npm run lint -- --fix
Executes commands with your permissions — read it before approving.
One workbench. Every route through Venice. Free, open source, yours.
Nothing routes anywhere you didn’t turn on — every session runs on Venice’s private tier, with no other provider account involved. Five engines in one workbench: Claude Code, OpenCode, Codex, Kimi Code and Grok Build.
Request Cloaky ProWhere your code goes PCloaky Base · macOS & Linux · free, open sourceCloaky Pro · private build · request access
running
What should we build in app?
Tests pass.Fixing the lint error next.
The retry helper is done — how far should the cleanup go?
Should the cleanup carry its own tests?
Pro
Cloaky Pro — local inference is a Pro feature. Base runs on Venice only. hello@cloaky.dev
Local inferenceSearch models…
Sessions on these models run on Ollama, on this machine — prompts never leave it.
ChipApple siliconApple silicon
Memory16 GBunified
Fits a model9.6 GBcomfortable · tight at 12.0 GB
LiveidleA local turn reports prompt and generate here.
Runtimes
checked just now
OllamaNo tool-capable modelsv0.32.5 · http://localhost:11434 · /Applications/Ollama.appOllama is serving and routed, but has no model that can call tools — an agent session needs one. Download one below.
LM StudioNot installedhttp://localhost:1234LM Studio publishes no release checksum, so Cloaky does not install it for you. Download it and this row picks it up.
Hybrid — a cloud session's subagents on a local model — unlocks once a runtime above serves a tool-capable model.
On this machine 1
llama3.212.4M pullsMeta’s small instruction models, quick on modest memory.1b1.2 GB3b1.9 GB✓Chat-only
Catalogue 3
qwen2.59.1M pullsBroad general family with sizes for every machine.0.5b0.4 GB1.5b0.9 GB3b1.8 GB7b4.4 GB14b8.4 GBtight+1 larger
gpt-oss5.2M pullsOpenAI’s open-weight models for agentic tasks.20b13.0 GBtoo big120b60.5 GBtoo bigqwen3-coder1.9M pullsAgentic coding in 100+ languages, tuned for tool use.30b17.7 GBtoo bigThe tests pass. Here is the final review.
The retry helper now retries network failures three times with exponential backoff and throws a typed ApiError. The implementation and its test remain reviewable in the diff pane; the lint issue is still open.
Edited
src/lib/api.tsRan
npm test -- --run apiRan
npm run lintfailedSession total: $0.42 · active time: 4m 12s · 18 turns
Add retry logic to the fetchUser helper and clean up the config.
Added the retry helper and updated three callers.
Now chase every fetch wrapper in src/lib.
Tests pass. Here’s where things stand:
- Retry helper — done — 3 retries, exponential backoff
- Lint — in progress — one unused-var error
Read
src/lib/api.tsRan
npm test -- --run apiRunning the suite twice: once scoped to the retry tests, then the full run with coverage so a regression outside api.ts cannot hide.
Bashnpm test -- api.test.ts
Bashnpm test -- --coverage
Default
engine configFollows the engine’s own config — no route is claimed until you pick one.
The standing pick until a model is chosen for this thread.
Effortfollows the engine
Today 8:00 AM
Run the suite and fix anything red.
Messaged Curie
Messaged Hopper & Curie Chat
2 messages with Curie
Suite is green again — one retry test asserted the old backoff cap; fixed in src/retry.ts.
Engines
Which agent runs a session. Claude Code is the built-in loop; the engines below replace it with another vendor's agent over ACP. Every one runs on this machine and sends its model calls to Venice — no other provider account involved. Engines are opt-in and ship disabled; pick per thread from the composer.
Prompts, project files the agent reads, and tool results go to Venice’s private tier — the route you see in Providers & privacy. The loop itself runs on this machine, and no other provider account is involved.
