“Local-first” is used for three different promises, and most comparisons collapse them into one. Your files and outputs can stay local. Your API keys and model traffic can stay out of the vendor's hands. The model weights themselves can run on your own hardware. A tool can do the first two and not the third — ours does exactly that. Here are eight tools with the three answered separately.
Does the tool read and write the folders on your own disk, and do the deliverables land there? This is the easiest of the three and nearly everything on this list passes it.
When you bring your own key, does the traffic go straight from your machine to your provider, or does it pass through the tool vendor first? Orkas does not proxy it: the key is sealed with AES-256-GCM in a local config and the call goes direct.
Can the model itself run on hardware you own — Ollama, llama.cpp, any local runtime — so that nothing leaves the machine at all? This is the strictest reading of local-first, and it is the one Orkas does not satisfy.
Telemetry, crash reporting and account behaviour are not scored here. They vary by build and by version, and a table that pretends otherwise would go stale quietly. Check each project's own current documentation before you rely on a number.
This list is published by Orkas, and Orkas is one of the tools on it. Every row that describes Orkas is marked. The four rules below were fixed before the list was written, and every tool here is measured against the same four.
All three promises kept
Local document question-and-answer that can run fully offline against a local model. If the requirement is that a set of documents never touches a network, this is the shortest path to it on the list.
It answers questions about documents. Turning those answers into a finished deliverable is a different job.
Read the full AnythingLLM comparison →An open-source local agent with a CLI and a desktop app, dozens of MCP extensions, and a wide provider list. Point it at Ollama and the weights stay on your hardware. It can also build a distribution with your own providers and branding.
One general-purpose agent throughout, and no media or office-document specialists.
Read the full Goose comparison →A self-hosted chat interface your whole team signs into, with wide model coverage including local runtimes. The only entry here built for several people at once rather than one person at a desk.
It is a chat surface. It does not hand you back finished files.
Read the full LibreChat comparison →An MIT-licensed terminal coding agent with dozens of providers, anything OpenAI-compatible, and local runtimes. A TUI first, with the desktop app newer. It separates a read-only plan agent from a full-access build agent, one key apart.
It is aimed at the repository. The half of a release that is not code is outside what a coding agent is for.
Read the full OpenCode comparison →A self-hosted gateway with channel clients, configured providers or local models, and a mature skill library that is strong at driving the machine itself. Background and nested sub-agent runs are built in.
You run the gateway, so there is infrastructure to keep alive.
Read the full OpenClaw comparison →An always-on autonomous agent you run on your own server, VPS or serverless infrastructure, for cron jobs and long-running headless work. Open source; bring your own model, including Nous Hermes.
No resume handle and no project-directory protocol, so a host has to replay history to continue a conversation. That is a limit only if you want a host in the loop.
Read the full Hermes Agent comparison →Files and keys local, model over the network
A desktop agent team on macOS and Windows. Your files, keys and outputs stay on your machine; with your own key the API traffic is never proxied through Orkas — the key is sealed with authenticated AES-256-GCM in a local config, excluded from cloud sync, and decrypted in memory only to call your provider directly. MIT licensed and auditable.
The third promise is the one it does not make: there is no local model runtime. Model calls go to a provider over the network — the managed official models, or your own key. If the weights have to stay on your hardware, pick one of the six above instead.
Read how that boundary works →Self-hosted workflow automation that moves data between systems on a trigger, on infrastructure you own. Included here because “local-first” is often really a question about where the pipeline runs, not where the model does.
It moves the data. It does not produce the deliverable a person reviews.
Read the full n8n comparison →| Tool | Files and outputs | Keys and model traffic | Model weights on your hardware |
|---|---|---|---|
| Orkas Ours | On your machine | Direct to your provider, never proxied | No — no local model runtime |
| Goose | On your machine | Your configured provider | Yes — via Ollama |
| AnythingLLM | On your machine | Your configured provider, or none at all | Yes — can run fully offline |
| LibreChat | On the host you run | Your configured provider | Yes — local runtimes supported |
| OpenCode | On your machine | Dozens of providers, anything OpenAI-compatible | Yes — local runtimes supported |
| OpenClaw | On the gateway you host | Configured providers | Yes — local models supported |
| Hermes Agent | On your own server or VPS | Your configured provider | Yes — bring your own model |
| n8n | On the host you run | Whatever each node is configured with | Depends on the node |
Read the last column as a requirement, not a score. Most people asking for a local-first agent want the first two columns — their documents and their keys under their own control — and are happy to call a hosted model. If you are in the minority who genuinely needs the weights on your own disk, that column is the only one that matters and it removes Orkas from consideration.
Orkas is free, MIT licensed, and open source, so none of the boundary above is a claim you have to take on faith — the key handling is in the repository you can read.
Checked on 11 September 2026, three independent runs the same day. The licence and liveness of all eight repositories came from the GitHub API; the Orkas column came from the shipping source at commit 54cab3031 — which is also where the “no local model runtime” limit in row seven comes from, rather than from a marketing decision to sound modest.
The third column — whether the weights can run on your own hardware — is cited from each project’s own documentation for the seven entries that are not ours. It is sourced, not measured, and it is the column most worth re-checking before you commit to a tool. Telemetry and account behaviour are out of scope entirely. One thing worth naming: n8n ships a sustainable-use licence, not a standard open-source one.
It is local-first for your files, your keys and your outputs, and it is not local for the model weights. Both halves of that sentence are on this page because collapsing them is how these lists usually mislead people. If your requirement is the second half, this page names six tools that meet it.
No. With your own keys, the key is encrypted with authenticated AES-256-GCM before it touches disk, written to a local config excluded from cloud sync, and decrypted in memory only to call your provider directly from your machine. The managed official models are the separate path, and they are opt-in.
AnythingLLM if the job is questions about documents, Goose if you want a general agent with Ollama behind it, OpenCode if the job is code. All three keep the weights on your hardware. Orkas is not the answer to that requirement and this page does not pretend otherwise.
Orkas is free and MIT licensed for macOS and Windows. Bring your own provider key and the model traffic never passes through us.
