HNHacker News
TopNewBestAskShowJobs

apiemotion

3 karma · joined September 16, 2025

submissionscomments
apiemotion··on Beta Agent Gateway – Anyone want to try it out?
Thanks for your comment by the way. I still won't open source the code just yet, but I have written a Medium article about it (https://medium.com/@stevef1uk/managing-ai-agents-and-your-cl...) which provides more explanation and details on the architecture etc. Anyone can vibe code this anyway :-)

In the public repo, for which all the code is open source I have included the script to run the security scan with ZAP Proxy the images and the results.

I have added demos that show OpenClaw running as a docker image within OpenShell with it's external access controlled by the AGW. All code open sourced. Then, I have just added a demo covering how MCP Servers and sensitive operations can be protected (Demo 9). The public repo contains a vanilla MCP Server for demo purposes.

If you have time to read my article you will get my LInkedIn profile and see my background where I have spend 15 years or so working in Financial Services so I am not likely to be a dangerous hacker as I still like to work from time to time and Banks perform lots of background checks ;-)

apiemotion··on Beta Agent Gateway – Anyone want to try it out?
It is in my personal locally hosted gittea repo. All written in Go and yes, I have looked at the code and directed OpenCode to produce it. I also ran security hardening on the code and images so I object to the term AI slop. However, feel free to ignore its a free world, at least in our part of the world.
apiemotion··on I want to try S.Y's GasTown to create code from a Spec without spending any $
The freeride proxy is fairly stable and easily configurable.

stevef@omen:~/dev/freeride$ ./scripts/freeride_proxy_model_stats.py +-------------------------------------------+-------------+----------+ | model | completions | attempts | +-------------------------------------------+-------------+----------+ | meta/llama-3.3-70b-instruct | 1679 | 2623 | | nvidia/llama-3.3-nemotron-super-49b-v1 | 1251 | 1501 | | nvidia/nemotron-3-super-120b-a12b | 555 | 846 | | cerebras/zai-glm-4.7 | 188 | 751 | | nvidia/llama-3.3-nemotron-super-49b-v1.5 | 176 | 295 | | ollama/qwen3-next:80b | 148 | 248 | | cerebras/gpt-oss-120b | 133 | 720 | | cerebras/qwen-3-235b-a22b-instruct-2507 | 129 | 711 | | nvidia/nemotron-3-super-120b-a12b:free | 86 | 137 | | qwen/qwen3-next-80b-a3b-instruct:free | 52 | 62 | | ollama/cogito-2.1:671b | 7 | 102 | | local/qwen3-coder-30b | 3 | 4 | | openrouter/owl-alpha | 3 | 4 | | openrouter/free | 1 | 1 | | meta-llama/llama-3.3-70b-instruct:free | 0 | 192 | | nousresearch/hermes-3-llama-3.1-405b:free | 0 | 192 | | ollama/kimi-k2:1t | 0 | 99 | | ollama/gpt-oss:120b | 0 | 96 | | ollama/qwen3-coder:480b | 0 | 96 | | ollama/qwen3-vl:235b | 0 | 96 | | ollama/qwen3-vl:235b-instruct | 0 | 96 | | ollama/deepseek-v3.1:671b | 0 | 95 | | ollama/mistral-large-3:675b | 0 | 95 | | ollama/qwen3.5:397b | 0 | 95 | | qwen/qwen3-coder:free | 0 | 69 | | google/gemma-4-31b-it:free | 0 | 36 | | nvidia/llama-3.1-nemotron-51b-instruct | 0 | 36 | | nvidia/nemotron-4-340b-instruct | 0 | 36 | | nvidia/nemotron-4-340b-reward | 0 | 36 | | nvidia/nemotron-parse | 0 | 36 | | arcee-ai/trinity-large-thinking:free | 0 | 6 | | openai/gpt-oss-120b:free | 0 | 6 | +-------------------------------------------+-------------+----------+ | TOTAL | 4411 | 9418 | +-------------------------------------------+-------------+----------+

apiemotion··on I want to try S.Y's GasTown to create code from a Spec without spending any $
Confession: all vibe engineered using OpenCode, Gemini & Cursor.

I started by wanting to try GasTown without needing a Claude Subscription. Freeride proxy was developed to use free Cloud hosted models. It started with NVIDIA NIM then I added OpenRouter. Now it also support Cerebras and Ollama Cloud. It manages fallbacks & cool downs. I recently added a local LLM if it is capable. There is flag to spend money if one wants to use the best models. Limited testing on that one as GasTown ate money using the traditional approach.

To get GasTown to work with freeride I needed to make quite a few architectural changes to GasTown. I have added gt-agents to replace Claude Code / OpenRouter. I used NATs instead of tmux for agent messaging and needed to add a gt-agent-console web app to visually see the agents (although gt status & gt feed show progress). I fixed loads of issues I encountered (>120) to try to work around hallucinations with the less capable cloud models. Then, I gave up and added an Orchestrator agent so I could move from the anarchic mailbox / nudge system to a more controlled workflow using configurable finite state machine(s) with Agent's working off configuration files to do one task in each state. The workflow rig is coding agents run using free LLM tiers with failover; Gas Town fork runs them through a spec → design → plan → code → QA pipeline instead of autonomous mail/hooks. This is more deterministic and less LLM intensive approach BUT still uses core GasTown agents e.g. the Mayor, Boot etc and the dolt / beads / gt approach of the original project. I added an Architect, Planner and QA Agents to the project.

I have just managed to get the e2e workflow to work for a non trivial project. It is still being actively developed / debugged as testing can take a while waiting for the LLMs to crack through all the tasks. However, I made more progress in a day or so that I did using the original approach. Using the new approach I add a rig as usual then type: gt mayor workflow start rig-flow --rig <rig name>'.

apiemotion··on Secure_packager: A simple tool for distributing data/models securely
Thanks!
apiemotion··on Local,private emotion detection API for text,powered by fine-tuned Gemma model
Interesting.

My first job was writing names of people in my office on a paper roll printer using Fortran 4. I think they were easier on me than your colleagues were on you ;-)

apiemotion··on Local,private emotion detection API for text,powered by fine-tuned Gemma model
Built an emotion detection API that runs on-prem (Docker container) or hosted to keep your sensitive data private. It's fast, accurate (beating Hugging Face and OpenAI models in our tests), and detects 14 emotions. Check out the demo and code here: https://emotion-service-api.sjfisher.com/ and https://github.com/stevef1uk/emotion-server-demo/blob/master...