OpenAI JavaScript SDK
Verified in the repository against the supported API surface.
Connect OpenAI-compatible clients and selected autonomous tools to registry-backed models with explicit verification status. Tsubasa does not persist prompt, output, or source-code content in usage history.
How do I connect an OpenAI-compatible coding agent to Tsubasa?
Preparing stream…
This animated interface is a product demonstration; it does not send a request.01 / MODELS
Tsubasa is built for unrestricted, abliterated, reduced-refusal, and experimental models used in coding and technical workflows. Every public claim comes from the same validated registry that resolves API requests.
Selected model
tsubasa-medium
Selected medium-tier candidate and default alias target: OrcaRouter's gated FP8 Qwen3.8 27B uncensored derivative. Local requests remain fake until exact-revision qualification passes.
This model is not available in this environment.
Known limitations
02 / CLIENTS & CODING TOOLS
For an OpenAI client, change the key, base URL, and model ID. For a coding tool, use only the interface its current release supports. Tsubasa provides model discovery and Chat Completions—not every OpenAI or Anthropic endpoint.
OpenAI-compatible clients
Create a named key, choose a model ID from the catalog, and point the client at api.tsubasa.sh/v1.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["TSUBASA_API_KEY"],
base_url="https://api.tsubasa.sh/v1",
)
stream = client.chat.completions.create(
model="tsubasa",
messages=[{"role": "user", "content": "Hello"}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")Admission acquires a lease and places a worst-case credit hold.
03 / PREPAID BILLING
Card-funded prepaid credit is the standard path. Admission, interruption, and settlement share one durable request identity so concurrent streams cannot spend the same credit twice.
BUILT FOR THE REQUEST PATH
Use case 01
Use a verified or explicitly experimental integration path, then select a model whose documented capabilities fit the workflow.
THE TSUBASA PATH
DATA HANDLING
Tsubasa does not persist prompt, completion, or source-code content in usage history. It retains the bounded metadata needed to account for, secure, and diagnose a request.
0
prompt or completion bodies written to usage history by default
API CONTRACT
The supported surface is deliberately narrow: model discovery and Chat Completions. Each client or tool is labelled by the evidence actually available today.
Verified in the repository against the supported API surface.
Documented client shape; Tsubasa-specific automation is still pending.
Call the JSON API without a client library.
Read incremental Chat Completion chunks.
PREPAID CREDITS
No subscription tier is implied. Choose a one-time amount, and let signed webhook evidence reconcile the balance before it can be spent.
Model
Token profile
One-time credit
Estimated request
$0.001400
About 17,857 requests at the selected token profile.
Estimate uses the selected catalog input/output rates. Settlement uses actual terminal token counts.
The estimator uses catalog rates and example token counts. The ledger settles the terminal token usage for each real request.
Open billingFAQ
No. Tsubasa currently implements model discovery and OpenAI-compatible Chat Completions, including streamed responses. Use only the endpoints shown in the catalog and API examples.
Availability is environment-specific and comes from the validated model registry. Local demonstration models are labelled as demos and do not imply production availability.
You purchase a fixed one-time credit amount. Before an admitted request reaches the provider, Tsubasa reserves its worst-case cost, then settles the hold against terminal token usage.
Review the model evidence, choose a compatible integration, and create a revocable key when you are ready to test the real path.
Familiar agent workflows
Codex, Claude Code, OpenClaude, Hermes Agent, and OpenClaw move through the same authorized red-team task. Tsubasa changes the model route only where the tool's current interface supports it.
Five coding tool terminal recreations rotate automatically while this carousel is visible. Use the pause, previous, and next buttons, or focus this carousel and use the left and right arrow keys.
Create named keys and revoke them independently.
Inspect tokens, latency, status, and settled cost.
Prompt and completion content is not persisted by default. Usage history retains operational metadata such as request ID, model, token counts, latency, cost, status, and timing.
Cancellation stops the response path and preserves partial output. After the provider confirms the job stopped, Tsubasa settles available usage and releases concurrency. If stop cannot be confirmed, the credit hold and concurrency lease remain temporarily reserved until their bounded expiry.
Yes. Create named API keys for separate applications or environments, copy each secret once, and revoke it without changing your dashboard session.
No. The balance changes only after signed payment-provider evidence is verified and reconciled. A browser return URL alone is not treated as payment evidence.