One coding agent.
Local or hosted. Fewer tokens.
VG Code maps your repository once, so the model works with focused context instead of repeatedly searching your code. Run a model locally or hosted. Every change arrives as a real diff and, in the default Agent mode, waits for your approval.
VG Code recommends what will run well on your machine, and shows hosted model prices before you choose.
Inside VS Code. From the terminal. For automation and CI.
npx @vibgrate/cli scanno install·Nothing is installed globally — ideal for CI or a one-off scan.
Fewer tokens. Easier local models.
Map the repository once. Give each task only the context it needs. That means less searching and fewer tokens spent finding the same code again.
Fewer tokens for the same work
In our published benchmark, focused context used about 34% fewer tokens for the same work on a ~300-file codebase.
See the methodology and full results →Local models without extra setup
VG Code recommends a local model that will run well on your machine and handles the setup for you. Use the recommended model, or keep using Ollama, LM Studio or a model you already have. You can always choose something else.
How local models work →Hosted models with visible cost
Because the model receives less context, hosted runs use fewer tokens at the same model rate. With Vibgrate Relay, you can see the price of each model before you choose it. Use Relay, or connect your own provider.
Explore Relay →~34% fewer tokens, same results
The same coding tasks, the same model, with and without the Vibgrate code map — on a ~300-file codebase. We count only the tasks both runs solved, so it is fewer tokens for the same result, never “cheaper because it gave up”.
Measured with a frontier coding model. On tiny repositories grep already finds the code and the map costs more — we publish that part of the curve too. See the full benchmark
Best experienced in VS Code
VG Code is easiest to use inside Vibgrate for VS Code. The same agent is also available from the terminal and for automation and CI.
Focused context for each task
The model works with the code and library versions relevant to the task instead of repeatedly searching the repository.
Local or hosted — same experience
Run locally or choose a hosted model without leaving the task. VG Code can recommend what will run well on your machine.
Nothing changes until you say so
Review a real diff before anything changes. Open it in VS Code’s diff view, approve it when you are ready, and undo an approved change if you need to.
See what else will be affected
See the related code and tests before a change lands. Afterwards, VG Code runs your checks and shows what failed if something goes wrong.
Inside VS Code, conversations stay available between tasks. Run chats in parallel, search or reopen earlier work, and fork a conversation when you want to try another approach.
The same VG Code agent also runs from the terminal and in VS Code Chat. The full VG Code panel includes the richer session history and review experience.
Works from the code you actually have
VG Code maps your repository once and uses that map to find the code relevant to each task. The model gets focused context instead of repeatedly searching the whole repository.
Your repository, mapped once
Before changing code, VG Code identifies the parts of the repository relevant to the task and how they connect. That gives the model a focused starting point.
Docs for the versions you actually have
VG Code uses documentation matched to the library versions in your project, helping it avoid APIs that do not match the code you run.
Your uncommitted work counts
The map includes your current working changes, not just the last commit. It also stays aware of edits made during the task.
You stay in control
In the default Agent mode every change waits for your approval. VG Code shows what it plans to do, what changed and what happened afterwards.
Review every change
In the default Agent mode every edit and command waits for your approval, while read-only work continues without interrupting you. There is no quick-apply path around that gate — the only way past it is switching to Auto accept yourself.
Changes land where they should
VG Code checks proposed changes against your repository before writing them. If a change depends on something that is not there, it stops before changing a file.
It checks the result
After a change, VG Code runs your checks. If something fails, it shows what failed and gives the model a focused result to work from.
Three permission modes, and you pick
Agent is the default: it works directly and asks you to approve each edit and command. Plan explores and writes you a plan, with file writes and commands blocked. Auto accept approves its own edits and commands — a command denylist, default-deny outbound network and shell isolation stand in for your review. Nothing switches modes but you.
Secrets stay out of the model
Known secret files and credentials are kept out of model context, and detected tokens are masked. Commands can run in an OS-level sandbox, with network access restricted in automated runs.
See security details →Where your code goes, plainly
With a local model, the task and its context stay on your machine. With a hosted model, the context for that turn is sent to the provider you selected. VG Code only uses endpoints you configure.
Free locally. Hosted when you want it.
VG Code is included with Vibgrate CLI and Vibgrate for VS Code, with no account required. Local models use your own hardware.
For hosted models, use Vibgrate Relay with visible per-model pricing, or connect your own provider and pay them directly.
Automated dependency remediation with vg fix is separate.
Describe the change. Read the diff. Decide.
VG Code is easiest to use inside VS Code. The same agent is available from the terminal and for automation and CI. In the default Agent mode, nothing changes until you approve it.
npx @vibgrate/cli scanno install·Nothing is installed globally — ideal for CI or a one-off scan.