@pivanov/agents-wire

One TypeScript SDK for every local coding agent. Drives Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Factory Droid, Pi, Cline, Kilo, Qwen Code, and Augment Code (Auggie) over the Agent Client Protocol, with cost budgets, structured JSON, tool middleware, multi-agent orchestration, and a Vercel AI SDK provider (LanguageModelV3, ai@^6).

import { agents } from "@pivanov/agents-wire";

const result = await agents.ask("claude", "Refactor src/auth.ts", {
  permission: "auto-allow",
  maxCostUsd: 0.5,
});

console.log(result.text, result.cost?.totalUsd);

Why

Local coding agents are powerful but awkward to drive programmatically. Each one ships a different CLI; their output formats drift; cost tracking is bring-your-own; structured outputs are a prompt-engineering project; and combining several agents takes a lot of glue. agents-wire fixes all of that with one transport and a small, focused API.

Features

• Twelve agents, one API - Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Factory Droid, Pi, Cline, Kilo, Qwen Code, Augment Code (Auggie)
• ask / stream / session - one-shot, streaming async-iterable, multi-turn with shared subprocess
• askJson with Standard Schema - Zod 4 / Valibot / ArkType auto-derived to JSON Schema. Strict CLI channel for Claude (via @pivanov/claude-wire), post-hoc validation for the rest.
• Cost tracker + budgets - per-agent breakdown, runtime pricing table, auto-abort when over budget
• Tool middleware - allowed / blocked / onToolUse decision pipeline plumbed through ACP permission requests
• Permission policies - auto-allow, auto-allow-once, auto-reject, stream (HITL), or custom function
• Orchestration - failover / race / cascade / pool for multi-agent workflows
• Vercel AI SDK provider (LanguageModelV3, ai@^6) - agentModel("claude") drops into streamText / generateText
• Typed errors - WireError with code field plus BudgetExceededError, JsonValidationError, AbortError, CapabilityNotSupportedError
• Fully typed - discriminated TAgentEvent union, full IntelliSense, no any
• Mock + transcript replay - @pivanov/agents-wire/testing for deterministic tests
• CLI - agents-wire ask | ask-json | stream | detect | agents

Install

bun add @pivanov/agents-wire
# or
npm install @pivanov/agents-wire

You also need the agent's CLI installed and authenticated. Run npx @pivanov/agents-wire detect to see which agents are available on your machine.

Agent	How it speaks ACP	Install
claude	bridge (@agentclientprotocol/claude-agent-acp, bundled)	Claude Code + claude /login
codex	bridge (@zed-industries/codex-acp, bundled)	OpenAI Codex CLI on PATH
cursor	native (agent acp)	Cursor Agent CLI
copilot	bridge (@github/copilot --acp, peer install)	npm i -g @github/copilot + gh auth login
gemini	bridge (@google/gemini-cli --acp, peer install)	npm i -g @google/gemini-cli + gemini auth login
opencode	native (opencode acp)	npm i -g opencode-ai
droid	native (droid exec --output-format acp)	npm i -g droid + FACTORY_API_KEY
pi	native (pi acp)	npm i -g @mariozechner/pi-coding-agent
cline	native (cline --acp)	npm i -g cline + cline auth or provider keys
kilo	native (kilo acp)	npm i -g @kilocode/cli + kilo auth login --provider <id> or KILO_API_KEY
qwen	native (qwen --acp --experimental-skills)	npm i -g @qwen-code/qwen-code + qwen auth qwen-oauth or BAILIAN_CODING_PLAN_API_KEY
auggie	native (auggie --acp)	npm i -g @augmentcode/auggie + auggie login (subscription required)

Quick start

One-shot

import { agents } from "@pivanov/agents-wire";

const result = await agents.ask("claude", "Summarize README.md in 3 bullets", {
  cwd: process.cwd(),
  permission: "auto-allow",
  maxCostUsd: 0.25,
});

console.log(result.text);
console.log(result.usage);          // { contextSize, contextUsed, costUsd }
console.log(result.cost?.totalUsd); // SDK-side cumulative cost

Streaming

import { agents } from "@pivanov/agents-wire";

const stream = agents.stream("claude", "Refactor src/auth.ts to use the new session API");

for await (const event of stream) {
  if (event.type === "text-delta") {
    process.stdout.write(event.text);
  }
  if (event.type === "tool-call") {
    console.log("\n[tool]", event.tool, event.input);
  }
}

const final = await stream.result();
console.log("\nfinished:", final.stopReason);

Multi-turn session

import { agents } from "@pivanov/agents-wire";

await using session = await agents.session("codex", { permission: "auto-allow" });

await session.ask("List all TODOs in the repo");
const fix = await session.ask("Now fix the highest-priority one");
console.log(fix.text);

console.log("session cost:", session.cost.snapshot.totalUsd);

Structured JSON with Standard Schema

import { agents } from "@pivanov/agents-wire";
import { z } from "zod";

const Issue = z.object({
  title: z.string(),
  severity: z.enum(["low", "medium", "high"]),
});

const { data } = await agents.askJson("claude", "Read src/auth.ts and report issues", Issue);
console.log(data.title, data.severity);

Works with Zod 4, Valibot (with @valibot/to-json-schema installed), and ArkType. You can pass a raw JSON Schema string for prompt guidance only — askJson will steer the agent toward that shape but cannot validate the response. For validated output, pass a Standard Schema instance.

