Replies: 2 comments 3 replies
-
|
Questions on AeroAgent:
I ask because this is a new topic for me. |
Beta Was this translation helpful? Give feedback.
-
|
Great questions, and a perfect excuse for a proper TL;DR. I will keep this out of the formal docs and just answer your four points directly. TL;DR. AeroAgent is the part of AeroFTP that turns plain language into real file actions, locally and on every server AeroFTP can reach. The "AI provider" you connect is only the brain. Whether that brain is online or offline is a separate axis from what AeroAgent does with it. With a local brain (Ollama) and local files, the whole loop runs offline with no account and no data leaving the machine. One thing to clear up first: two separate axesPeople keep conflating these, so it is worth stating plainly.
So online vs offline is a property of the brain, not of the model name. DeepSeek is the source of confusion here: DeepSeek-R1 is an open-weight model you absolutely can run offline, but only by pulling it into Ollama. The DeepSeek entry in the provider list is their cloud API, which is online. Same weights, two different doors. 1. Online vs offline column: yes, good ideaI will add a "Runs offline?" column to the guide above. In practice it is a single "yes" on the Ollama row (and Custom, if you point it at a local server), and "no" everywhere else, for the reason above. I will phrase it as offline-capable brain rather than per-model, so it does not mislead. 2. Why AeroAgent instead of a website or a separate offline appYou guessed the main one correctly: it operates on files directly. The difference from pasting a script out of a chat website is that AeroAgent has hands, not just a mouth.
Two properties matter more than the list itself:
And you are right that for a one-off "generate a script I will reuse forever" task, a plain chat website is perfectly fine. AeroAgent earns its place on the open-ended, stateful, multi-step jobs ("reconcile these two servers, then archive what only exists on the old one"), not on one-shot snippet generation. 3. What I personally use it for, honestlyAeroAgent was a deliberate component from day one. When I designed it, AI tools for managing files were not yet common, and for an app born in 2025 an integrated agent felt mandatory, especially since no comparable client had one. It started with the first few tools and grew through successive expansion phases. Today it covers the whole table above, and the goal is to make it fully capable for any task involving managing, editing, or creating files. The latest update is exactly that direction: it begins giving AeroAgent advanced coding functions. That phase just started and I will carry it forward in parallel with everything else. Do I use AeroAgent to develop AeroFTP itself? Honestly, no. For coding I use dedicated platforms, and today AeroAgent is not the best surface for that, even though it is true that a terminal is now enough to drive a coding agent. Could it already be used to develop AeroFTP or other apps today? Yes, but with real limitations compared to platforms built specifically for coding. The gap is not "they have more tools" in general, it is a few specific dedicated capabilities AeroAgent does not have yet: an MCP client (today AeroAgent is an MCP server, so it exposes its tools to Claude Code or Cursor, but it cannot yet consume external tool servers), parallel subagents, and a mature plan mode. The recent coding block already brought the foundations (checkpoints, patch application, git, diagnostics, @-mentions, rules files), so the direction is set, the depth is what is still growing. So for me the honest split is: AeroAgent for file and server management where having hands on the actual files is the point, and a dedicated coding platform for building AeroFTP, until AeroAgent's coding side matures. 4. Ready-made scripts for common tasks: I strongly agree, and a lot already exists without AIThis is exactly the right instinct, and it is already partly how the app is built. Deterministic, repeatable tasks should not need an LLM. AI is for the open-ended, "I am not sure what I want yet" cases. What already runs with zero AI:
What does not exist yet, and where your list is a genuinely useful roadmap:
Now the interesting part for "ready-made scripts": AeroFTP already has a plugin system that is precisely a manifest plus a script (bash, Python, anything), SHA-256 verified and sandboxed, with three reference plugins today (File Hasher, CSV Stats, Image Info). It is the natural home for exactly the tasks you describe: a jxl-convert plugin or a fuzzy-dedupe plugin would be a script, not an AI prompt. The honest status is that the plugin marketplace is currently gated (the signed-registry client is not finished) and CLI plugin support is still pending, so today plugins are more a reference mechanism than a shipped ecosystem. Your comment is a good push to either promote a couple of these to native AeroFile features (the way byte-identical dedupe already is) or to finally open the plugin lane for them. Thanks for asking, this is a genuinely useful framing and some of it is going straight into the tracker. |
Beta Was this translation helpful? Give feedback.
