SEEDANCE25APIINDEPENDENT
EN / 中文 获取 API Key

AGENT 模式 / SEEDANCE 2.5 API

在 agent 写代码前,先给它正确的上下文。

本页面面向 Cursor、Claude Code、Codex、ChatGPT 等替人类接入 Seedance 2.5 API 的 agent。复制指令块,让 agent 读取机器文件,等代码就绪后再创建 provider Key。

打开 llms-full.txt 打开 OpenAPI
01 / 上下文先复制这个
AGENT_PROMPT.TXTHUMAN -> AGENT
You are helping me integrate the Seedance 2.5 API.

Primary context: https://seedance25api.io/llms-full.txt
Request schema: https://seedance25api.io/openapi.json

Rules:
- Do not invent or hardcode an API key.
- Ask me to create EVOLINK_API_KEY before any live call: https://evolink.ai/dashboard/keys?utm_source=seedance25api.io&utm_medium=referral&utm_campaign=agent&utm_content=agent-prompt
- Keep the model configurable. Seedance 2.5 API access is expected; verify provider support before production.
- Public rates are not published yet — surface cost as an estimate and flag it.
- Validate duration, resolution, and reference counts before sending.
- Use polling or callback_url for async video jobs.
- Build behind a configurable model parameter so the same code can switch models.

为什么这样有效

agent 拿到稳定的上下文 URL、schema、Key 边界,以及明确的 provider 交接。它可以在不假装已有凭证的前提下准备好代码。

创建 API Key
02 / 机器文件AGENT 应读取什么
/llms.txt给 LLM 和搜索 agent 的简短索引。 /llms-full.txt完整上下文:事实、状态、schema、价格说明、交接。 /agent.md给 coding agent 的 Markdown 清单。 /openapi.json给 API 客户端和代码生成的请求契约。
03 / 事实状态感知的接入
Expected

Provider-mediated access

Seedance 2.5 API availability should be verified with the provider before production use. Build behind a configurable model parameter.

Expected

Reference-to-video model ID

seedance-2.5-reference-to-video

Confirmed in EvoLink API docs; model is pre-release ("not yet available"). References are optional, so a prompt-only call behaves as text-to-video.

Pending

Text-to-video / image-to-video IDs

Pending provider docs

Not yet confirmed. Keep the model parameter configurable so the same code can switch when the IDs are published.

Expected

Launch 限制

4-30 seconds · 480p and 720p

发送前校验时长、quality 和参考数量(最多 30 图 / 10 视频 / 10 音频)。

04 / AGENT 清单正式调用前
  1. Do not invent or hardcode an API key.
  2. Ask the human to create EVOLINK_API_KEY before live calls.
  3. Build behind a configurable model parameter.
  4. Surface cost as an estimate — public rates are not published yet.
  5. Validate duration, quality, and reference counts before sending.
  6. Use polling or callback_url for async completion.
  7. If the model is unavailable for the account, keep the request shape and switch models via the configurable parameter once access is enabled.