Tool middleware

const result = await agents.ask("claude", "Fix the build", {
  toolHandler: {
    blocked: ["Bash"],                              // hard block
    onToolUse: async (event) => {
      if (event.tool === "Write" && String(event.input).includes("secrets")) {
        return { decision: "deny", reason: "no secrets" };
      }
      return "allow";
    },
  },
});

Orchestration

Four primitives for combining agents.

failover

Try candidates in order; skip on transient errors, return the first success.

const result = await agents.failover("Classify this ticket", ["claude", "codex", "gemini"]);
console.log(result.winner, result.text);

race

All candidates in parallel; first to finish wins, losers get cancelled.

const result = await agents.race("Classify this ticket", ["claude", "gemini"]);
console.log(result.winner, "lost:", result.losers.map((l) => l.agent));

cascade

Escalation chain. Try cheaper/faster first; fall through if the result fails an accept predicate.

const result = await agents.cascade("Triage this issue", [
  { agent: "claude", options: { model: "haiku" }, accept: (r) => r.text.length > 20 },
  { agent: "claude", options: { model: "sonnet" }, accept: (r) => r.text.length > 50 },
  { agent: "claude", options: { model: "opus" } },
]);
console.log("won at stage", result.winningStageIndex);

pool

Warm subprocess pool with capacity limit. Concurrent prompts share the pool.

await using pool = await agents.pool({ agents: ["claude"], capacity: 4 });

const replies = await Promise.all(
  prompts.map((p) => pool.ask(p)),
);
console.log("total cost:", pool.cost.snapshot.totalUsd);

Vercel AI SDK provider

Use any agent as a LanguageModelV3 for streamText / generateText:

import { streamText } from "ai";
import { agentModel } from "@pivanov/agents-wire/ai-sdk";

const { textStream } = streamText({
  model: agentModel("claude", { permission: "auto-allow" }),
  prompt: "Refactor src/auth.ts",
});

for await (const chunk of textStream) {
  process.stdout.write(chunk);
}

For multi-turn sharing one subprocess across streamText calls:

import { createAgentModelSession } from "@pivanov/agents-wire/ai-sdk";

await using s = await createAgentModelSession("codex");
await streamText({ model: s.model, prompt: "list TODOs" });
await streamText({ model: s.model, prompt: "now fix the highest-priority one" });

CLI

npx @pivanov/agents-wire ask claude --prompt "explain this repo"
npx @pivanov/agents-wire ask-json claude --prompt "extract metadata" --schema-file ./schema.json
npx @pivanov/agents-wire stream gemini --prompt "summarize this PR"
npx @pivanov/agents-wire detect      # list available agents on this machine
npx @pivanov/agents-wire agents      # list all built-in agents

Testing

@pivanov/agents-wire/testing ships a mock agent and transcript record/replay so you can test consumers without spawning real processes.

import { createMockAgent, createRecorder, replayTranscript } from "@pivanov/agents-wire/testing";

const mock = createMockAgent({
  agent: "claude",
  turns: [
    { text: "ok" },
    { text: "porcupine" },
  ],
});

const turn1 = await mock.ask("remember 'porcupine'");  // → "ok"
const turn2 = await mock.ask("what was it?");          // → "porcupine"

Subpath exports

Subpath	What's there
@pivanov/agents-wire	the main facade and types
@pivanov/agents-wire/errors	typed WireError + subclasses + KNOWN_ERROR_CODES
@pivanov/agents-wire/ai-sdk	Vercel AI SDK provider
@pivanov/agents-wire/testing	mock agent + transcript replay
@pivanov/agents-wire/catalog	individual agent definitions + registry
@pivanov/agents-wire/orchestrate	failover, race, cascade, pool

Requirements

• Bun ≥ 1.0 or Node ≥ 22
• POSIX (macOS, Linux, WSL). Native Windows isn't supported.
• The agent CLIs you want to drive must be installed and authenticated on the host machine.

Troubleshooting

Global bin resolution

When agent bridges (@agentclientprotocol/claude-agent-acp, @github/copilot, @google/gemini-cli) are installed globally rather than as workspace deps, agents-wire walks npm root -g, your Node prefix, and conventional system roots (/usr/local, /opt/homebrew, ~/.npm-global, ~/.local). If your global root is in a non-standard location (custom npm config prefix, isolated Volta/asdf install), set:

export AGENTS_WIRE_GLOBAL_NODE_MODULES=/path/to/your/global/node_modules

The override is consulted first; resolution falls through to the standard search if the override doesn't contain the requested package.

Development

bun install
bun run typecheck
bun run lint
bun run --filter '@pivanov/agents-wire' build

# Run the smoke test against your installed Claude
bun packages/agents-wire/tests/smoke.ts
bun packages/agents-wire/tests/smoke-orchestrate.ts

Sponsors

Supported by LogicStar AI.

License

MIT - © Pavel Ivanov