-
|
I'm not sure where it's best to discuss this, but I think I prefer the term 'non-identical duplicates' to 'near-duplicates'. |
Beta Was this translation helpful? Give feedback.
-
|
Both adopted, thanks. I have taken "non-identical duplicates" as the name, it is more precise than "near-duplicates", and your modality point too: the plan now routes by file type rather than by one hash, perceptual hashing (pHash) for raster images, MinHash and SimHash for text including I have written the full detail into the tracker so this thread stays readable: #421 item 12. Still an idea, not built, but captured correctly. |
Beta Was this translation helpful? Give feedback.
-
|
A reader (A.K.) asked me by email, and it is worth answering in the open:
Short answer: the model does not execute anything, and it does not get to decide whether its own call is safe. The decision to call a tool and the execution of that tool are two different layers, and only the second one can move bytes. Here is the path every tool call travels. flowchart TD
A["You ask AeroAgent to do something"] --> B["Model proposes a tool call, just structured data, no I O yet"]
B --> C{"Host dispatcher reads the tool fixed danger tier"}
C -->|Read only| D["Runs automatically"]
C -->|Write| E["Approval gate"]
C -->|Destructive| E
E --> F{"You see the resolved arguments, the real path, destination and server"}
F -->|Approve| G["Executor performs the I O"]
F -->|Reject| H["Blocked, nothing runs"]
G --> I["Vault injects the credential only here, the model never sees it"]
The model can only propose. When AeroAgent chooses a tool it emits a structured tool call, which is just data, and that proposal never touches the filesystem or the network. A single host side executor is the only code that performs real work, and it runs outside the model entirely. Every tool has a fixed danger tier baked into the code, not chosen by the model at call time, so the model cannot relabel its own upload as harmless to slip it through.
That tier drives approval, and this is the actual boundary. The confirmation shows you the resolved arguments, the concrete destination, path and server profile, not a paraphrase from the model. So a hallucinated but valid looking argument is precisely what the prompt exposes, because you approve the real resolved target rather than the model's intent. The model never holds your credentials. It references a saved server by id, and the encrypted vault resolves and injects the secret only inside the connection, at execution time, so even a fully hallucinated tool call cannot carry a secret out, because the secret was never in the call. Paths are validated regardless of intent. Traversal, control characters, argument injection and a deny list of sensitive locations are refused before any write, so a bad path is rejected whether or not a human approved the action. If you have used agent mode in Cursor, in the VS Code Copilot agent, or in Claude Code, this shape will feel familiar, and that is on purpose. Those tools also treat a model tool call as a proposal and gate the risky ones behind a human accept. AeroAgent applies the same idea, tuned for file transfer and remote storage instead of editing a repository.
The difference is domain and blast radius, not philosophy. Where a coding agent asks before it runs a shell command or writes to your working tree, AeroAgent asks before it uploads, overwrites or deletes on a remote, it classifies each of those by a fixed tier, and it shows you the resolved destination and server before you approve. For unattended automation, where by design there is no human to click a dialog, the interactive prompt is replaced by deterministic guardrails, path allow and deny lists, rate limits and an audit trail, always opt in and never something the model can switch on by itself. So what enforces the boundary is not the model's judgment. It is a tier the model cannot alter, an approval gate that shows you the real resolved target, a credential layer the model cannot read, and path validation that runs no matter what. Thanks A.K. for a question worth publishing. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
AeroAgent is the AI assistant built into AeroFTP. You talk to it in plain language and it manages local and remote files, edits code, runs server operations, and more.