Prompts, project files the agent reads, and tool results go to Venice. OpenCode’s own loop runs on this machine; Cloaky points its model calls at Venice, so no OpenCode-side provider account is involved.
InstallRuns
curl -fsSL https://opencode.ai/install | bash, which fetches from opencode.ai. Cloaky types it into the terminal and stops there — read it before you press Return. Installers for these CLIs commonly edit your shell startup files, and not all of them verify a checksum.Prompts, project files the agent reads, and tool results go to Venice — no OpenAI account involved. This row runs the community ACP adapter @agentclientprotocol/codex-acp, because the Codex CLI has no ACP mode of its own — the version shown is the adapter’s.
CLI pathUsing /Users/demo/.local/bin/codex. Set a path only to override.auto-detect
ModelsFetched over ACP when this engine is enabled — Venice models only, the same Kimi K3, GLM 5.2, MiniMax M3 Preview and DeepSeek V4 Flash the picker pins. Detect re-runs the handshake.
What carries overStreaming and file tracking work. Approval cards appear only when the agent asks Cloaky before an action, and most do not: measured 2026-08-07, OpenCode reads and writes files entirely on its own, so no card appears and your permission mode does not reach it. Grok Build does route file access through Cloaky and does ask — but which tool an agent reaches for is its own decision, so treat that as behaviour observed today rather than a setting you control. Rewind, MCP toggles, verify-on-done, security scans and thread history are Claude Code engine features and stay off for these sessions.
Prompts, project files the agent reads, and tool results go to Venice — no Moonshot account involved. Kimi Code’s own loop runs on this machine.
Prompts, project files the agent reads, and tool results go to Venice — no xAI account involved. Grok Build’s own loop runs on this machine.
CursorPrime AgentLocal modelsOllama · LM StudioDirect providersyour own accounts
Local modelsOllama · LM StudioDirect providersyour own accountsWhich agent runs a session. The Claude Code engine is the default and routes through the providers you configure in Providers & privacy; the engines below replace the whole loop with another vendor's agent, under that vendor's account and policies. Pick per thread from the composer.
Prompts, project files the agent reads, and tool results go to Cursor under your Cursor account and draw on its plan. Cursor states that with Privacy Mode enabled on the account, prompts are not retained or used for training — that setting lives in your Cursor account, applies to the CLI, and Cloaky cannot see or verify it.
Sign-inAuth belongs to the CLI, not to Cloaky — run
cursor-agent login in a terminal once on this machine. Cloaky never stores or proxies that credential.CLI pathUsing /Users/demo/.local/bin/cursor-agent. Set a path only to override.auto-detect
ModelsFetched automatically when this engine is enabled — nothing yet. If that persists, the CLI is most likely not signed in (
cursor-agent login); Detect retries once it is.What carries overStreaming, file tracking and the phone companion work. Approval cards appear only when the agent asks Cloaky before an action, and most do not: measured 2026-08-07, Cursor and OpenCode read and write files entirely on their own, so no card appears and your permission mode does not reach them. Grok Build does route file access through Cloaky and does ask — but which tool an agent reaches for is its own decision, so treat that as behaviour observed today rather than a setting you control. Rewind, MCP toggles, verify-on-done, security scans and thread history are Claude Code engine features and stay off for these sessions.
Prompts go to whichever provider your OpenCode installation is signed into — that routing is configured in OpenCode itself, and its providers’ retention policies apply. Cloaky cannot see which provider that is.
InstallRuns
curl -fsSL https://opencode.ai/install | bash, which fetches from opencode.ai. Cloaky types it into the terminal and stops there — read it before you press Return. Installers for these CLIs commonly edit your shell startup files, and not all of them verify a checksum.Prompts, project files the agent reads, and tool results go to xAI under your xAI account and draw on its plan. Cloaky cannot see or verify what xAI retains.
Prompts, project files the agent reads, and tool results go to Moonshot AI under your Moonshot account and draw on its plan. Cloaky cannot see or verify what Moonshot retains.
