Voidleap Code
An IDE for coding
with AI agents
Voidleap Code is a desktop app for building software with AI agents. It is an integrated development environment (IDE), built for professional developers and AI engineers. You direct the agents and see everything they do: every prompt, every tool call, every changed file. It runs on your computer, with your own AI provider accounts or a local model. Your code and prompts do not pass through our servers.
News
The problem
Coding agents became black boxes
Agents write more of the code now. Most tools hide how. You see a summary, not the steps. When output degrades, you guess.
The tools are built to sell inference, not to make you better. And they try to work for everyone, from first prompt to production, so they commit to no one. Many still act like proofs of concept. A strong model is only half the equation, and the other half is missing.
What we built
A harness around the agents, not a wrapper around a model
Voidleap Code is a desktop IDE for working with coding agents. You give an agent a task. You watch it work: every prompt, every tool call, every file it touches. You can stop it, correct it, or fork the thread and try again.
It is not a fork of an existing editor, and it does not wrap another vendor's CLI. A wrapper never owns the loop. We built the harness from scratch and own every byte that goes to the model. That is why you can open the context and edit it. A real code editor is still there for when you want to type.
The harness is the same for every provider. An agent, a skill, or a workflow you build runs unchanged on any model. Swap providers mid-thread. Mix them in one task. Mixing is not a cost trick; it gets you a better result.
Run one agent or several in parallel, each in its own panel. The app keeps every thread in one view.
It is opinionated software, made to act like a professional tool, not a proof-of-concept toy. We build for professional developers and AI engineers.
How it works
Your machine runs the app. You pick the model.
The app installs on your computer and runs locally. Today that means Mac and Windows, and more platforms are in testing. Application data stays on your disk.
The model is separate. Connect a provider subscription or API key: Anthropic, OpenAI, Google Gemini, GitHub Copilot, AWS Bedrock, Azure AI Foundry, xAI, Meta, Kimi, OpenRouter, Opper, Antigravity, Moonshot AI, or Berget. Their terms apply to what you send. Or run a local model through Ollama or any OpenAI-compatible server, and no provider sees your work at all.
Voidleap sits outside that loop. We do not see your code or prompts. The app only calls our servers for a periodic license check.
Safety
Guardrails, not brakes
Agents run commands on your machine, so the rules are strict and visible. Four execution modes, from read-only to fully automatic. Approve every action, or let it run.
The app parses every bash and PowerShell command before it runs: paths, pipes, redirects, subshells. Output streams live. Sensitive paths like .ssh and .env are blocked out of the box.
An optional OS sandbox adds a hard wall, with strict modes and network allow and deny lists. If the sandbox is unavailable, it fails closed. All of it is managed visually, and a company policy can lock any of it for the whole team.
Control
See everything. Change anything.
The dashboard has 30+ charts: cost, tokens, latency, cache hits. A live feed shows every model call across every running thread. You learn what your work costs while it runs.
Context is visible too. Click any message, tool call, or file in a thread and tell the agent to forget it. The app classifies tool calls as it goes, with the agent's help: temporary results drop out on their own. Six operations, from a light trim to a full hand-off, keep only relevant work in the cached part of the context. Long threads stay useful without restarts.
Agents carry only what you allow. Allow and deny lists decide which tools and skills enter the system prompt. Skills and MCP tools load when needed, not up front. Unused baggage costs you no tokens.
Every agent, skill, and system prompt is a plain file with a visual editor on top. Override or replace, not just append. Three scopes: global, workspace, project. A workspace groups similar projects, so shared agents and skills live once instead of being copied around.
The built-in browser rounds it out. Annotate a page and hand the marks straight to the agent.
All of it adds up to token efficiency. You pay for work, not baggage.
The comparison
How it compares
Checked August 2026. These tools ship fast, so verify anything that matters to you.
| Voidleap Code | Claude Code CLI | Claude Code Desktop | Codex CLI | ChatGPT Desktop | Cursor | opencode / Pi | |
|---|---|---|---|---|---|---|---|
| Purpose-built agentic IDE | Terminal | Chat app | Terminal | Chat app | IDE fork / Chat app | Terminal / Chat app | |
| Any providers natively supported in one harness | Some via routing | ||||||
| Local models, full tool calling | |||||||
| Requests never pass through the product vendor | |||||||
| System prompt replacement per agent | CLI flag | Global | |||||
| Remove any item from context | |||||||
| Native local symbol index, code-navigation tools | Grep | Grep | Grep | Search | Cloud index | Grep | |
| Built-in code editor | |||||||
| Full custom grid layout for all agents | Tabs | Chat list | Tabs | Tabs (app) | |||
| Visual editor for agents, skills, system prompts | Partial | Partial | Partial | ||||
| Live token usage/cost dashboard | |||||||
| Workspace-scoped agents and skills | |||||||
| Browser with annotation | |||||||
| Multiple threads in one worktree |
Next
In internal testing
Three things run in internal testing: a desktop app for Unix, a standalone server you host yourself, and a mobile app for iOS and Android. The server and the mobile app are planned for Pro and up.
No dates. Details can change while we test. They ship when they work.
FAQ
Frequently asked questions
Where does my code go?
Voidleap Code uses a local agentic runtime on your machine. You choose where inference runs: through a provider or on local hardware. The architecture is locked local - on localhost only, all application data in ~/.voidleap/. The inference is your choice: a cloud provider for raw capability, or a local model for full provider privacy. Voidleap Code doesn't see your code either way.
How is Voidleap Code different from Claude Code, Cursor, or Pi?