A lot of people do not realize they do not need a paid OpenAI or Anthropic key to use it. AeroAgent speaks to 25 providers, and several of them offer a genuinely useful free tier, some with no credit card at all. This guide shows the best free options, the models you can actually select on each, and which one I would pick depending on what you want to do.
AeroFTP itself is open source (GPL-3.0), free, and ships with no tracking.
1. Open AeroAgent
In the GUI, open the AeroAgent panel from the AeroTools menu.
Inside the code editor you can also select a block of code and press
Ctrl+Shift+A(Cmd+Shift+Aon macOS) to send it straight to AeroAgent as "Ask AeroAgent".From the command line (the binary is
aeroftp-cli, also callable asaeroftp):2. Add a provider (this is the only required step)
Open Settings, AI, Providers, or use the Provider Marketplace. You only need to configure one provider to start. Paste an API key, add a model name, and you are ready.
3. Best FREE options to test AeroAgent (mid-2026)
Legend for the labels used below:
★ marks the options I recommend starting with. The Card? column is the important one: every provider below gives real value with no credit card. Watch out for the opposite case, the paid flagships (OpenAI, Anthropic, Google Vertex AI) work like an AWS or S3 free tier, they ask for a card up front and bill you the moment you go over. None of those are in this table.
Online vs offline note. "Offline" here is a property of the provider (the brain), not of a model name. Only Ollama runs fully offline (and a "Custom" endpoint pointed at a local OpenAI-compatible server like LM Studio or llama.cpp). Every other provider below is a cloud API. DeepSeek is the usual source of confusion: DeepSeek-R1 is an open-weight model you can run offline, but only by pulling it into Ollama; the cloud DeepSeek API is online. Whatever the brain, acting on a remote server still needs network to that server; acting on local files with a local brain is fully offline.
:freemodels (optional top-up to raise limits):free, rotating (DeepSeek V3, Llama 4 Scout, Gemma 3, Mistral Small, Qwen, and more)kilo-auto/free, 256K context, rotates free models), Cohere North Mini Code, NVIDIA Nemotron 3, Poolside, StepFun, plus paid Anthropic/OpenAI/MistralHow the OpenRouter free models really work
OpenRouter is worth a dedicated note because of a common misunderstanding. The models tagged
:freeare reachable even with a zero balance, but at a low quota of 50 requests per day. The trick: if you add credit just once (the threshold is $10, not $5), your free-model quota jumps to 1000 requests per day, and it stays there permanently even if your balance later drops back toward zero. The same balance also unlocks the paid models, charged per use. So topping up $10 once is the single best move to get serious daily mileage out of the free models.A note on Qwen and on Together AI
Qwen models are excellent and AeroAgent supports the Qwen (Alibaba) provider, but the free path changed in 2026. Alibaba closed the free OAuth quota on 15 April 2026, so Qwen now runs on trial credits and then pay per token. The good news: you can still use Qwen3 for free through Cerebras, Groq, and OpenRouter (see the table above), so you do not lose access to the family.
Together AI is also supported, but it is no longer a permanent free tier. New accounts get a small signup credit (about $5), which is fine for a quick try but runs out.
4. Which model should I choose?
qwen2.5-coderorllama3.2If you just want to start in five minutes with zero setup and zero cost: install Ollama for a fully local agent, or grab a free Groq or Cerebras key for blazing-fast cloud replies.
5. Example prompts
AeroAgent can:
6. Safety controls
AeroAgent has three approval levels for tools:
On the command line the same control is
--auto-approve safe|medium|high|all, and--plan-onlylets you preview the steps without running them.Resources
aeroftp-cli --helpIf this helped, a star on the repository is appreciated. Questions are welcome in this thread, and you can always ask AeroAgent itself: "How do I configure you to use Groq?"
Beta Was this translation helpful? Give feedback.
All reactions