There are a lot of coding agents now. Good. More people are building tools around the way they actually work. Every new approach gives the rest of us something to learn from, and the people using these tools get more choices.
Cloaky is our contribution: a desktop workbench where you can choose a coding engine, give it a task, follow the work, review the changes, and try what it built beside the conversation.
We’re preparing Cloaky Base v0.1.1 for macOS and Linux. It connects five coding engines to Venice with your own API key, and voice transcription runs on your machine. Pro, which remains a private build, adds local language models and direct provider options.
That is where we’re starting.
The bigger idea is personal: more of the intelligence you use should belong to you.
I don't only mean the model. I mean the examples you choose, the context you build, the ways you solve problems, and the routines you return to because they work. Right now, most of that lives inside other people's products.
I want people to build intelligence around their own lives and work, and keep more of it running on hardware and tools they control. Renting additional capability should be a choice you can make when it helps.
Cloaky starts with the workbench, because the workbench is the part you keep while the models change.
A place to stay with the work
When an agent changes a project, a lot happens between your request and its final message.
Files change. Commands run. An approach works, or it doesn’t. Sometimes you need to interrupt, ask a question, or point out that a perfectly reasonable-looking result has missed what you wanted.
Cloaky brings the conversation, file review, and preview into one place.
You can read a diff, comment on a line, reject a hunk, or return to a file checkpoint after a supported turn. Depending on the engine and permission mode, approval cards let you inspect requests to edit files or run commands and decide whether to allow them.
Those controls vary by engine. An approval is a review step, not a sandbox: once you approve an action, nothing contains it. Container isolation and the command sandbox, both off by default, are what limit what an engine can reach.
The experience I want is simple. You should be able to stay with the work as it happens.
See what changed. Give direction. Try the result. Keep the parts you want and work through the parts you don’t.
Whether you’re fixing something small or finally starting a project you’ve had parked for months, the workbench should help you turn an intention into something you can inspect and use.
Keep the workbench. Choose the intelligence.
An engine and a model do different jobs.
The engine runs the coding session and its tools. The model supplies the inference behind it. The workbench is where you give direction and review the result.
Cloaky Base works with Claude Code, OpenCode, Codex, Kimi Code, and Grok Build. Their model requests go through Venice, using your own Venice key and credit.
Base doesn’t run language models on your machine. Each request carries the context the task needs, including prompts, files the agent reads, and tool results. Tools and integrations can also make their own network requests. Cloaky’s model picker shows theVenice privacy tier
Figure 2. Base's model-request route. Hosted requests carry the task context they need. Tools and integrations can also make their own network requests.
What stays home matters too. Cloaky sends no telemetry, analytics, or crash reports, and it runs no server of its own. Your settings, sessions, and records live in a folder on your computer, outside the app.
Knowing where the intelligence runs is part of having a meaningful choice. Pro extends that choice to direct providers and to local language models through Ollama and LM Studio.
Over time, I want people to have more freedom to decide what they run locally, what they rent, and how they combine the two around their own work.
A useful workbench should keep earning its place as those choices change.
A small piece of ownership
Voice typing makes this idea tangible.
With whisper.cpp, Cloaky transcribes your recording on your machine. The text appears in the composer for you to read or change before sending.
Recording, transcription, and sending are separate steps. The audio stays local. When you send the resulting prompt in Base, that text follows the Venice route.
Figure 3. Recording and transcription stay local. You decide when to send the resulting text.
You can choose among four English speech models. You need whisper.cpp and a downloaded model, and transcription has no cloud fallback.
It is a small capability, but a useful one. You can speak a task, review the words, and decide when they become a request.
That is how I want to build toward the larger vision, through things people can understand and control.
More of your context staying useful to you. More of your workflows working the way you prefer. More capability running on machines you own.
There is plenty left to build. That is also what makes this interesting.
A second set of eyes
Cloakster is an optional second opinion on the session. Its session monitoring is off by default.
It looks for things such as repeated failed approaches, drift from the request, and claims the output doesn’t support. It can also flag pending actions that appear outside the task’s scope.
Cloakster uses Jev, built by TypeSafe AI, through Venice’s beta decision API. Cloaky uses it to ask questions such as whether work is complete or a proposed action goes beyond the task.
We tested it on labeled synthetic examples while building it.
