Scheduled Local AI Jobs on macOS with launchd

Branch·Provider: OpenAI API

Provider: OpenAI API

Direct API calls with fetch. You manage the key, you pick the model.

Get Your Agent to Help

Install the OpenAI API skill so your coding agent knows the current API patterns:

npx skills add jezweb/claude-skills@openai-api

Then ask your agent: "help me build the daily-digest run script using the OpenAI API"

Set Your Key

Get a key from platform.openai.com/api-keys:

export OPENAI_API_KEY="sk-..."
bun run sync

The Pattern

The chat completions endpoint hasn't changed — POST https://api.openai.com/v1/chat/completions with an Authorization: Bearer header:

async function ask(prompt: string): Promise<string> {
  const resp = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      Authorization: `Bearer ${process.env.OPENAI_API_KEY}`,
    },
    body: JSON.stringify({
      model: "gpt-4.1-mini",
      messages: [{ role: "user", content: prompt }],
      max_tokens: 1024,
    }),
  });

  if (!resp.ok) throw new Error(`OpenAI ${resp.status}: ${await resp.text()}`);
  const data = (await resp.json()) as any;
  return data.choices[0].message.content;
}

Current Models (March 2026)

Model	Speed	Cost	Best For
`gpt-4.1-mini`	Fast	Cheapest	Most scheduled jobs — summaries, classification
`gpt-4.1`	Mid	Mid	Complex reasoning
`gpt-5-mini`	Fast	Mid	Latest capabilities, still affordable
`gpt-5.3-codex`	Mid	Higher	Code-heavy tasks

GPT-4o was retired from ChatGPT in Feb 2026 but remains available in the API. Start with gpt-4.1-mini for scheduled jobs — it's fast and cheap.

Error Handling

LLM APIs fail. Rate limits hit. Always retry with backoff:

async function askWithRetry(prompt: string, retries = 3): Promise<string> {
  for (let i = 0; i < retries; i++) {
    try { return await ask(prompt); }
    catch (err) {
      if (i === retries - 1) throw err;
      await Bun.sleep(1000 * Math.pow(2, i));
    }
  }
  throw new Error("unreachable");
}

API Key at Runtime

The manager injects PATH into the plist but not arbitrary env vars. For the API key to be available when launchd runs your job:

// Read from a dotfile — most durable for scheduled jobs
const key = (await Bun.file(`${Bun.env.HOME}/.config/openai-key`).text()).trim();

Test It

bun run src/cli.ts kick daily-digest
bun run src/cli.ts logs daily-digest

Companion Notes

Branch: OpenAI API

Direct API calls with fetch. You manage the key, you pick the model.

Setup

# Set your API key — the manager injects it into the plist EnvironmentVariables
export OPENAI_API_KEY="sk-..."

The key needs to be available when you run bun run sync. The manager captures process.env.PATH into the plist — but custom env vars need to be in the shell environment at sync time OR set in the schedule file.

Option A: Export before sync

export OPENAI_API_KEY="sk-..."
bun run sync

Option B: Add to schedule file

{"type": "scheduled", "calendar": {"Hour": 8}, "env": {"OPENAI_API_KEY": "sk-..."}}

Note: Option B puts the key in a file. Fine for local-only, but don't commit it.

The Pattern

#!/usr/bin/env bun
const OPENAI_API_KEY = process.env.OPENAI_API_KEY;
if (!OPENAI_API_KEY) {
  console.error("[job] OPENAI_API_KEY not set");
  process.exit(1);
}

async function ask(prompt: string): Promise<string> {
  const resp = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      Authorization: `Bearer ${OPENAI_API_KEY}`,
    },
    body: JSON.stringify({
      model: "gpt-4o-mini",
      messages: [{ role: "user", content: prompt }],
      max_tokens: 1024,
    }),
  });

  if (!resp.ok) {
    const err = await resp.text();
    throw new Error(`OpenAI ${resp.status}: ${err}`);
  }

  const data = (await resp.json()) as any;
  return data.choices[0].message.content;
}

// Use it
const result = await ask("Summarize these notes...");
console.log(result);

Model Selection

Model	Cost	Speed	Best For
`gpt-4o-mini`	Cheapest	Fast	Most jobs — classification, summaries, short tasks
`gpt-4o`	Mid	Mid	Complex reasoning, longer outputs
`o3-mini`	Higher	Slower	Multi-step reasoning, code generation

Start with gpt-4o-mini. It handles 90% of scheduled job tasks.

Error Handling

async function askWithRetry(prompt: string, retries = 3): Promise<string> {
  for (let i = 0; i < retries; i++) {
    try {
      return await ask(prompt);
    } catch (err) {
      if (i === retries - 1) throw err;
      const waitMs = 1000 * Math.pow(2, i); // exponential backoff
      console.warn(`[job] Retry ${i + 1}/${retries} in ${waitMs}ms: ${err}`);
      await Bun.sleep(waitMs);
    }
  }
  throw new Error("unreachable");
}

Rate limits, transient errors, timeouts — scheduled jobs WILL hit these. Retry with backoff.

Verification

# Test the API key
curl -s https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY" | head -1

# Test through the job
bun run src/cli.ts kick daily-digest
bun run src/cli.ts logs daily-digest

Previous lesson

Declarative Sync

Next lesson

WatchPaths: Reactive AI