Before you start

You need Visual Studio Code 1.101 or newer, a model provider account or a local model runtime, and a repository you are allowed to modify. Work on a branch and keep a rollback path.

Install the agent

Bytos is installed from the Visual Studio Marketplace. Open the Extensions view in VS Code, search for Bytos, or run:

code --install-extension bytos.bytos

The extension ID is bytos.bytos. If the vscode: link from the website does not open, use the Marketplace listing or the command above.

Connect a provider

Open the Bytos panel, choose a provider and paste your own API key. Keys are stored in VS Code SecretStorage. Supported options include OpenRouter, Anthropic, OpenAI, Google Gemini, AWS Bedrock, Azure, GCP Vertex, Cerebras, Groq, any OpenAI-compatible endpoint, and local models through Ollama or LM Studio.

Start with a small, inexpensive model for setup. Switch to a stronger model for complex tasks once the workflow is working.

Run your first task

Describe a concrete task with a clear success condition, for example: “Add a unit test for the date parser and run the test file.” Attach relevant files or let the agent search the workspace. Bytos reads context and proposes a plan.

Good first tasks are small, testable and reversible. Avoid starting with a repository-wide refactor or a production migration.

A concrete first task

Open a small feature branch and ask for one specific outcome: “Add validation for an empty email field, update the existing test file, and run only that test.” A task like this has a clear success condition, a small blast radius and an easy rollback. It also exercises context reading, file editing, terminal approval and test output in one loop.

If the agent proposes a broader refactor, narrow the scope before approving. The plan is part of the output; correcting it early is cheaper than reviewing a large diff later.

Review and approve

  • Read the plan before execution and correct wrong assumptions.
  • Review diffs instead of accepting a summary.
  • Approve terminal commands individually until you trust the workflow.
  • Use checkpoints or Git to compare and restore.
  • Keep auto-approval scoped to the categories you actually need.

Bytos treats reads, edits, terminal commands, browser access and MCP tools as approval-gated by default. Auto-approval is opt-in.

Review checklist

  • Does the diff match the requested scope?
  • Are there unrelated formatting or dependency changes?
  • Did the agent run the relevant tests, not just a build?
  • Are secrets, local paths or credentials present in the diff?
  • Can you explain every changed line before merging?

Longer tasks with Auto Phase

Auto Phase lets you define up to ten sequential phases. Each phase can use its own provider, model, prompt, attachments and MCP allowlist. A phase can pass a short completion report to the next phase. This is useful for a multi-step feature: plan, implement, test, review, document.

It is sequential, not parallel. Start with two or three phases and inspect the reports before scaling up.

Troubleshooting

  • VS Code did not open: allow the vscode: hand-off in your browser, or install from the Marketplace.
  • code is not on PATH: run “Shell Command: Install 'code' command in PATH” from the VS Code Command Palette.
  • Provider error: check the key, model name, account limits and provider status.
  • The agent is too slow: use a smaller model for exploration and a stronger model for planning or review.
  • A command was blocked: review the approval category and the workspace boundary; do not disable safety controls globally.

Related guides

Try Bytos in VS Code

Bytos is an open-source, bring-your-own-key AI coding agent for Visual Studio Code. It plans first, asks before risky actions, and keeps your provider key in VS Code SecretStorage.

FAQ

Does Bytos require a Bytos account?

No. The extension runs with your own provider key. A Bytos account is optional and is used for website sign-in and the VS Code hand-off.

Can I use a local model?

Yes, through Ollama or LM Studio or any OpenAI-compatible local endpoint. Local inference keeps model requests on your machine, but tools you approve may still use the network.

Is the agent safe on production code?

Treat it as a new contributor with limited access. Use a branch, review diffs, approve terminal commands and keep secrets out of the workspace.

Sources and verification

Product details change. This guide was reviewed on 11 November 2026 against the linked official documentation; verify current features, pricing and licensing before choosing a tool. No benchmark scores or independent performance claims are used here.