Codex
Sign in with ChatGPT. Requires a compatible plan.
VEKTOR CLI · CODING AGENT
Meet VEKTOR: the agent that explores your project, works with your tools and brings your ideas to the terminal.
Your models. Your terminal.CONTEXT THAT CONTINUES
YOU STAY IN CONTROL
BUILT FOR THE WAY YOU CODE
MEET VEKTOR CLI
Explore a repository, plan a change or delegate research in the same session. Choose an example to see the experience.
Explain how authentication works in this project.
I’ll review the authentication flow and its references before proposing changes.
src/auth.ts
src/session.ts
tests/auth.test.ts
The flow starts in auth.ts and stores the session in session.ts. We can review credential renewal together.
ONE WORKSPACE
Browse files, edit and review changes without leaving the terminal.
1import { createSession } from './agent'23export async function start(project: string) {4const session = await createSession({5cwd: project,6permissions: 'ask',7})89return session10}
Vektor
I’ll review the sign-in flow and its tests before proposing changes.
Illustrative view. Tabs show sample files; the actual editor is in the CLI.
A WAY INTO YOUR WORKSPACE
Start the CLI in your terminal, pointed at your project directory.
$ vektorcli --cwd ./my-projectWindows x64 preview available below. macOS, Linux and npm publication are coming later.
AN OPEN ECOSYSTEM
Use your ChatGPT or Claude account through compatible agents, or connect your own API keys. Choose your connection with /connect.
Sign in with ChatGPT. Requires a compatible plan.
Sign in with Claude. Requires a compatible plan.
OR CONNECT YOUR OWN API KEY
GPT · Your API key
Claude · Your API key
Multi-model catalog · API key
DeepSeek models · API key
Groq inference · API key
Multi-model gateway · API key
ALSO THROUGH OPENCODE GO
Choose from these families in the OpenCode Go catalog. Availability and access depend on the provider and your plan.
ACP
Use an Agent Client Protocol compatible agent inside VEKTOR. Each agent keeps its own configuration and authentication.
MCP
Connect local MCP servers over stdio to make their tools available to the VEKTOR agent.
Each provider manages access, limits and billing. Subscriptions require a plan that supports Codex or Claude Code; API usage is billed separately.
THE VEKTOR ENGINE
Capabilities already implemented in the project. Built around everyday work with code.
PROVIDERS
Connect compatible providers and change models or reasoning levels within the session.
SESSIONS
Conversations and tabs persist per workspace so you can pick up the context when you return.
SUBAGENTS
Use subagents to investigate in parallel and bring their findings back into your conversation.
PERMISSIONS
The plan, ask and accept-edits modes let you decide when to review and approve an action.
MCP + SKILLS
Add MCP servers, skills and custom tools to work with the context of your project.
LSP
Query definitions, references and diagnostics through your configured language servers.
MEASURED ENGINEERING
We measure the work of the client and local harness. Model and network speed are evaluated separately.
1.43ms
8,010 history rows. Median of three warmed runtime runs.
20.26ms
Real ConPTY, a restored 8,010-row session and ten full repaints.
150
An initial 150-row window keeps the initial drawing work bounded when opening a long session.
Local measurement · Windows x64 · Bun 1.4.2 · Nov 16, 2026
The runtime benchmark reconstructs history and processes 2,000 synthetic chunks in 40 ms bursts. Each run takes the best replay of three internal iterations; the displayed value is the median of three runs. The terminal test uses a compiled binary, one 120×40 ConPTY run and ten keypresses per repaint class. No models or network requests are involved.
The first paint of that restored session took 1.77 s. Full repaint measured 23.94 ms p95. These are local idle-session results, not model generation latency.
VEKTOR can run its own agent or connect to an external agent through ACP. Here is how each part of the system changes.
| Architecture | VEKTOR local | VEKTOR + ACP |
|---|---|---|
| Where the loop lives | Inside the TUI process | In the configured external agent |
| Main interface | Native terminal | The same VEKTOR terminal |
| Communication | Direct in-process calls | JSON-RPC with the agent process |
| Comparable timings | Local reports available | Pending the same workload |
Architecture alone does not determine which is faster. Model, provider and workload also matter.
GET STARTED
Download the Windows x64 preview. The editor, files and agent share one workspace.
Preview 0.2.0 · ZIP · Unsigned executable
macOS and Linux: public distribution coming later. This preview does not require Bun to run the CLI.
SHA-256Extract the ZIP and run the command in PowerShell. Use your project path.
.\vektorcli.exe --cwd C:\ruta\mi-proyectoChoose Codex or Claude Code. The first connection installs its adapter and needs Node.js 22+ and npm. The provider handles sign-in.
/connectF3 opens files, Ctrl+S saves, F4 returns to the agent and F6 shows Git changes. In a wide terminal, code and agent sit side by side.
F3 files F4 agent F6 reviewYes, through the Codex and Claude Code adapters, with a plan that includes access to those products. VEKTOR does not turn a chat subscription into an API key.
Yes. OpenAI, Anthropic, OpenRouter, DeepSeek, Groq and OpenCode Go are available through API connections. The provider bills that usage separately.
The preview combines a local editor, explorer, tabs, Git review and agent in the terminal. The editor does not yet offer semantic completion or an integrated debugger.
You choose the agent permission mode. The editor saves when you ask and detects whether another process changed the file before overwriting it.
THE NEXT CHAPTER
VEKTOR CLI is the current product. The IDE will be the next interface to keep building in the same direction.
A new space for working with code. More details when it is ready.
THE TEAM BEHIND IT
At Vektorial Labs, we are building tools for coding with agents. VEKTOR brings that vision together: project context, connected tools and decisions in the hands of the person creating.
GET IN TOUCH
Want to talk about VEKTOR, collaborate or learn about what we are building?