AI SDK Provider for Codex CLI

A community provider for Vercel AI SDK v6 that integrates OpenAI's Codex CLI with GPT‑5.1 / GPT‑5.2 class models (gpt-5.1, gpt-5.2, the Codex-specific gpt-5.3-codex / gpt-5.2-codex, the flagship *-codex-max, and the lightweight *-codex-mini slugs) using your ChatGPT Plus/Pro subscription.

This package ships two provider modes:

• codexExec: non-interactive codex exec (spawn a new process per call)

• codexAppServer: persistent codex app-server JSON-RPC client (shared process, true delta streaming, optional stateful threads)

• Works with generateText, streamText, and generateObject

• Uses ChatGPT OAuth from codex login (tokens in ~/.codex/auth.json) or OPENAI_API_KEY

• Node-only (spawns a local process); supports CI and local dev

• v1.0.0: AI SDK v6 stable migration with LanguageModelV3 interface

• v0.5.0: Adds comprehensive logging system with verbose mode and custom logger support

• v0.3.0: Adds comprehensive tool streaming support for monitoring autonomous tool execution

Version Compatibility

| Provider Version | AI SDK Version | NPM Tag | NPM Installation | | ---------------- | -------------- | ----------- | ----------------------------------------------------- | | 1.x.x | v6 | latest | npm i ai-sdk-provider-codex-cli ai@^6.0.0 | | 0.x.x | v5 | ai-sdk-v5 | npm i ai-sdk-provider-codex-cli@ai-sdk-v5 ai@^5.0.0 |

Installation

For AI SDK v6 (default)

1. Install and authenticate Codex CLI

npm i -g @openai/codex
codex login   # or set OPENAI_API_KEY

2. Install provider and AI SDK v6

npm i ai ai-sdk-provider-codex-cli

For AI SDK v5

npm i ai@^5.0.0 ai-sdk-provider-codex-cli@ai-sdk-v5

⚠️ Codex CLI Version: Requires Codex CLI >= 0.105.0 for full support of both provider modes (codexExec and codexAppServer). If you supply your own Codex CLI (global install or custom codexPath), check it with codex --version and upgrade if needed. The optional dependency @openai/codex in this package pulls a compatible version automatically.

npm i -g @openai/codex@latest

Quick Start

Exec provider (codexExec) — process-per-call

import { generateText } from 'ai';
import { codexExec } from 'ai-sdk-provider-codex-cli';

const model = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
  approvalMode: 'on-failure',
  sandboxMode: 'workspace-write',
});

const { text } = await generateText({
  model,
  prompt: 'Reply with a single word: hello.',
});
console.log(text);

App-server provider (createCodexAppServer) — persistent process

import { streamText } from 'ai';
import { createCodexAppServer } from 'ai-sdk-provider-codex-cli';

const provider = createCodexAppServer({
  defaultSettings: {
    minCodexVersion: '0.105.0',
    autoApprove: false,
    personality: 'pragmatic',
  },
});

const { textStream } = await streamText({
  model: provider('gpt-5.3-codex'),
  prompt: 'Write two short lines of encouragement.',
});
for await (const chunk of textStream) process.stdout.write(chunk);

await provider.close();

App-server stateful threads (optional)

By default, codexAppServer is stateless (new ephemeral thread per call). To continue a prior conversation, pass threadId in providerOptions['codex-app-server'].

import { generateText } from 'ai';
import { createCodexAppServer } from 'ai-sdk-provider-codex-cli';

const provider = createCodexAppServer();

const first = await generateText({
  model: provider('gpt-5.3-codex'),
  prompt: 'Start a migration checklist.',
});

const threadId = first.providerMetadata?.['codex-app-server']?.threadId;

const second = await generateText({
  model: provider('gpt-5.3-codex'),
  prompt: 'Continue from step 2.',
  providerOptions: {
    'codex-app-server': { threadId },
  },
});

await provider.close();

Object generation (Zod)

import { generateObject } from 'ai';
import { z } from 'zod';
import { codexExec } from 'ai-sdk-provider-codex-cli';

const schema = z.object({ name: z.string(), age: z.number().int() });
const { object } = await generateObject({
  model: codexExec('gpt-5.3-codex', { allowNpx: true, skipGitRepoCheck: true }),
  schema,
  prompt: 'Generate a small user profile.',
});
console.log(object);

