Mercury CLI

Notes

TensorFold in Mercury CLI and 1.0.6

· by @dubsthedev · 1.0.6

Mercury CLI 1.0.6 makes TensorFold a first-class local model source in Mercury. Before getting into the controls, I want to shout out @ashxhart (ashhart on GitHub), who made TensorFold. I am very grateful for the work he is doing in the local-model community. I build Mercury CLI on my own as one person, and his work aligns closely with what I value in Mercury. I think he is pushing the local community in the right direction, and seeing this happen is something I look forward to building further on, especially because TensorFold is open source under the Apache License 2.0, which I really appreciate now that Mercury CLI is open source under Apache 2.0 as well.

Running models with /tensorfold and /model

/tensorfold and the TENSORFOLD and TENSORFOLD HOST groups in /model (Providers and sign-in, Models and effort) set up and run the TensorFold 1.2.2 engine (or Mercury's llama.cpp engine on Windows until TensorFold ships for Windows) across four places:

  • On a Mac or Linux GPU: On Apple silicon (Metal) and Linux NVIDIA GPUs (CUDA), Mercury installs TensorFold through Homebrew on macOS (brew install ashhart/tensorfold/tensorfold) or the SHA-256-verified release archive on Linux, and offers TensorFold first on first-run sign-in and in /model on a fresh install of a capable computer. It pulls and serves the qualified 256k starting points (Nemotron Lightning, ordered first on Macs under 64 GiB of memory, including 48 GB Macs, and on pre-M5 Apple silicon; Qwen3.8 27B with its DFlash2 speculative drafter, ordered first on 64 GiB+ M5 Macs and Linux CUDA; and Qwen3.5 2B for 8 to 16 GB Macs) or any supported MLX, NVFP4 or EXL3 checkpoint on disk or from Hugging Face. You can step the context window across 32k, 64k, 128k and 256k against usable memory, copy the macOS GPU wired-memory sysctl command with w when needed, copy the one-line support string with c, and read per-turn decode, time to first token, prefill, cache hit rate, reasoning tokens and draft acceptance in /usage.
  • On a Mac or Linux server on your network: Press n on a Mercury-started local server to share it on your local network with a scoped mercury[inference] key and a 6-digit one-time pairing code, or connect from another computer by host:port under TENSORFOLD HOST.
  • On a rented GPU: Under TENSORFOLD HOST, rent a GPU pod rents a Runpod or Vast.ai CUDA pod through /llmhost that verifies the TensorFold archive by SHA-256 and Ed25519 signature, pulls and serves over an SSH tunnel, shows live spend on its card, requests a stop with s, and terminates the pod and its disk with t.
  • On Windows x64: Until TensorFold ships a native Windows binary, /tensorfold and /model serve Qwen3.5 4B, Qwen3.5 2B and Nemotron Lightning on CPU through Mercury's own llama.cpp engine (dist/tensorfold-adapter.mjs), downloadable in place on Windows x64 with SHA-256 verification and translating reasoning_effort and thinking_budget.

Local sessions and rented TensorFold pods also use a new local prompt profile (MERCURY_LOCAL_PROMPT_PROFILE, on by default) that keeps a compact, cacheable system prompt prefix and sends the core tools in full while deferring the rest behind ToolSearch. In /config, Helper model and Helper model for compaction (both off by default) can route background jobs and compaction to 2B on TensorFold or Mercury's engine, 4B on llama.cpp, or the session's local model, and /health (Health check) adds the TensorFold, support and Helper model rows.

Also in 1.0.6

The Session Concourse coordinator is now off by default (mercury --coordinator-on turns it on), leaving a two-pane concourse with the selected session's live view on the left and the session list on the right (Sessions). /goal joins /mission as a second name for a standing session goal checked at the end of each turn (Slash commands). In /llmhost, a private recipes file now joins the shipped catalogue with your private rows ordered first instead of replacing the shipped list. The release also fixes OpenAI HTTP 4xx diagnostics and resume recovery, crew-board rows reappearing after compaction, dropped large @-mentioned image notices, ApolloReview's model-facing schema, and four tool descriptions. The full list of changes is in the 1.0.6 release notes.

Please give TensorFold and /tensorfold in Mercury CLI a try and send feedback. I am happy to see what else I can build around it from here.