Messages API

POST /v1/messages in the Anthropic format: streaming, tools, caching, thinking.

POST https://api.egrtgghfghtytgb.space/v1/messages is the Anthropic Messages API. Point the Anthropic SDK or any Messages API client at Unspent and keep the rest of your code:

python

client = Anthropic(base_url="https://api.egrtgghfghtytgb.space", api_key=os.environ["UNSPENT_API_KEY"])

Request bodies and the anthropic-version and anthropic-beta headers reach the model unchanged. Responses and streams come back unchanged, event by event.

What works

Feature	Notes
Streaming	Server-sent events, forwarded as they arrive
Prompt caching	`cache_control` with the 5-minute and 1-hour TTL, billed at cache prices
Extended thinking	Billed as output tokens
Images and documents	As in the Anthropic API
Your own tools	Tools without `type` or with `type: "custom"`
Client tools	`bash`, `text_editor`, `memory`, `computer` and other tools your client runs, billed as normal tokens
Tool search	`tool_search_tool_regex` and `tool_search_tool_bm25`
Web search	`web_search_*`, billed per search, see Models & pricing
Token counting	`POST /v1/messages/count_tokens`, free

What is rejected

Anything that changes the price in a way Unspent doesn't meter, or runs outside the request on the provider's side, gets 400 invalid_request_error before anything is held or charged:

other tool types: web fetch, code execution, the MCP connector and new server tools;
mcp_servers and container;
speed other than standard (fast mode);
service_tier other than auto or standard_only;
inference_geo other than global;
fallbacks, fallback_credit_token, compaction and compaction edits in context_management.

An unsupported tool type is rejected in the provider's own wording (Input tag '…' found using 'type' does not match any of the expected tags), so clients such as Claude Code recognize it and retry without that tool.

The Batch API and the Files API are not available. Other paths under /v1 return 404.

How a request is billed

Before the model runs, Unspent holds the request's worst-case cost on your API balance with a reserve transaction on Solana. The response starts once that transaction confirms.
The request goes to the model once. Unspent never retries it at the provider.
When the response ends, Unspent prices the tokens actually used and settles onchain: the program burns that many credits and releases the rest of the hold.

You are never charged more than the hold. If the real cost comes out higher, Unspent pays the difference. How holds and costs are computed: Models & pricing.

If your client disconnects mid-stream, the model still runs to the end. You pay for everything it generated, up to max_tokens.

Response headers

Every response carries x-unspent-request-id: the request's onchain id, 32 bytes in hex. With it you can find the request's two transactions, see Onchain.

The provider's request-id, retry-after, x-should-retry and anthropic-* headers are forwarded.

Retries and Idempotency-Key

Send an Idempotency-Key header (8–128 characters) to make retries safe. The key is tied to your wallet and the request body for 24 hours, and a request with it never reaches the model twice:

Retry with the same key	Result
The first request is still running	`409 request_in_progress`
The first request reached the model and finished	`409 already_processed` with its `request_id` and `credits_charged`. Responses are not stored, so it can't be returned again
A different body	`409 idempotency_key_reused`
The first request ended before reaching the model	Runs as a new request

The Anthropic and OpenAI SDKs retry 429 and 5xx responses by default. Unspent's own errors cost nothing, so these retries are safe.

Errors

Unspent's errors use the Anthropic error format here. Errors from the model provider are forwarded unchanged. The full list: Errors.