Thread

A local desktop client for coding agents — Claude and OpenAI Codex — in one chat UI.

Pick a project folder, open a thread, and pair with the agent: streaming responses, live tool calls and reasoning, tool-use approvals, Plan/Build modes, and a per-turn git diff viewer. Runs locally — no cloud backend, no accounts.

Built on Electron + a local event-sourced server + a WebSocket RPC, driving the Claude Agent SDK and the OpenAI/Codex agents.

Run

pnpm install
pnpm run dev

The selected model in the composer picks the provider for that turn.

Features

• Streaming chat — token-level output, thinking blocks, and tool/file-edit activity shown live as work items
• Approvals — the agent pauses for permission (Approve / Always allow / Decline); modes: Supervised, Auto-accept edits, Full access
• Plan mode — read-only planning (Shift+Tab to toggle)
• Diffs — per-turn git checkpoints and a working-tree diff panel (inline/split), without touching your real index
• Model picker — Claude and Codex models, with per-model reasoning effort
• Multi-turn sessions — session resume per thread; auto-generated thread titles

The renderer never touches Node/OS directly (contextIsolation: true); everything flows over the WS RPC, with a minimal preload bridge for dialogs and links.

Scripts

pnpm run dev        # Vite dev server + Electron with HMR
pnpm run build      # bundle main / preload / renderer into out/
pnpm run start      # preview the production build
pnpm run typecheck  # tsc over both tsconfigs
pnpm run lint       # eslint