Foyle

Foyle is a developer-operations copilot built by Jeremy Lewi, a former Google engineer who co-founded the Kubeflow project and previously worked on YouTube and Google Cloud Platform ML infrastructure. The project sits inside a VSCode notebook through the Runme extension: an engineer writes a natural-language intent in a markup cell, and Foyle suggests the shell commands or API calls in adjacent code cells. As the operator executes (and edits) those commands, Foyle records the trajectory and uses it to ground future suggestions on the same environment.

The architecture is a self-hosted local server (Go for the backend, default port 8877, with a TypeScript frontend layer) that mediates between the notebook and an external LLM provider, with OpenAI as the documented integration. The implementation language split is roughly 85% Go and 13% TypeScript. The learning loop is the central design idea: rather than treating the assistant as a stateless chat client, Foyle persists the pairing of operator intents and the commands that actually worked, then conditions later completions on that history. The result is a notebook-shaped substitute for the wiki of runbooks teams accumulate over time.

The Mozilla award placed Foyle in the first Builders Accelerator cohort on September 23, 2024 with funding of up to $100K plus the 12-week mentorship program. The pitch fits the cohort's local-AI theme: a notebook the operator owns, an LLM key the operator owns, and a feedback record that stays on the operator's machine rather than feeding a vendor model.

In the stack, Foyle sits at the agents and runtime layers. It is a narrow agent (deploy-and-operate, not general coding) that exercises real shell privileges, which is unusual for OSS copilots and pushes against the trust boundary most cloud-hosted assistants sidestep by sandboxing.