Features

• AI SDK v6 compatible (LanguageModelV3)
• Dual provider architecture:
  • codexExec / createCodexExec for codex exec
  • codexAppServer / createCodexAppServer for codex app-server
• Backward-compatible aliases: codexCli / createCodexCli map to exec mode
• Streaming and non‑streaming
• Configurable logging (v0.5.0+) - Verbose mode, custom loggers, or silent operation
• Tool streaming support (v0.3.0+) - Monitor autonomous tool execution in real-time
• Native JSON Schema support via --output-schema (API-enforced with strict: true)
• JSON object generation with Zod schemas (100-200 fewer tokens per request vs prompt engineering)
• Safe defaults for non‑interactive automation (on-failure, workspace-write, --skip-git-repo-check)
• Fallback to npx @openai/codex when not on PATH (allowNpx)
• Usage tracking from experimental JSON event format
• Image support - Local binary images in both providers, plus remote HTTP/HTTPS image URLs in app-server mode

Image Support

The provider supports multimodal (image) inputs for vision-capable models:

import { generateText } from 'ai';
import { codexExec } from 'ai-sdk-provider-codex-cli';
import { readFileSync } from 'fs';

const model = codexExec('gpt-5.3-codex', { allowNpx: true, skipGitRepoCheck: true });
const imageBuffer = readFileSync('./screenshot.png');

const { text } = await generateText({
  model,
  messages: [
    {
      role: 'user',
      content: [
        { type: 'text', text: 'What do you see in this image?' },
        { type: 'image', image: imageBuffer, mimeType: 'image/png' },
      ],
    },
  ],
});
console.log(text);

Supported image formats:

• Base64 data URL (data:image/png;base64,...)
• Base64 string (without data URL prefix)
• Buffer / Uint8Array / ArrayBuffer

Remote image URLs:

• codexExec mode: HTTP/HTTPS image URLs are not supported (provide binary/image data)
• codexAppServer mode: HTTP/HTTPS image URLs are supported and forwarded to app-server as remote image inputs

Local image data is written to temporary files and passed to Codex CLI via --image (or app-server localImage). Temp files are automatically cleaned up after each request.

See examples/exec/image-support.mjs and examples/app-server/image-support.mjs for complete working examples.

Tool Streaming (v0.3.0+)

The provider supports comprehensive tool streaming, enabling real-time monitoring of Codex CLI's autonomous tool execution:

import { streamText } from 'ai';
import { codexExec } from 'ai-sdk-provider-codex-cli';

const result = await streamText({
  model: codexExec('gpt-5.3-codex', { allowNpx: true, skipGitRepoCheck: true }),
  prompt: 'List files and count lines in the largest one',
});

for await (const part of result.fullStream) {
  if (part.type === 'tool-call') {
    console.log('🔧 Tool:', part.toolName);
  }
  if (part.type === 'tool-result') {
    console.log('✅ Result:', part.result);
  }
}

What you get:

• Tool invocation events when Codex starts executing tools (exec, patch, web_search, mcp_tool_call)
• Tool input tracking with full parameter visibility
• Tool result events with complete output payloads
• providerExecuted: true on all tool calls (Codex executes autonomously, app doesn't need to)

Current behavior:

• codexExec: tool outputs are delivered in final tool-result events.
• codexAppServer: when Codex emits tool output delta notifications, the provider surfaces tool-result parts with result.type === 'output-delta' during streaming.

See examples/exec/streaming-tool-calls.mjs, examples/exec/streaming-multiple-tools.mjs, and their app-server counterparts under examples/app-server/.

Logging Configuration (v0.5.0+)

Control logging verbosity and integrate with your observability stack:

import { codexExec } from 'ai-sdk-provider-codex-cli';

// Default: warn/error only (clean production output)
const model = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
});

// Verbose mode: enable debug/info logs for troubleshooting
const verboseModel = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
  verbose: true, // Shows all log levels
});

// Custom logger: integrate with Winston, Pino, Datadog, etc.
const customModel = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
  verbose: true,
  logger: {
    debug: (msg) => myLogger.debug('Codex:', msg),
    info: (msg) => myLogger.info('Codex:', msg),
    warn: (msg) => myLogger.warn('Codex:', msg),
    error: (msg) => myLogger.error('Codex:', msg),
  },
});

// Silent: disable all logging
const silentModel = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
  logger: false, // No logs at all
});

Log Levels:

• debug: Detailed execution traces (verbose mode only)
• info: General execution flow (verbose mode only)
• warn: Warnings and misconfigurations (always shown)
• error: Errors and failures (always shown)

Default Logger: Adds level tags [DEBUG], [INFO], [WARN], [ERROR] to console output. Use a custom logger or logger: false if you need different formatting.

See examples/exec/logging-*.mjs and examples/app-server/logging-*.mjs for complete examples, and docs/ai-sdk-v5/guide.md for detailed configuration.

