@fusedio/agentbridge-hub

High-level host-facing API for agentbridge: run local coding-agent CLIs with normalized event streaming and session continuity, from one small interface.

Install

npm install @fusedio/agentbridge-hub @fusedio/agentbridge-adapter-claude-code

Usage

import { createAgentHub } from "@fusedio/agentbridge-hub";
import { createAdapter as claudeCode } from "@fusedio/agentbridge-adapter-claude-code/server";

const hub = createAgentHub({ adapters: [claudeCode()] });

const run = hub.run({
  adapter: "claude_code",
  prompt: "Summarize the files in this directory.",
  cwd: process.cwd(),
  taskKey: "my-task", // runs sharing a key resume the same agent session
});

for await (const ev of run.events()) {
  // ev.type: "log" | "transcript" | "meta" | "status"
}

const result = await run.result; // ExecutionResult { usage, costUsd, summary, ... }

• run.events() replays from the start and is SSE-friendly — pipe it straight to an EventSource endpoint.
• The terminal status event and the resolved result carry token usage (input / output / cached) and costUsd.
• Sessions persist through a pluggable SessionStore (a JSON-file store by default), so a follow-up run with the same taskKey continues the prior conversation. Cancellation is hub.cancel(runId).
• Pass runId to own run identity end-to-end: when set it becomes run.runId, the cancel() key, and the runId on every emitted event — so a host can correlate and cancel using its own durable id. Omit it and the hub mints a UUID. It is per-execution (1:1 with the run() call), not a session key — that stays on taskKey. A supplied id that collides with a still-active run throws DuplicateRunIdError; re-use is allowed once a run finishes.

ESM-only, Node >= 20.

License

MIT © Fused