@tanstack/ai-code-mode

Code Mode for TanStack AI — let LLMs write and execute TypeScript in secure sandboxes with typed tool access.

Overview

Code Mode gives your AI agent an execute_typescript tool. Instead of one tool call per action, the LLM writes a small TypeScript program that orchestrates multiple tool calls with loops, conditionals, Promise.all, and data transformations — all running in an isolated sandbox.

Installation

pnpm add @tanstack/ai-code-mode

You also need an isolate driver:

# Node.js (fastest, uses V8 isolates via isolated-vm)
pnpm add @tanstack/ai-isolate-node

# QuickJS WASM (browser-compatible, no native deps)
pnpm add @tanstack/ai-isolate-quickjs

# Cloudflare Workers (edge execution)
pnpm add @tanstack/ai-isolate-cloudflare

Quick Start

import { chat, toolDefinition } from '@tanstack/ai'
import { createCodeMode } from '@tanstack/ai-code-mode'
import { createNodeIsolateDriver } from '@tanstack/ai-isolate-node'
import { z } from 'zod'

// Define tools that the LLM can call from inside the sandbox
const weatherTool = toolDefinition({
  name: 'fetchWeather',
  description: 'Get weather for a city',
  inputSchema: z.object({ location: z.string() }),
  outputSchema: z.object({ temperature: z.number(), condition: z.string() }),
}).server(async ({ location }) => {
  // Your implementation
  return { temperature: 72, condition: 'sunny' }
})

// Create the execute_typescript tool and system prompt
const { tool, systemPrompt } = createCodeMode({
  driver: createNodeIsolateDriver(),
  tools: [weatherTool],
})

const result = await chat({
  adapter: yourAdapter,
  model: 'gpt-4o',
  systemPrompts: ['You are a helpful assistant.', systemPrompt],
  tools: [tool],
  messages: [
    { role: 'user', content: 'Compare weather in Tokyo, Paris, and NYC' },
  ],
})

The LLM will generate code like:

const cities = ['Tokyo', 'Paris', 'NYC']
const results = await Promise.all(
  cities.map((city) => external_fetchWeather({ location: city })),
)
const warmest = results.reduce((prev, curr) =>
  curr.temperature > prev.temperature ? curr : prev,
)
return { warmestCity: warmest.location, temperature: warmest.temperature }

API Reference

createCodeMode(config)

Creates both the execute_typescript tool and its matching system prompt. This is the recommended entry point.

Config:

• driver — An IsolateDriver (Node, QuickJS, or Cloudflare)
• tools — Array of ServerTool or ToolDefinition instances. Exposed as external_* functions in the sandbox
• timeout — Execution timeout in ms (default: 30000)
• memoryLimit — Memory limit in MB (default: 128, supported by Node and QuickJS drivers)
• getSkillBindings — Optional async function returning dynamic bindings

createCodeModeTool(config) / createCodeModeSystemPrompt(config)

Lower-level functions if you need only the tool or only the prompt. createCodeMode calls both internally.

Advanced

These utilities are used internally and exported for custom pipelines:

• stripTypeScript(code) — Strips TypeScript syntax using sucrase (edge-safe, no native binary).
• toolsToBindings(tools, prefix?) — Converts tools to ToolBinding records for sandbox injection.
• generateTypeStubs(bindings, options?) — Generates TypeScript type declarations from tool bindings.

Driver Selection Guide

Custom Events

Code Mode emits custom events during execution that you can observe via the TanStack AI event system:

Models eval (development)

The benchmark lives in a separate workspace package so @tanstack/ai-code-mode does not depend on @tanstack/ai-isolate-node (avoids an Nx build cycle). See models-eval/package.json (@tanstack/ai-code-mode-models-eval).

1. packages/ai-code-mode/models-eval/pull-models.sh — pull recommended Ollama models
2. pnpm --filter @tanstack/ai-code-mode-models-eval eval:capture — run models and capture raw outputs/telemetry only (no judge LLM call)
3. pnpm --filter @tanstack/ai-code-mode-models-eval eval:judge — judge latest captured session from logs (no model rerun)
4. pnpm --filter @tanstack/ai-code-mode-models-eval eval — single-pass run+judge (legacy convenience mode)
5. pnpm --filter @tanstack/ai-code-mode-models-eval eval -- --ollama-only — only Ollama models from eval-config.ts
6. pnpm --filter @tanstack/ai-code-mode-models-eval eval -- --ollama-only --models qwen3-coder — one or more model ids (comma-separated)

Judge-phase flags:

• --judge-latest judge latest captured session
• --rejudge re-run judging even if logs already contain judge fields

The default list omits some small Ollama models that rarely complete code-mode successfully (see comments in eval-config.ts). You can still benchmark them with --models granite4:3b etc. if pulled locally.

Model comparison metrics

The models eval now tracks seven decision-oriented metrics plus an overall rating:

• accuracy (1-10): numerical/factual correctness vs gold report
• comprehensiveness (1-10): whether the response covers everything requested by the user query
• typescriptQuality (1-10): quality/readability/type-safety of generated TypeScript
• codeModeEfficiency (1-10): how efficiently the model uses code-mode/tooling to reach the answer
• speedTier (1-5): relative wall-clock speed against peers in the same category (local or cloud)
• tokenEfficiencyTier (1-5): relative token efficiency against peers in the same category
• stabilityTier (1-5): success consistency over the latest 5 logged runs for that model
• stars (1-3): weighted rollup score across all metrics

Raw run telemetry also includes compile/runtime failures, redundant schema checks, total tool calls, TTFT, token totals, stability sample size/rate, and per-model logs.

Methodology

Canonical output is written to packages/ai-code-mode/models-eval/results.json after each capture or judge run.

• Benchmark: single code-mode benchmark prompt over the in-memory customers / products / purchases dataset
• Primary quality scores (judge): accuracy, comprehensiveness, typescriptQuality, codeModeEfficiency
• Computed comparative scores: speedTier, tokenEfficiencyTier, stabilityTier
• Stability definition: a run is "stable" if it has no top-level run error, produces a non-empty candidate report, and has at least one successful execute_typescript call
• Star rollup weights:
  • accuracy: 25%
  • comprehensiveness: 15%
  • typescriptQuality: 15%
  • codeModeEfficiency (with compile/runtime failure penalty): 10%
  • speedTier: 10%
  • tokenEfficiencyTier: 10%
  • stabilityTier: 15%

Model comparison table

The table below is transcribed from canonical models-eval/results.json (session 2026-03-26T15:38:44.006Z).

Suggested interpretation:

• Local-first: favor stars >= 2 with high speed tier.
• Cloud-first quality: favor high accuracy + typescriptQuality, then compare stars.
• Cost-sensitive: prioritize tokenEfficiencyTier and speedTier together.

License

MIT