Text Streaming behavior

codexExec mode: Incremental streaming is not currently available with codex exec --experimental-json.

The --experimental-json output format (introduced Sept 25, 2025) currently only emits item.completed events with full text content. Incremental streaming via item.updated or delta events is not yet implemented by OpenAI.

What this means in exec mode:

• streamText() works functionally but delivers the entire response in a single chunk after generation completes
• No incremental text deltas—you wait for the full response, then receive it all at once
• The AI SDK's streaming interface is supported, but actual incremental streaming is not available

codexAppServer mode: supports true incremental text deltas via item/agentMessage/delta, so streamText() emits progressively as tokens arrive.

When OpenAI adds streaming support to codex exec --experimental-json, this provider will surface those deltas in exec mode as well.

Documentation

• Getting started, configuration, and troubleshooting live in docs/:
  • docs/ai-sdk-v5/guide.md – full usage guide and examples
  • docs/ai-sdk-v5/configuration.md – all settings and how they map to CLI flags
  • docs/ai-sdk-v5/troubleshooting.md – common issues and fixes
  • docs/ai-sdk-v5/limitations.md – known constraints and behavior differences
  • docs/ai-sdk-v5/migration-app-server-v2.md – app-server v2 migration notes
• See examples/ for runnable scripts covering core usage, streaming, permissions/sandboxing, and object generation.
• Validation helpers:
  • npm run validate:docs checks markdown links and example command paths
  • npm run validate:examples:app-server runs all app-server examples with intent checks
  • npm run validate:full runs build/type/lint/test plus docs and app-server example validation

Authentication

• Preferred: ChatGPT OAuth via codex login (stores tokens at ~/.codex/auth.json)
• Alternative: export OPENAI_API_KEY in the provider’s env settings (forwarded to the spawned process)

Configuration (high level)

• allowNpx: If true, falls back to npx -y @openai/codex when Codex is not on PATH
• cwd: Working directory for Codex
• addDirs: Extra directories Codex may read/write (repeats --add-dir)
• Autonomy/sandbox:
  • fullAuto (equivalent to --full-auto)
  • dangerouslyBypassApprovalsAndSandbox (bypass approvals and sandbox; dangerous)
  • Otherwise the provider writes -c approval_policy=... and -c sandbox_mode=... for you; defaults to on-failure and workspace-write
• skipGitRepoCheck: enable by default for CI/non‑repo contexts
• color: always | never | auto
• outputLastMessageFile: by default the provider sets a temp path and reads it to capture final text reliably
• Logging (v0.5.0+):
  • verbose: Enable debug/info logs (default: false for clean output)
  • logger: Custom logger object or false to disable all logging

See docs/ai-sdk-v5/configuration.md for the full list and examples.

App-server settings highlights

createCodexAppServer({ defaultSettings }) accepts app-server specific options:

• connectionTimeoutMs: initialize handshake timeout
• requestTimeoutMs: default per-request JSON-RPC timeout
• idleTimeoutMs: close idle app-server process after inactivity
• minCodexVersion: minimum supported app-server version (semver)
• includeRawChunks: emit raw JSON-RPC notifications as raw stream parts by default
• serverRequests: typed handlers for server-initiated JSON-RPC requests
• autoApprove: default approval response when no custom handler is provided
• persistExtendedHistory: request extended thread history persistence
• threadMode: stateless (default) or persistent automatic thread reuse
• resume: shorthand to resume an existing thread id
• onSessionCreated: receive a session object for injectMessage() / interrupt()

Per-call app-server overrides use providerOptions['codex-app-server'] (for example threadId, threadMode, includeRawChunks, personality, approvalPolicy, sandboxPolicy, serverRequests, configOverrides).

Additional app-server helpers:

• listModels(): query available models via a temporary app-server process (or use provider.listModels() to query through an existing provider/client)
• tool(), createLocalMcpServer(), createSdkMcpServer(): define and expose local MCP tools

Local MCP security defaults:

• createLocalMcpServer() binds to loopback hosts by default and rejects non-loopback host values unless you set allowNonLoopbackHost: true.
• createLocalMcpServer() generates a per-server bearer token and expects Authorization: Bearer <token> on direct HTTP calls. The token is available at server.config.bearerToken.
• createSdkMcpServer() propagates this auth config automatically, so provider-level MCP wiring works without extra manual headers.
• Without cacheKey, SDK MCP server/tool function identity participates in persistent keying to avoid conflating closure-dependent tool behavior.
• Use createSdkMcpServer({ cacheKey }) when you intentionally recreate equivalent SDK MCP definitions per call and want stable persistent model reuse.

