Torana Edge

Torana Edge

Torana Edge is a **local-first, programmable reverse proxy for AI coding agents**. It sits between your harness (Claude Code, Codex, OpenCode, Aider) and your provider, and gives you a place to observe, redact, route, veto, or rewrite traffic — without replacing the agent or the model.

All request/response mutations are handled by **WebAssembly (WASM) plugins** running in a sandboxed `wazero` runtime, communicating with the host via **Protobuf** serialization. This architecture enables hot-loadable, language-agnostic plugins with zero-downtime updates.

[harness] ←→ [Torana Edge :8080] ←→ [LLM Providers]
             /provider/deepseek/...   → DeepSeek (OpenAI format)
             /provider/anthropic/...  → Anthropic
             /provider/openai/...     → OpenAI
             /provider/gemini/...     → Google Gemini API / Vertex AI

For harnesses that can't be pointed at a base URL — notably the **Antigravity CLI (`agy`)** — Torana also offers an optional TLS-terminating MITM ingress. See [docs/GEMINI_ANTIGRAVITY.md](docs/GEMINI_ANTIGRAVITY.md).

Key Features

• WASM Plugin Ecosystem: Write ABI-v1 plugins in Go or Rust with the torana-plugin-sdk and compile them to .wasm. Hot-loaded, no proxy restart. Both SDKs run through the host's executable conformance harness in CI.
• Operator-approved plugins: A plugin declares the capabilities it wants; it does not receive them. You inspect and approve each bundle, and the approval is bound to that bundle's SHA-256 digest — rebuild or change permissions and it needs approving again. Sandboxed in wazero: a plugin gets the IR and nothing else, no filesystem, no sockets you did not grant.
• Tool-aware compaction: Explicit policies keep source and failure evidence exact while allowing recoverable searches/listings to be reduced deterministically—even on first expo