ir — Claude Code through inferroute

Tiny launcher: one command (ir) spawns claude with the model you pick and your inferroute key, so you don't have to set env vars every time. Your normal claude is never touched — ir only configures the subprocess it spawns.

You choose the model per session — there's no auto-routing. Optionally, an on-device recorder logs your choices privately (see Local recording).

Install

pipx install inferroute
ir setup

That's it. ir setup logs you in (key saved to ~/.config/inferroute/credentials, mode 600) and enables private on-device recording — full by default, since it never leaves your machine; change it with ir add recording --level minimal or turn it off with ir remove recording. Update any time with pipx upgrade inferroute.

Sign up for a key at https://inferroute.ai if you don't have one. (Prefer the steps separately? ir login then ir add recording do the same thing.)

Commands

ir setup            first-time onboarding: log in + optionally enable recording
ir                  open the model picker, then launch (same as `ir choose`)
ir --model NAME     pin a model — short alias or canonical id; any claude flag passes through
ir choose           interactive picker — pick a model, then launch
ir anthropic        escape hatch — plain Claude, your own setup
ir pi / ir opencode same, with Pi or OpenCode instead of Claude Code
ir --plain          standard lane (InferRoute can read the requests); the default for
                    enclave-backed models is the CONFIDENTIAL lane — see below
ir confidential …   show | card | models — re-print the last panel, export it, list models
ir status           personal usage view (TUI)
ir login            save / refresh your inferroute API key
ir help             one-screen explainer

ir add recording    turn on private on-device recording (see below)
ir data show        what's been recorded (counts, models, size)
ir data export DIR  copy the metadata layer  ·  ir data wipe  delete it all
ir remove recording stop recording

NAME is a short alias (minimax, minimax-m3, kimi, glm) or any canonical id (e.g. claude-opus-4-8), which passes through verbatim.

Every command bakes in --dangerously-skip-permissions, which is the right default for an agentic workflow. If you want the prompts back, just use plain claude.

What does each command do under the hood?

ir --model NAME      →  ANTHROPIC_BASE_URL=<recorder daemon if running,         \\
                                            else https://api.inferroute.ai>      \\
                        ANTHROPIC_AUTH_TOKEN=$saved_key                          \\
                        exec claude --dangerously-skip-permissions              \\
                                    --model <resolved-id>
ir anthropic         →  exec claude --dangerously-skip-permissions
                        (no env mutation — pure pass-through)

The env vars are scoped to the spawned claude process. Your shell doesn't see them. When the local recorder daemon is running it sits in front of api.inferroute.ai — it records locally, then forwards there (the cloud still does its own backend fallbacks).

Local recording (optional)

ir add recording runs a small daemon on localhost:5005 that records — on this machine only — which model you pick per task and how the turn went, so you (or a future personal router) can learn your preferences. It is never uploaded; we never see it.

• Default is full — choices, outcomes, and prompt text. The prompt text is what actually lets it learn your preferences, and it never leaves this machine. Pick minimal at install for choices + outcomes only (no prompt text), which is lighter but can't train a personal router.
• Inspect it any time with ir data show, copy the metadata layer with ir data export DIR (never raw prompt text), or delete everything with ir data wipe.
• It lives under ~/.inferroute. Remove the daemon with ir remove recording (--purge also deletes the recorded data).

Where your config lives

Privacy

The CLI talks to:

• https://api.inferroute.ai for your traffic (everything except ir anthropic)
• https://inferroute.ai only on ir help — to print the signup URL

It does not phone home, no telemetry. If you enable ir add recording, that data is written only to ~/.inferroute on your machine and is never uploaded — inspect or delete it any time with ir data show / ir data wipe. Whatever your claude --model … sends to whatever upstream is on you.

ir anthropic is literally exec claude plus one flag — the inferroute service sees nothing.

For enclave-backed models (Kimi, GLM, DeepSeek) ir runs on the confidential lane by default — --plain opts out. It goes further than any of the above: your requests are encrypted on your machine to a hardware enclave (Intel TDX + NVIDIA confidential computing) whose attestation your machine verifies itself, against a fresh challenge, before the first byte is sent. The inferroute service relays ciphertext and sees sizes, timing and the model — never the words. What is checked, what is not, and the receipt every session writes: docs/confidential-lane.md. Requires pip install 'inferroute[confidential]'.

Source

https://github.com/InferRoute/inferroute

Issues / feature requests / PRs welcome.

License

MIT — see LICENSE.