Why not just use print() or logging?

Logging shows fragments. Tether connects the prompt, action, file diff, command output, failed test, and recovery point in one local execution graph.

Can I use this with production code?

Yes. It's a local proxy, so your real code doesn't change. Use it locally for debugging, or keep it running. Tether only stores traces locally and never sends anything anywhere.

How much money does caching actually save?

It depends on your agent. If you're iterating on prompt logic and re-running the same retrieval steps, caching saves you 50-90% of API spend while you debug. Each cached hit costs $0.0000.

Does Tether add latency to my agent?

Negligible. Tether runs locally on your machine. The only overhead is the proxy hop, which is less than 1ms. Real LLM calls are the bottleneck, not Tether.

Can I share traces with my team?

Not yet. Each developer runs their own Tether instance locally. Export as JSON is coming in a later release.

How does Tether intercept LLM calls without SDK changes?

Tether runs a local HTTP proxy on your machine. You point your AI client's base_url at http://localhost:8080/v1 — that's the only change required. Tether forwards requests to the real provider and records everything locally.

Does Tether send my prompts or API keys to the cloud?

No. Tether is fully air-gapped. Your prompts and responses stay on your machine. macOS secrets use Keychain-aware storage, Linux uses the same local proxy model, and traces stay local.

Which LLM providers does Tether support?

Tether supports OpenAI, Anthropic, Ollama, LM Studio, and any provider that accepts an OpenAI-compatible base_url. It also works with LangChain, LangGraph, LlamaIndex, and similar frameworks.

Yes. Tether is free during the alpha period. The core proxy is open source.

Home/Docs/Rust proxy architecture

Architecture

Rust proxy architecture

The backend is a modular monolith: small Rust modules with explicit service boundaries and a single binary.

Install Setup steps

Release: Alpha channel
Runtime: macOS + Linux + local proxy
Docs: 19 guides

Request path

gateway parses the incoming path and chooses OpenAI-compatible or Anthropic upstreams.
context derives model, prompt, preview, and input hashes from provider payloads.
trace ingestion queues persistence work so the proxy hot path does not block on SQLite.
cache stores and replays identical request bodies for local iteration.

Trace services

Module	Responsibility
capture	Build TraceCapture from request metadata and body previews.
summarize	Extract response text, token counts, language, and tool ids.
store	Convert captures and outcomes into trace_calls rows.
query	Read trace snapshots and session lists for the UI API.
replay	Replay selected nodes and invalidate downstream branches.