Claude Code and Codex are built around their vendors' own models. Cursor is an IDE, but your prompts route through Cursor's servers, and it now belongs to a model vendor too. opencode and Pi live in the terminal, come minimal by design, and you configure them by hand, per tool. Voidleap Code is one desktop harness for all of it, built from scratch rather than forked: the same agents, skills, context tools, and charts work on every provider or local model, configured visually, and you can mix providers in one task. The short version: they build for everyone. We build for professionals.
Can I work without a constant internet connection?
Yes, for agent execution. With local models, your prompts and code stay on your machine - no provider in the loop. Voidleap Code only requires periodic license validation; once validated, you can stay offline for 24h.
Can I see what I'm spending?
Yes, 30+ charts cover cost trend, per-model, per-agent, per-thread, per-tool, and cache hit rate - exportable.
What happens when my thread gets long?
You stay in control. The visual context lets you click any message, tool call, or file and tell the agent to forget it - no thread restart. AI Prune in suggest mode so you review before anything is removed. Four slash commands: /context, /trim, /prune, /compact.
Which models can I use?
Anthropic, OpenAI, Google Gemini, GitHub Copilot, AWS Bedrock, Azure AI Foundry, xAI, Meta, Kimi, OpenRouter, Opper, Antigravity, Moonshot AI, and Berget, plus local models through Ollama or any OpenAI-compatible server. Switch per thread, per agent, or per turn. Local models get full tool-call parity.
What can the agent run on my machine?
You decide. Four execution modes: Default, Read-only, Careful (approve every action), Yolo. AST-aware command parsing understands paths, redirects, pipes, subshells, and heredocs. Sensitive-path defaults block .ssh, .env*, .aws, .gnupg out of the box. Optional macOS Seatbelt sandbox.
How customizable is it?
Every agent, skill, hook, slash command, security rule, and system prompt is editable through the library or markdown. Three scopes: Global, workspace, project lets you configure once and override anywhere. Override or replace, not just append.
How do I install it?
A desktop app. Mac and Windows today, more platforms in testing. No bun, no npm, no build step. Auto-updates with signed releases.
Download it. Judge for yourself.
Bring a task you already solved with your current tool. Run it in Voidleap Code and read every step the agents take. Then compare.
Free plan. No credit card, no trial clock.
Say hello
Agents write more of the code now. Most tools hide how. You see a summary, not the steps. When output degrades, you guess.
The tools are built to sell inference, not to make you better. And they try to work for everyone, from first prompt to production, so they commit to no one. Many still act like proofs of concept. A strong model is only half the equation, and the other half is missing.
Voidleap Code is a desktop IDE for working with coding agents. You give an agent a task. You watch it work: every prompt, every tool call, every file it touches. You can stop it, correct it, or fork the thread and try again.
It is not a fork of an existing editor, and it does not wrap another vendor's CLI. A wrapper never owns the loop. We built the harness from scratch and own every byte that goes to the model. That is why you can open the context and edit it. A real code editor is still there for when you want to type.
The harness is the same for every provider. An agent, a skill, or a workflow you build runs unchanged on any model. Swap providers mid-thread. Mix them in one task. Mixing is not a cost trick; it gets you a better result.
Run one agent or several in parallel, each in its own panel. The app keeps every thread in one view.
It is opinionated software, made to act like a professional tool, not a proof-of-concept toy. We build for professional developers and AI engineers.
The app installs on your computer and runs locally. Today that means Mac and Windows, and more platforms are in testing. Application data stays on your disk.
The model is separate. Connect a provider subscription or API key: Anthropic, OpenAI, Google Gemini, GitHub Copilot, AWS Bedrock, Azure AI Foundry, xAI, Meta, Kimi, OpenRouter, Opper, Antigravity, Moonshot AI, or Berget. Their terms apply to what you send. Or run a local model through Ollama or any OpenAI-compatible server, and no provider sees your work at all.
Voidleap sits outside that loop. We do not see your code or prompts. The app only calls our servers for a periodic license check.
Agents run commands on your machine, so the rules are strict and visible. Four execution modes, from read-only to fully automatic. Approve every action, or let it run.
The app parses every bash and PowerShell command before it runs: paths, pipes, redirects, subshells. Output streams live. Sensitive paths like .ssh and .env are blocked out of the box.
An optional OS sandbox adds a hard wall, with strict modes and network allow and deny lists. If the sandbox is unavailable, it fails closed. All of it is managed visually, and a company policy can lock any of it for the whole team.
The dashboard has 30+ charts: cost, tokens, latency, cache hits. A live feed shows every model call across every running thread. You learn what your work costs while it runs.
Context is visible too. Click any message, tool call, or file in a thread and tell the agent to forget it. The app classifies tool calls as it goes, with the agent's help: temporary results drop out on their own. Six operations, from a light trim to a full hand-off, keep only relevant work in the cached part of the context. Long threads stay useful without restarts.
Agents carry only what you allow. Allow and deny lists decide which tools and skills enter the system prompt. Skills and MCP tools load when needed, not up front. Unused baggage costs you no tokens.
Every agent, skill, and system prompt is a plain file with a visual editor on top. Override or replace, not just append. Three scopes: global, workspace, project. A workspace groups similar projects, so shared agents and skills live once instead of being copied around.
The built-in browser rounds it out. Annotate a page and hand the marks straight to the agent.
All of it adds up to token efficiency. You pay for work, not baggage.
Three things run in internal testing: a desktop app for Unix, a standalone server you host yourself, and a mobile app for iOS and Android. The server and the mobile app are planned for Pro and up.
No dates. Details can change while we test. They ship when they work.
