of the installer put it. It still runs, but the current installer will not touch it and --uninstall will not undo it — those fields have to be moved into .claude/settings.local.json by hand. Say that; do not offer to do it, and do not reinstall.

This is a precondition, not a mode decision — the job never changes with what is on disk, the skill either does it or refuses to start.

1 — Ask the router what it serves

curl -s https://router-api.0g.ai/v1/models | python3 -c "
import json,sys
rows=[m for m in json.load(sys.stdin)['data']
      if 'anthropic' in (m.get('supported_formats') or [])
      and m.get('type') == 'chatbot' and m.get('context_length')]
if not rows: sys.exit('router returned no anthropic-format chat model - stop and report')
for m in sorted(rows, key=lambda x:-(x.get('context_length') or 0)):
    v=m.get('verifiability') if m.get('tee_attested') else None
    tee={'TeeML':'TEE, model in enclave','TeeTLS':'TEE, proxied upstream'}.get(v,'NO TEE')
    a=m.get('architecture') or {}
    img='text+image' if 'image' in (a.get('input_modalities') or []) else 'text only'
    print(f\"{m['id']:22} ctx={m.get('context_length'):>8}  {tee:22} {img}\")"

Only what this prints can be reached. A model the router serves in openai format only needs a local translation layer, which is a different setup and not this one. If the endpoint is unreachable, stop and report — do not fall back to a list written down somewhere, which is how a model retired last month gets recommended today.

Read the three TEE tiers out loud and do not merge them. They are different promises:

This column is the reason someone is here. Flattening it to "TEE ✓" tells them something false about where their prompt gets processed — and TEE is what they are paying for.

Mention text only / text+image as well. The shipped default is text-only: it cannot take an image at all, and fallbackModel does not rescue that, because it fires on the router being overloaded rather than on the request being the wrong shape.

2 — Let them choose, and wait

Put the list to the user with AskUserQuestion, carrying context length, TEE tier and the image column into what they see. Recommend the shipped default, glm-5.3 — TEE with the model in the enclave, the largest context among those, text only — but present it as the default, not as the answer.

Write nothing and suggest no command until the answer comes back.

3 — Hand them the command

curl -fsSL https://raw.githubusercontent.com/0gfoundation/0g-pc-skills/main/install.sh | bash -s claude --key sk-…

Tell them, in this order:

1. Run it from the project they want on 0G. It writes into that folder and nowhere else, so the directory is the one thing that matters.
2. Replace sk-… with their own key, from pc.0g.ai → Dashboard → API Keys. This is their step; hard rule 3.
3. --key - prompts for the key instead, with echo off, so it stays out of shell history. Worth mentioning to anyone who shares a machine or keeps history in a synced dotfile.
4. It writes .claude/settings.local.json (mode 600) and one line in .claude/.gitignore to keep it out of git. .claude/settings.json is not touched — going on 0G is a personal, per-machine choice, and a committed config points teammates who have no 0G key at the 0G router. It checks the key against the router before writing anything, so a rejected key says so with nothing on disk instead of surfacing later as a failure to start. An account with no balance (402) is not a failure: the config is written and correct, and it says so.

If their chosen model is not the shipped default, they install first and then run /0g-pc-switch-model <model> (skills/private-computer/0g-pc-switch-model/SKILL.md here). Do not rewrite the model fields from here. That rewrite has one owner, and duplicating it in two skills is how the two drift apart.

4 — Hand off

Restart Claude Code, then /status: the Base URL should read https://router-api.0g.ai. The startup line [claude-code:unrecognized_model] is expected and harmless — Claude Code does not know 0G's model names.

To confirm the parts that are easy to get wrong:

curl -fsSL https://raw.githubusercontent.com/0gfoundation/0g-pc-skills/main/check-0g.sh | sh

Silence is the pass.

When it goes wrong