This week MeshHold learned to compute, not just store. It can now run AI models on your own hardware, and hand the heavy jobs — transcoding, transcription, captioning, indexing — to whichever of your machines is up to the task. Private AI, on hardware you own, with the data never leaving it.
AI on your own hardware
You can now run local models inside MeshHold. Downloading a few gigabytes of weights is a managed job with a progress bar rather than a goroutine that might die halfway; there's a registry of what your build can actually run, and a runtime manager that decides which model gets the graphics card when there's one to share.
The part I like most: your AI agent can answer from this node's own model. Add a Claude or OpenCode instance and point it at the machine's local model instead of a cloud provider — "local models only" means the text you type stays on the machine, full stop, with no sign-in and no account. And if the node in front of you has no GPU and no weights, it can borrow a model from a peer that does, over the mesh. The weights live where the hardware is; the answer comes to you.
A mesh that does the work
Underneath that sits a proper jobs system, and it's the more general win. Heavy work — transcoding a video, pulling subtitles out of an audio track with Whisper, captioning a photo, computing embeddings, OCR-ing a scanned PDF — is now a job rather than something the daemon does on the calling thread and hopes finishes. A scheduler auctions each job to whichever node can actually do it and has the spare capacity, background work politely stands aside for whatever you're using the machine for, and — my favourite shape — your phone can ask your desktop to do the lifting and just collect the result. Your mesh stops being only a place to keep things and becomes a small, private compute cluster.
Search inside your files
The first thing the jobs system bought, almost for free: real full-text search inside your documents. Not just filenames — the words in a PDF, a Word doc, or a photograph of a page run through OCR, indexed on the node that holds them. And a related one, filed the same way: antivirus scanning as provenance. A file can carry a scan verdict as an overlay next to its other metadata, so you can see what looked at it and what it said — information, not a gatekeeper.
The office kept maturing
Not everything this week was new. The office suite that landed last week spent this one growing up: charts you can actually make and shape (scatter, lines that curve or step), a real formula library (dates, the time-value-of-money functions a spreadsheet is expected to have), and a pile of small fixes from the unglamorous act of actually using it — a ruler that counts in millimetres, right-click menus that offer the document's own verbs, and a Print button that prints your document instead of a picture of the editor.
And a small delight I couldn't resist: the boot animation now draws MeshHold's logo mark from your actual mesh — its corners are your real nodes, lighting up as they come online.
That's very likely the last of the big new surface before 0.8. From here, the job is making all of it solid.
Since last week
- Local AI models: run models on your own hardware — weights fetched as a managed job, a registry of what a build can run, and a runtime manager that arbitrates the GPU
- Agents on your own model: a Claude or OpenCode instance can answer from this node's local model, with a "local models only" mode that keeps your text on the machine — no sign-in, no cloud; a node with no weights can borrow a peer's model over the mesh
- A distributed jobs system: transcode, Whisper transcription, photo captioning, embeddings and OCR run as jobs auctioned to whatever node can do them, with load-aware scheduling that yields to interactive work and lets a phone offload to a desktop
- Search inside documents: full-text search across PDFs, Word docs and OCR'd scans, indexed per node
- Antivirus as provenance: an ICAP / clamd scan verdict is stored as an overlay beside a file's other metadata — informational, not a gate
- Office suite maturing: spreadsheet charts (scatter / line / curve), a fuller formula library (dates, financial), and ten usability fixes from real use — a millimetre ruler, right-click verbs, and printing that prints the document
- A living boot mark: the startup animation renders the logo from your real mesh topology, corners lighting up as nodes come online
- Windows networking robustness: stale VPN sessions no longer steal the live one's routes, switching networks no longer leaves the host with no DNS, and one PowerShell handles every route instead of one per route