Shuttle: in-editor agentic tooling for a private LLM orchestrator
VS Code extension talking to a self-hosted Ollama-backed orchestrator. Tool-calling loop with approval UI for read/write/run, token streaming, image attachments.
The brief
Run a real agentic coding assistant entirely on local hardware, with approvals for any write or shell action, and no data leaving the network.
What I built
- Server-side tool-calling orchestrator using Ollama’s tools API, with per-task event and tool_result queues.
- VS Code client (Shuttle) wiring tool executors for read_file, write_file, run_shell, glob, grep, get_diagnostics and image attachments.
- Approval UI driven by an allowlist setting, with hard prompts for write_file and run_shell regardless of allowlist.
Outcome
Functional offline coding assistant with full audit trail. Now used daily.