v0.6.4

• Pair without a terminal: with no terminal to answer prompts, pair asks ssh for none and trusts a host it meets for the first time, so the agents panel can pair in the background. A host whose key has changed is still refused, now with how to forget the old key (#32). The panel side is omacom/omarchy#18641.
• Tell a host that refuses this machine's SSH keys apart from one with SSH off, and point at the SSH agent (like 1Password's) instead of saying to turn SSHD on (#31)
• Name the machine short in a pairing error's first line, so it fits on the model's line in the panel, and set errors apart from the output above them on a terminal (#31, #32)

v0.6.3

• Keep a pi extension, ~/.pi/agent/extensions/omarchy-local.ts, listing every model set up here and every paired machine's, so plain pi and its model picker reach them with no hand-written models.json (#29, #30). The overlays' provider is now omarchy-local, and omarchy-local-<host> for a paired machine, instead of llamacpp. The usage side is omacom/omarchy#18506 and omacom/omarchy#18515.
• Add overlay qwen, so omarchy agent --local can open Qwen Code on a local or paired model (#30). The launcher side is omacom/omarchy#18515.

v0.6.2

• Leave a model on another machine off the list once that machine stops it, and bring it back when it starts again (#25). The panel side is omacom/omarchy#18475 and omacom/omarchy#18486.
• Remove remote start|stop: a model on another machine runs on that machine's terms, and the panel now offers Disconnect instead (#28)
• Drop the catalog version: catalog.toml keeps only its schema, and a release is just this tag (#27)

v0.6.1

• Count the machines using each shared model, through a share proxy in front of Tailscale Serve, so the agents panel can say "1 user" or "2 users" (#23). The panel side is omacom/omarchy#18467.
• Pair and disconnect one model at a time, keeping every connection when pairing again or disconnecting from a whole host (#22)
• Point to Setup > Security > SSHD when pairing can't connect to the other machine (#21)

v0.6.0

• Share models across the tailnet, with a network survey of what other machines share and whether this one can share (#7). The agents panel shows and controls them with omacom/omarchy#18296.
• Run a model across two DGX Sparks cabled through their ConnectX-7 ports, with GLM-5.3 Flash as the first paired recipe (#8)
• Notice a newly cabled DGX Spark and offer to set it up (#10)
• Report a paired model's memory across both machines (#11)
• Fix the record_pair test on machines that aren't a 128 GB Spark (#12)
• Link model failures to their logs (#13)
• Give agents the server's reply cap instead of a fixed 16k, with the tensorfold recipes raised to 64k (#14)
• Let tensorfold models see up to 50 images per request (#16)

v0.5.0

Brings local models to 64 GB GB10-class machines and shows a setup's download in the agents panel. Nemotron 3.5 Bolt is the recommendation for 64 GB: one stream at the full 262K window, 91 tok/s on short prompts and 82 tok/s at 32K, peaking under 40 GiB, with Qwen3.8 27B (131K window) after it. Its Bolt image now counts reclaimable memory as free, so a unified-memory machine no longer refuses requests past 4,096 tokens once the page cache fills. The N1x shares the GB10's class, so only memory decides what runs; a variant can carry its own server arguments sized to its memory; and use --detach runs a setup in the background, reporting its download to the panel as an installing row.

v0.4.0

Adds Nemotron 3.5 Bolt on TensorFold, the recommended model for NVIDIA N1x laptops with 128 GB: 86 tok/s on short prompts and 69 tok/s at 32K, loaded in about 30 seconds in 32 GiB, and it takes images. Recipes can hand a server a checkpoint directory as a bare argument, and measurements record the NVIDIA driver, an ARM CPU and the system, with one pin per artifact instead of every file.

v0.3.2

Detects the NVIDIA N1x laptop GPU (RTX Spark N1X) as its own unified-memory family, so models set up for nvidia-n1x fit and start again.

omarchy-local-ai 0.3.0

Both recipes on Gufo 0.28.1: Flash Next moves from halogen (47.5 vs 28-41 tok/s at 32K, 105 vs 50 on code edits); the 27B moves to 4-bit (20.7 vs 13.7 tok/s, 40 instead of 55 GiB). tools/bench also measures a code edit.

omarchy-local-ai 0.2.1

A model's first start after a setup saved while memory was busy now records its memory, so the agents panel shows it.