Prompts, project files the agent reads, and tool results go to OpenAI under your OpenAI account and draw on its plan. Cloaky cannot see or verify what OpenAI retains. This row runs the community ACP adapter @agentclientprotocol/codex-acp, because the Codex CLI has no ACP mode of its own — the version shown is the adapter’s.
Prompts go to whichever provider you configured inside Prime Agent — Cloaky cannot see which one, and that provider’s terms apply. Prime Agent works by writing and running Python with your permissions; Prime Intellect states its worker and kernel processes are not a security sandbox. Configure a provider with prime-agent’s own /login before enabling.
esc
Cloaky Base is coming to macOS and Linux.
Free and open source, with every session running on Venice’s private tier.
Cloaky Pro has no public download.
Pro is the full workbench — every engine, local models, direct providers, fleet — and it is in a private build for the next three months. Email hello@cloaky.dev with what you would do with it; the user study group gets Pro.
Add retry logic to fetchUser+24−6
Read · src/lib/api.tsdone
Edit · src/lib/api.tsupdated
npm test -- --run apirunning
Describe the work — / commands, @ files
A complete desktop workbench
Threads, diffs, subagents, readable approvals, and the exact cost of every turn — packaged as a native desktop app.
v0.1.1 · not published yetRequest Cloaky Pro
Bashwrite
npm run lint -- --fix
Runs with your permissions. Review before allowing.
DenyAllow once
Needs input · 1
Readable approvals before execution
File changes and shell commands pause with the exact action in view. Approve once, for the session, or for the project.
Veniceprivate tier · zero retention by contract
GLM 5.2via Venice · private tier
Kimi K3via Venice · private tierQwen3 Codervia Venice · private tier
DeepSeek V4 Flashvia Venice · private tierVeniceprivate tier · zero retention by contract
Ollamalocal inference · stays on this machineLM Studiolocal inference · stays on this machine
OpenAI · Codexyour account · OpenAI termsGLM 5.2your account · provider termsKimi K3your account · provider terms
Claude directyour subscription · Anthropic termsOne workbench, one route
Every session runs on Venice’s private tier. Pick the model per session; the route never changes underneath you.
One workbench, your provider
Start with Venice, run locally, or use a provider you already pay for. Cloaky Pro keeps the active route visible in every session.
Get one email the day v0.1.1 ships
There is also Cloaky Pro.You are looking at Cloaky Pro.
Pro is the full workbench: every engine, including Cursor and Prime Agent; local models through Ollama and LM Studio; direct providers — a Claude subscription, an OpenAI-compatible endpoint, a GLM or Kimi coding plan; and fleet, bots and squads. Pro is in a private build. Base is the same bench with one route: every session on Venice's private tier, with no other provider account involved.
Cloaky Base download, free
- Engines
- Five: Claude Code, OpenCode, Codex, Kimi Code, Grok Build.
- Inference route
- Venice's private tier, every session. No other provider account involved.
- Local models
- Not in Base.
- Direct providers
- Not in Base.
- Bots & fleet
- Bots and squads with routines. Subagents in every session.
- Price & access
- Free and open source. v0.1.1 downloads here the day it ships, macOS and Linux.
Cloaky Pro request access
- Engines
- Every engine, including Cursor and Prime Agent.
- Inference route
- Venice, a provider you already pay for, or your own machine.
- Local models
- Ollama and LM Studio.
- Direct providers
- A Claude subscription, OpenAI-compatible endpoints, GLM and Kimi coding plans.
- Bots & fleet
- Fleet, bots and squads.
- Price & access
- Private build. No public download — ask for it by email.
Private buildPro is in a private three-month build, from mid-September 2026. There is no public download while it runs, and we will describe the features when they exist.
Say what you would do with it. We are picking a small first-class user study group — a private net of users who help us build Pro — and its members get Cloaky Pro.