coding-agent-runner

English | 简体中文

Run local coding-agent CLIs from Node.js through one small Node-only API.

coding-agent-runner wraps the process and protocol details for:

• codex: Codex CLI via codex app-server --listen stdio://
• claude: Claude Code via native stream-json
• cursor: Cursor Agent via cursor-agent acp
• opencode: OpenCode via opencode acp
• pi: Pi via pi-acp

It is positioned next to ACP-style local agent interoperability, but with a friendlier application API. Instead of making every app deal with provider-specific commands, stdio transports, JSON-RPC shapes, stream-json differences, cancellation, and session ids directly, this package exposes a small SDK surface: pick a provider, pass cwd and prompt, then consume normalized events or a final result.

It does not provide a UI, database, task queue, sandbox, credential manager, or memory layer. It starts local CLI processes, speaks their stdio protocols, normalizes streaming events, and returns turn results that local apps can render or store however they want.

When To Use It

Use this package when you are building a local app, desktop app, daemon, or developer tool that needs to call installed coding agents from Node.js.

Good fits:

• Run Codex or Claude Code from a desktop app.
• Stream Cursor/OpenCode/Pi tool progress into your own UI.
• Detect which coding CLIs are installed on a user's machine.
• Keep a lightweight multi-turn session without adopting a larger agent platform.

Not a fit:

• Browser-only applications.
• Hosted LLM API calls.
• Sandboxed execution by default.
• Provider installation, login, billing, or credential management.

Install

npm install coding-agent-runner

Install and authenticate the underlying CLI separately. For example, codex, claude, cursor-agent, opencode, or pi-acp must be available on PATH.

Runtime requirements:

• Node.js 20 or newer
• ESM project or dynamic import()
• At least one supported provider CLI installed locally

Quick Start

import { runCliAgent } from "coding-agent-runner";

const result = await runCliAgent({
  provider: "codex",
  cwd: process.cwd(),
  model: "gpt-5.5",
  prompt: "Inspect this repository and summarize the test command.",
});

console.log(result.output);

runCliAgent() is the simplest API. It creates a provider process, runs one prompt, consumes the stream, closes the process, and returns the final output.

Model Selection

Pass model on the top-level runner options:

await runCliAgent({
  provider: "claude",
  cwd: process.cwd(),
  model: "sonnet",
  prompt: "Review the staged diff.",
});

model is passed directly to Codex thread startup and Claude Code's --model flag. ACP providers (cursor, opencode, and pi) expose model switching inconsistently today; when a provider supports a model flag or env var, pass it through spawn.args or spawn.env.

System Prompt

Pass systemPrompt on the top-level runner options:

await runCliAgent({
  provider: "codex",
  cwd: process.cwd(),
  model: "gpt-5.5",
  systemPrompt: "You are a concise senior TypeScript reviewer.",
  prompt: "Review this repository.",
});

Provider behavior:

• Codex: mapped to app-server developerInstructions.
• Claude: mapped to Claude Code --append-system-prompt.
• ACP providers (cursor, opencode, and pi): wrapped into the prompt as explicit <system> and <user> blocks because ACP does not currently define a portable system-prompt field.

MCP Servers

Pass local stdio MCP servers with mcpServers when a turn needs extra tools or context:

await runCliAgent({
  provider: "claude",
  cwd: process.cwd(),
  model: "sonnet",
  mcpServers: [{
    name: "docs",
    command: "npx",
    args: ["-y", "@acme/docs-mcp"],
    env: { DOCS_TOKEN: process.env.DOCS_TOKEN },
  }],
  prompt: "Use the docs MCP server and review this module.",
});

Provider behavior:

• Codex: mapped to app-server config.mcp_servers.
• Claude: written to a temporary Claude Code --mcp-config file.
• ACP providers (cursor, opencode, and pi): passed to ACP session setup as stdio MCP servers.

undefined environment values are dropped before provider config is created.

