Debug what your AI coding agent actually did.
Tether is a local execution debugger for AI coding agents. It captures the prompt, action, file diff, failed command or test, and recovery point in a desktop execution graph for macOS and Linux.
Security review asking where prompts live? Prompts and traces stay localSee the request, the actions, the diff, and the breakage.
Tether turns one messy agent run into an execution graph: user request, tool calls, file reads and writes, shell commands, test results, git diff, errors, replay, and rollback evidence.
Execution graph, not an LLM tree
Nodes represent the agent run: LLM requests, tool calls, file reads, file writes, shell commands, tests, diffs, errors, replay, and rollback.
File and command evidence
Node cards keep changed files, line counts, command output, exit status, timing, tokens, model, and source metadata together.
Replay where supported
Proxy-captured requests can be replayed from a selected node. Local log sources stay inspectable and route clearly to supported recovery flows.
Local evidence vault
macOS secrets use Keychain-aware storage. Linux uses the same local proxy model. Prompts, responses, and traces stay in a local SQLite database that never leaves the machine.
Inspect the debugger before you wire it into your agent.
The right pane mirrors the app inspector. Pick an evidence type and watch the node details change the way they do in a live execution graph.
{
"intent": "order_status",
"confidence": 0.97,
"entities": {
"sentiment": "calm", <- mocked
"order_id": "4471"
}
}Wrap one run and keep the rest of your stack.
Run tether capture -- <agent command>. The wrapper starts or uses the local proxy, captures agent traffic plus tool, file, shell, test, and diff events, then stores the trace for the desktop graph.
Wrap the agent command
Use tether capture -- with supported agent CLIs, custom CLIs, or an adapter-backed framework run.
Run your agent
Tether normalizes LLM requests, tool calls, file changes, shell commands, tests, and errors.
Inspect and recover
Open the desktop graph, click the broken node, inspect file impact and output, then replay supported requests or roll back from local evidence.
Questions engineers ask before installing a local proxy.
Hosted trace tools are useful for team observability. Tether is narrower: a local execution debugger for coding agents that shows the request, actions, file diffs, failed commands, and recovery point on the developer machine.
No. Tether is local-first on macOS and Linux. Prompts, responses, and traces stay in local storage. macOS uses Keychain-aware secrets, Linux uses the same proxy-first local model, and the proxy only talks to the providers you configure.
Use tether capture -- <agent command> or route OpenAI-compatible traffic through the local proxy. Source adapters normalize supported agent CLIs, LangChain, LangGraph, OpenAI/OpenGPT-style agents, and custom CLI events into one execution graph.
Yes. It is a local proxy, so your real code does not need SDK instrumentation. Use it locally while debugging sensitive flows, demos, eval runs, or production-like traces that cannot be copied into a hosted dashboard.
Negligible. Tether runs locally on your machine. The only overhead is the proxy hop, which is usually less than 1ms. Real LLM calls are the bottleneck, not Tether.
Tell me where your agent actually breaks.
Use this only if the alpha is missing a workflow you would need. Specific traces beat vague feature requests.
Install the local execution debugger
before your next agent run.
Free during alpha. Run Tether locally, point your SDK at the proxy, and keep the whole debugging record on your machine.
Tether works on macOS and Linux today, with one shared Rust proxy and platform-specific desktop clients.
curl -fsSL https://tetherapp.vercel.app/install.sh | bashThe installer checks macOS or Linux, downloads the matching release, and only uses sudo when a Linux package install needs it.