Pinvou Agent ships preview builds, a private-runtimes directory and a home folder named for an older number
Open-source desktop AI agent for tools, files, knowledge, workflows, and real deliverables.
At a glance
- What is it?
- A Rust and Tauri desktop agent workspace organised around three modes of work, with a design surface, a coding agent brought in over a protocol, and a connector store. The readme is candid about what leaves your machine, and the packaging has three separate version numbers that do not agree.
- Who is it for?
- Pinvou Agent suits someone who wants an agent that ends a task with a file rather than a paragraph, and who wants a local model to be a first class option rather than an afterthought. Three things to know before you rely on it.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Rust, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- Pinvou/pinvou-agent3,032 ★ · +1,198
| Repository | 2026-09-10 | 2026-11-15 |
|---|---|---|
| Pinvou/pinvou-agent | 1.8k | 3k |
1,198 Stars in the last 66 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The last three releases are all labelled preview
Every recent tag carries the same suffix in its title, translated as a preview version: 0.10.4, then 0.10.5 a week later, then 0.10.6 two days after that. The download link in the readme header is labelled as a preview download and points at the releases page. So the published artefacts are previews by the project's own account, not a stable channel, and they are moving fast enough that a week is a full patch cycle. The version line is still below 1.0, which is consistent with a desktop shell of this size, but the combination means there is no release a cautious user can point at and say this is the one. The last recorded commit on the main branch is dated 2026-10-02, the day before this was written, so the branch is moving at the same pace as the tags.
The data folder, the application folder and the release line carry three different numbers
The readme says sessions, settings, knowledge and runtime extensions all live under one directory in the user's home folder. The application's own directory in the repository is named for that same number with a suffix. The releases are on a completely different minor line. So a user reporting a bug, a maintainer reading a log path and someone pinning a release are all looking at different numbers, and none of them is wrong on its own. Two more version sources sit at the root: a version file, and the application package manifest, which one of the badges on the readme points at directly so that the header shows the app's own version rather than the repository's. That is a reasonable thing to do for a monorepo, and it means the single number in the badge and the single number in the release tag are not necessarily the same number.
Pointing it at a local server means using variables named after a hosted provider
The readme's local model example is instructive because of what it does not call things.
export DEEPSEEK_BASE_URL="http://127.0.0.1:8000/v1"
export DEEPSEEK_API_KEY="local-no-auth"
export DEEPSEEK_MODEL="your-model-name"The endpoint is a local server on the loopback interface, and the key is a literal placeholder meaning there is no authentication. Both are named after one specific hosted provider, which is the configuration the application was built around, so a user running something else locally inherits a variable name that describes a service they are not using. The page does say that the same values can be managed inside the application's settings instead, and that a template exists for a local server, so the variables are one route among several. The provider list the templates cover is long, spanning the major hosted services plus a local runtime.
Plain HTTP off the loopback interface is refused unless you ask for it
One sentence in the model access section is a deliberate safety default. For a plain HTTP endpoint that is not on the loopback interface, sitting on a network the developer trusts, you have to set a specific variable to a specific value to allow it. So the failure mode of pointing the agent at a colleague's development server is a refusal rather than a silent plaintext connection, and the opt-in is explicit and documented rather than buried. That is a small, well-handled thing and it is worth noting because most of the rest of the readme is feature description: this is the one place where the page is about a risk rather than a capability, and the accompanying privacy note takes the same tone. Whether data leaves the machine depends on which model and which tools are enabled, and the page is direct that cloud models, remote tool servers and third party connectors send requests to their respective services.
The auxiliary chat is described as enforced by the server, not by the model
One feature entry carries a parenthetical that does the real work. Alongside a main task, you can ask a side question in a per-task side session, and the entry states that this session is zero-tool and isolated from the task's execution context, and that both properties are enforced by the server. That distinction matters. If the isolation were a prompt instruction, an agent under pressure would eventually ignore it, and a side channel with no tool access is a much smaller risk than one that can read the same files. Stating the enforcement point says the property is architectural. It is also the kind of claim worth verifying in the code rather than taking from a bullet list, because the whole safety argument rests on the side session genuinely being unable to call tools rather than merely being told not to.
