vercel-ai

Vercel AI SDK provider for hum — pure, no OC coupling

A LanguageModelV3 provider that lets the Vercel AI SDK drive a hum daemon. Code written against ai / @ai-sdk/* can call hum models without learning thrum.

Propensity

statefulness	richness	wire shape	hides
stateless per-call	lean	Vercel AI SDK LanguageModelV3	pulse, breath, drone, perf-mark, tendril, permission-ask, tool-meta, session-ready, log, echo

Each call mints a fresh sid; history flows only through what the caller re-supplies. No persistent server-side state.

Install

# Once published to npm:
npm install @hum/vercel-ai ai

# Today (workspace / git):
# (in a sibling repo, depend on this dir or vendor it)

Use

import { humProvider } from "@hum/vercel-ai";
import { generateText } from "ai";

const hum = humProvider({
  // socket path; defaults to $XDG_RUNTIME_DIR/hum/hum.sock.thrum
});

const { text } = await generateText({
  model: hum("sonnet"),
  prompt: "explain monorepos in 3 sentences",
});

Streaming:

import { streamText } from "ai";

const { textStream } = await streamText({
  model: hum("sonnet"),
  prompt: "hello",
});
for await (const chunk of textStream) process.stdout.write(chunk);

What flows where

Vercel AI surface	hum chi
generate / stream request	chi:"prompt"
streamed text part	chi:"chunk" (text)
streamed tool-call part	chi:"tool-call"
finish	chi:"finish"
abort	chi:"cancel"
tool result feedback	chi:"tool-result"

Everything else thrum carries (pulses, performance marks, drone events, permission asks) gets dropped at the seam — the SDK has no place to surface them.

Status

Reference implementation. Provider surface tracks @ai-sdk/provider v3+. Older SDK versions need an adapter shim.

See also

• thrum — the npm package this bee imports.
• openai-server — for an HTTP/SSE surface instead of an SDK provider.
• adiled.github.io/hum — docs site.