This is an experimental repo where I FAFO trying out different workflows, automations, agent configurations and pipelines to figure out an optimal method to simulate backend APIs for popular software (like Linear, Notion, Figma, Intercom).
These simulated backends can then run inside of a sandbox along with pre-populated data and an MCP server on top of them.
Currently this repository houses an attemp at simulating Intercom's backend. This was entirely made by Claude Code / Agent. More on this soon.
So that these APIs can be used with an MCP server in an RL environment.
In terms of RL environments, traditional MCP servers / real-world APIs bring a lot of instability via;
- network latency & instability
- API rate limits
- difficult state reversion
- difficulty in seeding and creating tasks
This makes RL environments very brittle to train on.
Hence, the motivation to simulate these APIs so that they be scaled up for RL training easily.
- A pipeline to create these simulations with minimal / reasonable human review.
- Manual task creation format.
- Training an RL agent on 2 of these backends.