Per-user sessions (self-hosted proxy)

The Chat Completions API continues a conversation when you send back metadata.session_id. If you would rather not track session ids at all — for example you have many end users behind one integration — you can deploy a small self-hosted proxy that does it for you: pass a stable user string and each user gets one rolling session, resumed automatically.

The proxy is a Cloudflare Worker that ships in the Manyfold repository under apps/chat-session-proxy. It sits in front of /api/v1/chat/completions, maps user to a session, and is otherwise a transparent passthrough. The core API and your API tokens are unchanged.

Use it

Point your OpenAI-compatible client at the deployed Worker and add user:

import OpenAI from 'openai'

const client = new OpenAI({
    apiKey: process.env.MANYFOLD_API_TOKEN,
    baseURL: 'https://<your-worker>.workers.dev/v1'
})

await client.chat.completions.create({
    model: 'agt_your_agent_id',
    user: 'end-user-123',
    messages: [{ role: 'user', content: 'Continue from last time.' }]
})

Call again with the same user and the conversation resumes — no session id to store. If you do send metadata.session_id yourself, that takes precedence.

Deploy

See apps/chat-session-proxy/README.md in the Manyfold repository for the full setup: create a Workers KV namespace, set your upstream origin, then wrangler deploy. The folder is self-contained and can be copied out and deployed on its own.

Notes

• Your API token is forwarded to Manyfold unchanged; the proxy stores no secret.
• Sessions are namespaced per token and per agent, so the same user talking to different agents stays in separate conversations.
• The mapping is stored in Workers KV (eventually consistent). For strict per-user serialization, the README explains swapping in a Durable Object.