On September 24, 2026, Venice’s jev-latest returned answers for 103 of 104 scenarios; one request hit a rate limit. The recorded median decision-request latency was 421 millisec
Figure 4. Results at the selected thresholds. Blue segments are detected positive examples; sand segments are misses. Thresholds were chosen using the same dataset.
At the thresholds selected for that set, Cloakster detected 17 of 18 completion examples and 22 of 23 actions outside the task’s scope. The two misses were one finished task without a check and one proposed file write in another project.
We chose those thresholds using the same examples. These are calibration results, and we still need to see how Cloakster behaves in real sessions.
When session monitoring is on for an eligible session, selected excerpts go to Venice’s decision model. Each Cloakster check can be turned off in Settings.
The purpose is to give you another perspective while you stay in charge of the work.
Cloakster is an optional second opinion on the session. Its session monitoring is off by default.
It looks for things such as repeated failed approaches, drift from the request, and claims the output doesn’t support. It can also flag pending actions that appear outside the task’s scope.
Cloakster uses Jev, built by TypeSafe AI, through Venice’s beta decision API. Cloaky uses it to ask questions such as whether work is complete or a proposed action goes beyond the task.
We tested it on labeled synthetic examples while building it.
On September 24, 2026, Venice’s jev-latest returned answers for 103 of 104 scenarios; one request hit a rate limit. The recorded median decision-request latency was 421 milliseconds.Building something worth coming back to
The public beta has closed. Cloaky Base v0.1.1 is coming next, free for macOS 13 or later and Linux. Its source opens under the MIT license at v0.1.10, after ten releases.
Apple Silicon has received the most testing. Intel Mac and Linux feedback will be especially useful.
Figure 5. Two editions, with different model options. Local voice transcription is available in Base; local language-model inference is a Pro capability.
The engines are separate programs. Cloaky can install Claude Code, OpenCode, Codex, and Kimi Code, or use an existing installation. Grok Build uses a guided terminal install.
We’ve used an earlier Linux build for timers, notes, boards, and other small projects, and watched how it handled real coding sessions.
In one slugify task using GLM-5.3-Flash through OpenCode Go, Cloaky took 13 seconds and the OpenCode CLI took 14. Both passed the same four tests. One run tells us very little about relative speed, but it was good to see the workbench carry a task from a request to a checked result.
A harder task exposed more to investigate. OpenCode Go through Cloaky wrote files, as did the standalone Grok Build CLI. Grok Build through Cloaky’s ACP connection produced none in that run. That earlier build had other problems too. Opening a second instance hit a profile lock and showed a black window, and provider-balance errors needed a clearer explanation.
I want to keep showing what happens in practice, including attempts that fail.
Earning a place in someone’s daily work takes attention to those details: getting started, understanding an error, reviewing a change, and coming back tomorrow.
That is the work we need to keep doing.
There is room in the harbor
This space is crowded with people doing work I respect. I want them to succeed.
OpenCode is the standard we keep learning from. Cloaky gets to build on a world these teams helped create.
A personal thank-you to Theo, the T3 Code team, and their community. I’ve learned a lot from Theo’s channel and from studying T3 Code.
Cloaky is not a T3 Code fork. Probably should’ve been. Would’ve made life a lot easier.
I have real respect for the care they put into their work and for sharing it so openly.
Thank you to OpenCode, Cline, Aider, Zed and the ACP community, Ollama, LM Studio, and Venice. Cloaky ships with notices crediting the maintainers whose work it includes.
And credit to Diogo Almeida and the TypeSafe AI team for Jev. Their work gives us another building block; the Cloakster calibration measurements above are our own. We’ll update these numbers as we learn.
We all benefit when people have better tools and more freedom to build with them. There is room for different approaches, different workflows, and different ideas about where this can go.
For Cloaky, the ambition reaches beyond the coding session. I want it to help people build intelligence around their own work and keep more of that intelligence in their own hands.
If that sounds like a direction you want to help shape, welcome.
Leave your email at cloaky.dev and you'll get one email on release day, in the next few days, no newsletter. Follow @cloakyapp for practical demonstrations, development updates, and the next measurements.
Pro remains a private build. If you would like to help shape it, contact hello@cloaky.dev.
Another boat in the harbor. Let’s see how she does.
GIF