Model Parameters & Advanced Options (v0.4.0+)

Control reasoning effort, verbosity, and advanced Codex features at model creation time:

import { codexExec } from 'ai-sdk-provider-codex-cli';

const model = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  skipGitRepoCheck: true,
  addDirs: ['../shared'],

  // Reasoning & verbosity
  reasoningEffort: 'medium', // none | minimal | low | medium | high | xhigh (xhigh on codex-max and newer models that expose it)
  reasoningSummary: 'auto', // auto | detailed (Note: 'concise' and 'none' are rejected by API)
  reasoningSummaryFormat: 'none', // none | experimental
  modelVerbosity: 'high', // low | medium | high

  // Advanced features
  profile: 'production', // adds --profile production
  oss: false, // adds --oss when true
  webSearch: true, // maps to -c tools.web_search=true

  // MCP servers (stdio + HTTP/RMCP)
  rmcpClient: true, // enables HTTP-based MCP clients (features.rmcp_client=true)
  mcpServers: {
    local: {
      transport: 'stdio',
      command: 'node',
      args: ['tools/mcp.js'],
      env: { API_KEY: process.env.MCP_API_KEY ?? '' },
    },
    docs: {
      transport: 'http',
      url: 'https://mcp.my-org.com',
      bearerTokenEnvVar: 'MCP_BEARER',
      httpHeaders: { 'x-tenant': 'acme' },
    },
  },

  // Generic overrides (maps to -c key=value)
  configOverrides: {
    experimental_resume: '/tmp/session.jsonl',
    sandbox_workspace_write: { network_access: true },
  },
});

Nested override objects are flattened to dotted keys (e.g., the example above emits -c sandbox_workspace_write.network_access=true). Arrays are serialized to JSON strings. MCP server env/header objects flatten the same way (e.g., mcp_servers.docs.http_headers.x-tenant=acme).

Per-call overrides via providerOptions (v0.4.0+)

Override these parameters for individual AI SDK calls using the providerOptions map. Per-call values take precedence over constructor defaults while leaving other settings intact.

import { generateText } from 'ai';
import { codexExec } from 'ai-sdk-provider-codex-cli';

const model = codexExec('gpt-5.3-codex', {
  allowNpx: true,
  reasoningEffort: 'medium',
  modelVerbosity: 'medium',
});

const response = await generateText({
  model,
  prompt: 'Summarize the latest release notes.',
  providerOptions: {
    'codex-cli': {
      reasoningEffort: 'high',
      reasoningSummary: 'detailed',
      textVerbosity: 'high', // AI SDK naming; maps to model_verbosity
      rmcpClient: true,
      mcpServers: {
        scratch: {
          transport: 'stdio',
          command: 'pnpm',
          args: ['mcp', 'serve'],
        },
      },
      configOverrides: {
        experimental_resume: '/tmp/resume.jsonl',
      },
    },
  },
});

Precedence: providerOptions['codex-cli'] > constructor CodexCliSettings > Codex CLI defaults.

App-server per-call overrides use providerOptions['codex-app-server']:

import { createCodexAppServer } from 'ai-sdk-provider-codex-cli';

const appServerProvider = createCodexAppServer();

const response = await generateText({
  model: appServerProvider('gpt-5.3-codex'),
  prompt: 'Continue this task.',
  providerOptions: {
    'codex-app-server': {
      threadId: 'thr_existing',
      personality: 'pragmatic',
      approvalPolicy: 'on-request',
    },
  },
});

Zod Compatibility

• Peer supports zod@^3 || ^4
• Validation logic normalizes v3/v4 error shapes

Limitations

• Node ≥ 18, local process only (no Edge)
• Codex --experimental-json mode emits events rather than streaming deltas; streaming typically yields a final chunk. The CLI provides the final assistant text in the item.completed event, which this provider reads and emits at the end.
• Some AI SDK parameters are unsupported by Codex CLI (e.g., temperature/topP/penalties); the provider surfaces warnings and ignores them

JSON Schema Limitations (v0.2.0+)

⚠️ Important: OpenAI strict mode has limitations:

• Optional fields NOT supported: All fields must be required (no .optional())
• Format validators stripped: .email(), .url(), .uuid() are removed (use descriptions instead)
• Pattern validators stripped: .regex() is removed (use descriptions instead)

See LIMITATIONS.md for comprehensive details and migration guidance.

Disclaimer

This is a community provider and not an official OpenAI or Vercel product. You are responsible for complying with all applicable terms and ensuring safe usage.

License

MIT