OpenClaw mini gateway

A minimal headless gateway that routes agent/LLM traffic on a port, with SSE streaming and OAuth device login.

TypeScriptTelegramLLM gateways

Dev-story01

Challenge

Full agent platforms are heavy when all you need is a small always-on gateway to route model and channel calls for an experiment.

What we did

A trimmed OpenClaw fork with the UI and extras removed — backend only, started with `OPENCLAW_PORT=18889 pnpm start` — so it runs as a lightweight always-on service.

Result

A small gateway you can stand up on a port to front agent or LLM traffic without running the whole OpenClaw stack.

Dev-story article

OpenClaw mini gateway: how the project was built

We started from a full agent runtime built around a chat UI and a zoo of messaging channels, but what we actually needed was none of that: a headless process that binds to a port, forwards coding-agent and LLM traffic, streams tokens back, and handles login. The tension was surgical — keep the routing, streaming, and auth internals intact while cutting out the UI, the web modules, and every channel except the one used as storage, without the app-shaped TypeScript tree refusing to emit a standalone dist.

Sections

Modules

Stack

TypeScript + Telegram

Why the project exists

Full agent platforms are heavy when all you need is a small always-on gateway to route model and channel calls for an experiment.

What was built

A trimmed OpenClaw fork with the UI and extras removed — backend only, started with `OPENCLAW_PORT=18889 pnpm start` — so it runs as a lightweight always-on service.

A minimal backend-only gateway that listens on OPENCLAW_PORT and proxies requests to coding-agent providers, streaming responses over SSE. It carries device-code OAuth login for the upstream providers, per-user context keyed by an x-user-id header, and an auth mode that can be switched off entirely for local runs. Telegram remains only as a storage/channel backend; there is no UI.

Main modules and user path

M01

Backend-only reduction: removed the UI, web modules, and every channel except Telegram, then added a minimal single-purpose gateway entry that binds on OPENCLAW_PORT and avoids the old channel dependencies

M02

Device-code OAuth login helpers for the coding-agent providers, plus an auth.mode of 'none' so a local instance can run with no login at all

M03

SSE streaming was the finicky part — real-time stream flushing and SSE parsing went through a revert-and-reapply cycle before the token pipe held under load

M04

Per-user profile and memory context routed by an x-user-id header, so one gateway serves distinct users without cross-talk

M05

Build surgery to make an app-shaped TS tree emit a clean dist: rewriting .ts import paths to .js, tuning tsc emit config, and defaulting config/state to ~/.openclaw-mini

Architecture and technology decisions

Built with TypeScript, Telegram, LLM gateways.

TypeScript on Node, built with pnpm, fed by device-code OAuth against upstream coding-agent providers and streaming back over Server-Sent Events; the bulk of the work was deletion plus getting tsc to emit a runnable standalone dist.

Result and lessons

A small gateway you can stand up on a port to front agent or LLM traffic without running the whole OpenClaw stack.

A lean headless gateway you start with one env var and one command — no UI, no channel sprawl — that logs in, routes agent traffic, and streams responses; honestly a fork trimmed to a single job rather than a new product.