Skills

Pass local skill references with skills:

await runCliAgent({
  provider: "codex",
  cwd: process.cwd(),
  skills: [{ name: "code-review", path: "/path/to/skills/code-review" }],
  prompt: "Review the staged diff.",
});

Provider behavior:

• Codex: sent as structured app-server skill input blocks before the user text and also referenced in developerInstructions for compatibility with CLI builds that do not materialize ad-hoc skills automatically.
• Claude and ACP providers: added to the system prompt as explicit skill references, instructing the provider to read the skill path when relevant.

Streaming

import { streamCliAgent } from "coding-agent-runner";

for await (const event of streamCliAgent({
  provider: "claude",
  cwd: process.cwd(),
  prompt: "Add tests for the auth module.",
})) {
  if (event.type === "text_delta") process.stdout.write(event.text);
  if (event.type === "tool_start") console.log("tool:", event.name);
}

streamCliAgent() is still one-shot, but it yields normalized progress events as the provider runs.

Stateful Multi-Turn Runner

import { createCodingAgentRunner } from "coding-agent-runner";

const runner = await createCodingAgentRunner({
  provider: "cursor",
  cwd: "/path/to/project",
});

try {
  for await (const event of runner.stream({ prompt: "Inspect the codebase." })) {
    console.log(event);
  }

  for await (const event of runner.stream({ prompt: "Now make the change." })) {
    console.log(event);
  }
} finally {
  await runner.close();
}

The stateful runner keeps the last provider session id in memory and reuses it on the next turn.

Detect Local CLIs

import { detectCliAgents } from "coding-agent-runner";

const agents = await detectCliAgents();
console.table(agents);

Detection is best-effort. It scans PATH, runs each provider's version command with a timeout, and does not try to log in or mutate provider state.

Command Overrides

import { runCliAgent } from "coding-agent-runner";

await runCliAgent({
  provider: "opencode",
  cwd: process.cwd(),
  prompt: "Summarize this package.",
  spawn: {
    command: "/custom/bin/opencode",
    args: ["acp"],
  },
});

Use command overrides when your host app bundles a CLI adapter or stores it outside PATH.

Cancellation

import { runCliAgent } from "coding-agent-runner";

const controller = new AbortController();
const timeout = setTimeout(() => controller.abort(new Error("Timed out")), 30_000);

try {
  await runCliAgent({
    provider: "claude",
    cwd: process.cwd(),
    prompt: "Run the test suite and fix failures.",
    signal: controller.signal,
  });
} finally {
  clearTimeout(timeout);
}

Cancellation is best-effort and provider-specific. The package forwards the abort signal to the active transport and cleans up child processes it owns.

Wrappers And Sandboxes

This package does not sandbox provider processes. If you need isolation, pass a wrapper command:

await runCliAgent({
  provider: "codex",
  cwd: process.cwd(),
  prompt: "Inspect this repository.",
  spawn: {
    wrapper: {
      command: "sandbox-exec",
      args: ["-f", "/path/to/profile.sb"],
    },
  },
});

The wrapper is invoked as:

<wrapper.command> <wrapper.args...> <provider.command> <provider.args...>

Events

The friendly streaming API emits:

type CodingAgentEvent =
  | { type: "text_delta"; text: string }
  | { type: "thinking_start"; id: string }
  | { type: "thinking_delta"; text: string }
  | { type: "thinking_end"; id: string }
  | { type: "tool_start"; id: string; name: string; input?: unknown }
  | { type: "tool_update"; id: string; name: string; input?: unknown; output?: string }
  | { type: "tool_end"; id: string; name: string; output: string; isError: boolean }
  | { type: "done"; output: string; sessionId: string | null; stopReason: string }
  | { type: "error"; error: Error };

Lower-Level API

The package also exports the lower-level building blocks:

• runAgentTurn() for the existing one-shot provider dispatcher
• runCodexTurn() and acquireCodexAppServer()
• runClaudeNative()
• AcpConnection
• mapStreamEventToAgentEvents()
• buildDefaultSpawn() and listProviderConfigs()

Lower-level APIs use the internal provider ids: codex-cli, claude-code-cli, cursor-cli, opencode-cli, and pi-cli.

