Connect your agent

The three values

Setup in any tool

Tool by tool

Where the setting lives in each tool. Menus change between versions, so check the tool's own docs for the exact field names.

API reference

Send your key as Authorization: Bearer sealrelay-… (or x-api-key: sealrelay-…). Requests and responses follow the OpenAI format.

The model

Model id

qwen3.8-27b-uncensoredqwen3.8-27b-uncensored-2026-09-15 pins this version; the plain id always means the newest.

Context window

262,144 tokensPrompt and output together.

Max output

32,768 tokensmax_tokens (or max_completion_tokens) is capped at this, and set to it when you leave it out.

Price per million tokens

$0.30 input · $0.13 cached input · $2.50 output

Input

Text, and imagesImages go inline as data:image/… URLs; links to remote images are refused.

Thinking

Qwen 3.8 27B thinks before it answers. The thinking comes back in the message's (or the stream delta's) reasoning field, apart from content, and is billed as output. Set how much it thinks with reasoning_effort, at the top level of the request or in chat_template_kwargs:

• noneNo thinking: the answer starts right away. Top level only.
• lowShort thinking. minimal is read as low.
• mediumThe default, when reasoning_effort is left out.
• xhighThe longest thinking. high and max are read as xhigh.

"chat_template_kwargs": {"enable_thinking": false} also turns thinking off, whatever the effort. Any other effort is refused with a 400.

Sampling

What you leave out takes the model's own defaults:temperature 1.0, top_p 0.95 and top_k 20. Besides the OpenAI parameters you can set top_k, min_p, repetition_penalty and min_tokens. Fields the API doesn't know, and user and metadata, are dropped before the request reaches the model.

Usage and billing

Every response is billed from its token usage: uncached input, cached input and output, each at the model's price. Agents that send a session id (opencode and pi do) keep a conversation on the server that has its context cached, so more of each turn is billed as cached input. To see usage on a stream, set stream_options.include_usage to true. If you disconnect mid-stream, the tokens generated so far are charged; failed requests are not.

Example

curl https://api.sealrelay.com/v1/chat/completions \
  -H "Authorization: Bearer $SEALRELAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.8-27b-uncensored", "stream": true,
       "reasoning_effort": "low",
       "stream_options": {"include_usage": true},
       "messages": [{"role": "user", "content": "Hello"}]}'

Errors

Errors use the OpenAI shape, with a stable code you can match on. Every response carries an X-Request-Id to quote to support.

• 400invalid_requestThe request is malformed: bad JSON, n other than 1, a remote image URL.
• 400upstream_rejectedThe model server refused it, for example a prompt longer than the context window. Not charged.
• 401invalid_api_keyThe key is missing, wrong, expired or revoked.
• 402insufficient_balanceYour balance can't cover the request. Add credit or redeem a code.
• 402key_cap_reachedThis key hit its monthly cap.
• 403account_suspendedThe account is suspended, for example for breaking the acceptable use policy.
• 404model_not_foundThe model id doesn't exist.
• 413request_too_largeThe request body is larger than 8 MB.
• 429too_many_concurrent_requestsYour account already has 2 requests running. Wait for one to finish.
• 429capacity_exceededEvery model server is full. Retry after the Retry-After header's seconds.
• 429rate_limitedYour account is starting requests too fast. Retry after the Retry-After header's seconds.
• 502upstream_errorThe model server failed. You aren't charged; retry after a moment.