ChatGPT Agent APIs Alternative: Open Source TypeScript Harness

1import { createAgent } from 'smoke-monkey-harness';
2
3// 1. Initialize with zero cloud dependencies
4const agent = createAgent({
5  provider: 'gemini', // or 'nvidia', 'openai', 'ollama'
6  model: 'gemini-1.5-pro',
7  workspacePath: process.cwd(),
8  permissions: {
9    run_command: 'ask',      // Human-in-the-loop pause
10    replace_file_content: 'allow',
11  },
12});
13
14// 2. Stream tokens and observe tool execution
15agent.on('tool.started', (e) => console.log('Executing:', e.data.toolName));
16agent.on('permission.required', async (e) => {
17  // Grant or deny dangerous bash commands
18  await agent.resolvePermission(e.data.toolCallId, 'allow');
19});
20
21// 3. Run multi-step engineering loop
22const result = await agent.run('Refactor auth module and run jest tests');
23console.log('Completed in phase:', result.phase);

1import OpenAI from 'openai';
2
3const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
4
5// 1. Create assistant on OpenAI cloud servers
6const assistant = await openai.beta.assistants.create({
7  name: 'Coding Assistant',
8  instructions: 'Refactor code and execute tools',
9  model: 'gpt-4o',
10  tools: [{ type: 'code_interpreter' }],
11});
12
13// 2. Manage thread, run, and polling cycles
14const thread = await openai.beta.threads.create();
15const run = await openai.beta.threads.runs.createAndPoll(thread.id, {
16  assistant_id: assistant.id,
17});
18
19// 3. Deal with cloud function calling polling loops
20if (run.status === 'requires_action') {
21  // Manual dispatch of tool outputs back to OpenAI cloud
22  // High latency, restricted filesystem, no native terminal access
23}