Skip to content

Guide CLIProxyAPI Computing Provider

Jump to bottom
charles edited this page Jul 12, 2026 · 3 revisions

Guide: CLIProxyAPI — Serve GPT-5.x as a Computing Provider

Turn a ChatGPT account into a Swan Chain computing provider in five minutes — no GPU required.

The idea

The computing-provider forwards inference requests to any OpenAI-compatible HTTP endpoint listed in models.json. CLIProxyAPI turns a ChatGPT OAuth session into exactly that — a local OpenAI-compatible server serving gpt-5.5, gpt-5.4, and gpt-5.4-mini.

Neither binary needs modification. The integration is two config files.

swan-inference
      │  WebSocket dispatch
      ▼
computing-provider
      │  POST /v1/chat/completions
      ▼
CLIProxyAPI :8317           ← no GPU, no weights
      │  ChatGPT OAuth
      ▼
gpt-5.5 / gpt-5.4 / gpt-5.4-mini

Step 1 — Run CLIProxyAPI

Clone and build CLIProxyAPI, authenticate with your ChatGPT account, then start the server:

# config.yaml
port: 8317
auth-dir: ~/.cli-proxy-api
api-keys:
  - sk-swan-local
./CLIProxyAPI serve

Verify it lists models:

curl http://localhost:8317/v1/models \
  -H "Authorization: Bearer sk-swan-local"

You should see gpt-5.5, gpt-5.4, gpt-5.4-mini in the response.

Step 2 — Create a provider account

# Sign up
curl -X POST http://localhost:8100/api/v1/user/signup \
  -H "Content-Type: application/json" \
  -d '{"email":"cliproxy@example.com","password":"YourPass1","display_name":"CLIProxy Provider"}'

# Upgrade — note the sk-prov-* key in the response
TOKEN="<token from signup>"
curl -X POST http://localhost:8100/api/v1/user/upgrade-to-provider \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"name":"CLIProxy Provider"}'

Step 3 — Configure the computing-provider

Initialize a new config directory:

computing-provider --repo ./cp-cliproxy init --node-name cliproxy-provider --port 9086

Edit cp-cliproxy/config.toml — update the [Inference] block:

[Inference]
  Enable = true
  WebSocketURL = "ws://localhost:8081"   # no /ws suffix
  ApiKey = "sk-prov-<your-key>"
  Models = ["gpt-5.5", "gpt-5.4", "gpt-5.4-mini"]

Create cp-cliproxy/models.json:

{
  "gpt-5.5": {
    "endpoint": "http://localhost:8317",
    "api_key": "sk-swan-local",
    "category": "text-generation"
  },
  "gpt-5.4": {
    "endpoint": "http://localhost:8317",
    "api_key": "sk-swan-local",
    "category": "text-generation"
  },
  "gpt-5.4-mini": {
    "endpoint": "http://localhost:8317",
    "api_key": "sk-swan-local",
    "category": "text-generation"
  }
}

Note — local_model for name-prefixed deployments: Some swan-inference deployments register models with a vendor prefix (e.g. openai/gpt-5.5) in the marketplace. CLIProxyAPI doesn't recognise that prefix — it only serves gpt-5.5. Add a local_model field to rewrite the model name before the request is forwarded:

{
  "openai/gpt-5.5": {
    "endpoint": "http://localhost:8317",
    "api_key": "sk-swan-local",
    "category": "text-generation",
    "local_model": "gpt-5.5"
  }
}

Without this, inference requests and benchmarks will silently fail (CLIProxyAPI returns a model-not-found error). The file is watched for changes — no restart required.

Step 4 — Start and verify

computing-provider --repo ./cp-cliproxy run

Expected log output:

Connected to Swan Inference
Registration successful: registered successfully
Model gpt-5.5 health changed: unknown -> healthy
Model gpt-5.4 health changed: unknown -> healthy
Model gpt-5.4-mini health changed: unknown -> healthy

End-to-end test through the marketplace:

curl http://localhost:8100/v1/chat/completions \
  -H "Authorization: Bearer sk-swan-localtest" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4-mini",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_tokens": 20
  }'

What you get

Feature Detail
Models in marketplace gpt-5.5, gpt-5.4, gpt-5.4-mini
Health monitoring computing-provider polls GET /v1/models every 30s
Streaming SSE streaming works end-to-end
Hot-reload Edit models.json to add/remove models without restart
Dashboard http://localhost:9086 — provider metrics UI

Limitations

Rate limits — a single ChatGPT account has per-minute caps. For higher throughput, run multiple CLIProxyAPI instances (each with its own account) and register them as separate providers.

Model verification — swan-inference's deterministic/logprob challenges expect a local model checkpoint. Disable verification in dev_mode or use the external endpoint registration path if your deployment enforces it.

Latency — ChatGPT adds ~200–800 ms round-trip vs a local server. Suitable for development and low-volume production.

Troubleshooting

Symptom Fix
Registration failed: invalid API key Check ApiKey in config.toml matches the sk-prov-* from your provider signup
Models stay unknown health Confirm CLIProxyAPI is running and GET /v1/models returns 200
WebSocketURL connection refused Confirm swan-inference is running on port 8081; URL must not have /ws suffix
Requests return wrong model / benchmarks score 0% Add local_model to models.json if the marketplace model name has a prefix (e.g. openai/gpt-5.5) that CLIProxyAPI doesn't recognise
permission denied on docker.sock at startup You are in the docker group but the session hasn't picked it up. Run sg docker -c "computing-provider --repo $CP run" to activate the group for the subprocess, or re-login

Clone this wiki locally