CNB provides verified organizations with 500 to 1,166 AI Credits and 1,600 compute core-hours monthly for free. Combined with upstream prompt caching discounts, organizations can unlock up to 200 million free DeepSeek tokens per month. However, the completion endpoint is locked inside isolated workspace networks using ephemeral tokens, and containers are recycled daily:

Client (OpenAI Client / Claude Code / API Gateway)
  │  https://ai.example.com/v1 (Fixed domain, self-hosted Nginx relay)
  ▼
https://<dynamic-subdomain>-9001.cnb.run (CNB port-proxy, auto-synced via /ops/register on boot)
  │
  ▼
node src/server.mjs (Workspace port 9001, cnb2api reverse proxy)
  │  Bearer CNB_TOKEN (Auto-minted by CNB pipeline, internal network)
  ▼
https://api.cnb.cool/<org>/<repo>/-/ai/chat/completions (Official AI core endpoint)

1. Automated Recovery: An internal cron pipeline checks workspace connectivity and triggers start sequences if stopped.
2. Subdomain Sync: On boot, deploy/start.sh registers the new assigned subdomain with /ops/register, immediately updating the Nginx routing upstream.
3. External Fail-safe: An external watchdog (cnb-watchdog.sh) monitors the fixed domain to guard against CI platform cron dormancy.

1. Fork Private: Fork this repository as a private repository on CNB.
2. Configure API Key: In .cnb.yml, define your custom PROXY_KEY in the env: configuration block (e.g., generated via openssl rand -hex 24).
3. Launch Cloud Workspace: Click "Start Cloud Development" on CNB. The build console outputs PROXY_URI=https://<subdomain>-9001.cnb.run once online.

curl https://<subdomain>-9001.cnb.run/v1/chat/completions \
  -H "Authorization: Bearer your-proxy-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": false
  }'

Connect Anthropic Claude Code directly to your CNB DeepSeek endpoints:

export ANTHROPIC_BASE_URL="https://ai.example.com"
export ANTHROPIC_AUTH_TOKEN="your-proxy-key"
export ANTHROPIC_MODEL="deepseek-v4.1-flash"

claude -p "Explain the actor model in distributed computing."

Production billing data measured over months of live operations:

1. Uncached Base Rate: On deepseek-v4.1-flash, an uncached 9,000-token prompt costs ~0.39 credit (~23,000 tokens/credit).
2. Prompt Cache Discount: When prompt prefixes hit the cache, cost falls to ~0.01 credit per request (a 30x discount, yielding ~700,000 tokens/credit).
3. Blended Real-World Workloads: In typical coding assistant, Claude Code, and multi-turn Agent conversations with ~90% prompt cache hit rates: $$\text{Blended Cost} = 10% \times 1 + 90% \times \frac{1}{30} \approx 13% \text{ of base cost} \implies \approx 177,000\text{ tokens / credit}$$ $$\text{Monthly Total} = 1,166\text{ credits} \times 177,000\text{ tokens} \approx \mathbf{206\text{ Million tokens / month}}$$

Inspect real-time quota balances and settled consumption directly from your shell:

npm run quota
# or: npx cnb2api-quota

  CNB Quota Balance  org-slug

  Credits  ███████▋───────────────────  32%   320.0 / 1,000.0 cr
  Dev      ██████▏─────────────────────  25%   406.4 / 1,600.0 core-h
  CI       █▋──────────────────────────   8%    13.0 / 160.0 core-h

  In-flight (unsettled): 12.0 cr, 0.8 core-h
  Remaining balance:     680.0 cr as of 2026-11-09 08:30:00 UTC

• cnb2api-quota --line: Single-line summary, ideal for prompt status lines.
• cnb2api-quota --json: Structured JSON stream for external automation monitors.

Why does the model list initially include GLM or Kimi?
The /v1/models endpoint uses autonomous model discovery. The upstream CNB gateway lacks a standalone model catalog API, so the proxy sniffs actual model names from completion headers. The static fallback list ensures standard client dropdowns function immediately on first start.

Are streaming tool calls and multi-turn functions supported?
Yes. In streaming mode, partial input_json_delta chunks are assembled into complete JSON arguments. In non-streaming mode, standard OpenAI tool_calls payloads are emitted.

How does the Outbound Neutralizer mitigate 11128 errors?
Upstream gateways enforce client-fingerprint moderation on system and assistant texts. CNB2API mitigates this via a three-layer pipeline: ① rewriting signature phrases and adding zero-width spaces to identity tokens; ② normalizing developer roles to system; ③ executing an automated retry ladder folding system prompts into user messages on HTTP 400/11128.

Why does the model not think / reasoning never appear?
Thinking toggles (reasoning_effort, thinking, enable_thinking, …) are stripped by default. On the upstream thinking variant the reasoning phase is billed against max_tokens; with a small budget it consumes everything and the answer never starts (finish_reason=length, empty content). Set CNB2API_ALLOW_THINKING=1 to forward the toggles — the proxy then raises max_tokens to CNB2API_THINKING_MIN_TOKENS (32k) and folds Anthropic thinking.budget_tokens into it, while a larger client budget is kept as-is. Requests without a toggle are unaffected.

Does this service violate CNB platform policies?
No. CNB2API runs within your own authorized cloud workspace, calls documented platform completion APIs using platform-issued ephemeral tokens, and strictly consumes your own allocated organization quotas.

Contributions and improvements are welcome. Please ensure that all test suites pass before submitting changes:

node --test test/*.test.mjs

Licensed under the MIT License.