agent-bridge-kit

What this is

A reference implementation of the bridge architecture for running AI agents in production on a bare Linux VPS. Model tier configuration, formatting enforcement, truncation detection, token logging, and a systemd service file.

Companion to agent-spec-kit. The spec tells you what the agent is. This kit is how you actually run it without it failing silently in three different directions at once.

The problem

Most agent tutorials end at "call the API and print the response." Production is different. In production:

• Models ignore formatting directives unless you enforce them at the right point in the message chain
• Models hit token limits and stop mid-response with no warning unless you check finish_reason
• Model IDs drift out of sync with what you think you are running
• Processes crash and restart into port conflicts
• Token costs accumulate invisibly until they do not

None of these are capability problems. They are infrastructure problems. This kit addresses them.

What is in the kit

agent-bridge-kit/
README.md                   (this file)
bridge_reference.py         (annotated reference bridge implementation)
agent.service               (systemd service file with watchdog)
FORMATTING.md               (the formatting directive problem and fix)
MODEL_SELECTION.md          (how to pick models, verify IDs, handle drift)
CHECKLIST.md                (pre-deployment checklist)

How to use it

1. Read FORMATTING.md and MODEL_SELECTION.md first -- these are the failure modes most people hit
2. Use bridge_reference.py as a starting point, not a framework
3. Run through CHECKLIST.md before going live
4. File agent.service under systemd and adapt paths to your setup

Platform compatibility

The bridge uses the OpenAI-compatible API format. Works with any provider that supports it:

• OpenRouter (recommended -- single API, 200+ models)
• Anthropic direct
• Any open-weight model via Ollama or vLLM

Who this is for

Independent operators running agents on their own infrastructure. If you are on managed hosting with a GUI, this is probably more than you need. If you are on a VPS with a terminal and a budget, this is built for you.

About loopwright

We build automation tools for independent operators. Everything we sell, we use ourselves. This kit is the architecture we run in production.

Questions: loopwrightdev@proton.me