streamkit-ui

Appearance

Anthropic adapter ​

Translates the Anthropic Messages API raw stream events into StreamChunk sequences.

Import ​

typescript
import { fromAnthropic } from "streamkit-ui/adapters/anthropic";
// or (includes all adapters):
import { fromAnthropic } from "streamkit-ui";

Usage ​

typescript
import Anthropic from "@anthropic-ai/sdk";
import { fromAnthropic } from "streamkit-ui/adapters/anthropic";

const client = new Anthropic();

const stream = await client.messages.create({
  model: "claude-opus-4-6",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Hello" }],
  stream: true,
});

const source = fromAnthropic({ stream });

for await (const chunk of source) {
  console.log(chunk);
}

Translation notes ​

Tool call arguments ​

Anthropic's API streams tool-use arguments as input_json_delta events, each carrying a partial JSON string fragment (partial_json). Individual fragments are not safely parseable — a fragment can split a JSON token at an arbitrary byte boundary (e.g. the fragment "cit might be followed by y").

The adapter maintains a Map<index, BlockTracker> for the duration of the stream. On content_block_start for a tool_use block, a tool-call-start chunk is emitted immediately. input_json_delta fragments are accumulated in the tracker's buffer. On content_block_stop, the full buffer is JSON.parse()d and a tool-call-ready chunk is emitted with the complete args.

If the accumulated JSON is malformed (truncated stream, model-emitted invalid JSON), the adapter emits a tool-result chunk with isError: true rather than throwing, so one bad tool call doesn't abort the entire stream.

Events not translated ​

reasoning-delta, thinking-delta, signature-delta, citations-delta, message_start, message_delta, file, and all extended server-tool events are silently dropped — streamkit is a text + tool-call rendering layer, not a full reimplementation of every Anthropic API feature.

Tool results ​

Tool results do not come back through the assistant-turn stream in the Anthropic Messages API. They are sent by the caller as tool_result content blocks in the next user-turn message. This adapter has no tool-result-from-execution case — use useToolCallState with your own tool implementations to handle execution.

MIT License