Linux gets two architecture builds, macOS gets one universal build
The platform section is specific enough to be useful. The source builds on all three desktop systems. Linux release packages target Ubuntu 22.04 or newer, which means a specific glibc floor, on two architectures, and the package additionally needs a minimum version of the web toolkit that ships with the desktop runtime, so a fresh install that has not applied the usual updates will not run it. macOS packages are single universal binaries covering both Apple Silicon and Intel, for a much older system version. So the two architectures that matter for Linux get separate artefacts while macOS gets one. The same section notes that speech recognition engines can be packaged per build configuration, which means the voice input feature is not present in every artefact, and that document parsing depends on optional external tools. The sentence describing those tools is cut off on the page.
A public MIT repository holding a directory called private runtimes
The top level is wider than an application shell needs to be. Alongside the application directory there is a command line package, a knowledge package, a relay for the phone remote control feature, a scripts directory, a third party documentation directory, and a directory whose name is a private runtimes folder. There is also a submodule entry and a gitmodules file, so at least one component lives in another repository. The licence is MIT and there are separate trademark and third party notice files, which is what a project shipping bundled third party material should carry. Contributions are handled through a certificate of origin rather than a contributor licence agreement, which means sign-off rather than a copyright assignment. A secret scanner is configured, and three readmes in three languages are maintained side by side. For a project inviting extensions, the interesting question is what the private runtimes directory actually contains, and the repository does not say.
Editorial conclusion
Pinvou Agent suits someone who wants an agent that ends a task with a file rather than a paragraph, and who wants a local model to be a first class option rather than an afterthought. Three things to know before you rely on it. The releases the project publishes are labelled previews by its own account, there is no in-application update check, and upgrading is a manual download, so pin a version deliberately. Its privacy story is honest and conditional: a local model with local tools stays on the machine, and the moment you add a cloud model, a remote tool server or a third party connector, requests leave it. And the repository ships a directory named for private runtimes under an MIT licence without saying what is in it, which is the sort of question to ask before building on the extension story.
Frequently asked questions
What is Pinvou Agent?
An open source desktop agent workspace written in Rust on a desktop shell framework, organised around three modes of work: giving the agent a task that ends in a deliverable, turning a prompt into an editable visual design, and bringing an external coding agent into a real project. It runs on a local model or any OpenAI compatible endpoint and extends through tool servers, command line connectors, skills and workflows.
Does Pinvou Agent send my data to a server?
It depends on what you enable, and the readme says so directly. A local model with local tools stays on the machine. Cloud models, remote tool servers and third party connectors send the relevant requests to their respective services. The ready-made connectors cover chat platforms, a note application, meeting software and enterprise knowledge bases.
How do I point Pinvou Agent at a local model?
It works with a local serving runtime and any OpenAI compatible endpoint. The readme's example uses environment variables named after one hosted provider, with a literal placeholder for the key because a local server has no authentication, and says the same values can be set in the application settings instead. A plain HTTP endpoint that is not on the loopback interface additionally needs an explicit opt-in variable.
What does ACP mean in Pinvou Agent?
It is the protocol through which an external coding agent is driven from inside the desktop workspace. The agent can read and edit either a real project or an isolated temporary workspace, run commands, and surface plans, tool steps, permission requests and file changes, with sessions bound to their workspace and resumable after the application restarts.
Where does Pinvou Agent keep its data, and how does it update?
Sessions, settings, knowledge and runtime extensions all live under a single directory in the user's home folder. Updates come from the project's release page, and the readme states plainly that in-application update checks are not enabled yet, so upgrading is a manual download.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/pinvou-pinvou-agent)Star history chart for your README
[](https://hysenlabs.com/projects/pinvou-pinvou-agent)Was this useful?