Provider IDs

Friendly APIs accept short provider ids:

Public id	Default command	Internal id
codex	codex app-server --listen stdio://	codex-cli
claude	claude -p --output-format stream-json --input-format stream-json --verbose	claude-code-cli
cursor	cursor-agent acp	cursor-cli
opencode	opencode acp	opencode-cli
pi	pi-acp	pi-cli

Safety

This library does not sandbox provider processes. The spawned CLI runs with the working directory, environment, credentials, and filesystem permissions you give it. If you need isolation, pass a wrapper command in spawn options or run this package inside your own sandbox/container.

Claude Code permission bypass is not enabled by default. To add --dangerously-skip-permissions, call runClaudeNative() with dangerouslySkipPermissions: true.

Development

npm install
npm run typecheck
npm test
npm run build
npm pack --dry-run

npm run check runs typecheck, tests, and build.

Interactive Demo

Use the demo when you want to type real prompts and watch normalized runner events as they happen, similar to inspecting a lightweight TUI trace:

npm run demo -- codex --model gpt-5.5
npm run demo -- claude --model sonnet
npm run demo -- codex --model gpt-5.5 --system-prompt "You are concise"

Inside the demo, type a prompt and press Enter. It prints streamed answer text plus process events such as thinking_start, thinking_delta, thinking_end, tool_start, tool_update, tool_end, done, sessionId, and elapsed time. The same runner instance is reused, so follow-up prompts continue the same session when the provider supports it.

Codex often emits thinking_start and thinking_end lifecycle events without exposing private reasoning text; the demo shows those lifecycle markers instead of inventing hidden chain-of-thought content.

To pass one input and exit:

npm run demo -- codex --model gpt-5.5 --prompt "Reply with exactly DEMO_OK"

Useful demo commands:

/session   Show the current provider session id
/cwd       Show the working directory
/help      Show interactive commands
/exit      Quit

For machine-readable traces:

npm run demo -- codex --model gpt-5.5 --prompt "List two test commands" --json-events

Real CLI Smoke Tests

The default test suite uses mocks and protocol fixtures so CI does not need local agent logins. To verify a real installed provider on your machine, run the manual smoke scripts:

npm run smoke:claude
npm run smoke:codex
npm run smoke:cursor
npm run smoke:opencode
npm run smoke:pi

Each smoke script builds the package, starts the real local CLI, checks streaming output, verifies one session resume turn, and checks cancellation. The scripts require the underlying CLI to be installed, authenticated, and available on PATH.

To run only the streaming check:

npm run smoke:claude -- --stream-only

To run with an explicit model for Codex or Claude:

npm run smoke:claude -- --model sonnet
npm run smoke:codex -- --model gpt-5.5

To test a CLI outside PATH, pass a command override:

CAR_OPENCODE_COMMAND=/custom/bin/opencode npm run smoke:opencode
PI_ACP_BIN=/custom/bin/pi-acp npm run smoke:pi

Real Feature Smoke Tests

Use these when you want a real, non-mocked check for higher-level runner options:

npm run smoke:real
npm run smoke:real:codex
npm run smoke:real:claude

smoke:real defaults to Codex and Claude Code. It builds the package, starts the real installed CLI, and verifies:

• systemPrompt exact-token behavior
• skills behavior through a temporary local SKILL.md

The script requires the selected CLIs to be installed, authenticated, and available on PATH. You can tune models and providers:

npm run smoke:real -- --providers codex,claude --codex-model gpt-5.5 --claude-model sonnet
npm run smoke:real -- --providers cursor,opencode,pi --skip-missing

To include a real MCP server config in the smoke run:

npm run smoke:real:claude -- \
  --mcp-name docs \
  --mcp-command npx \
  --mcp-args-json '["-y","@acme/docs-mcp"]' \
  --mcp-prompt "Use the docs MCP server and reply with DOCS_OK" \
  --mcp-expected DOCS_OK

License

MIT