Vector

One workspace

Everything shipping takes.

Not a chat box bolted onto an editor.

Point a swarm of agents at separate branches and let each one spawn subagents for the parts. Set up the database, deploy to your own cloud account, and have the change reviewed before it reaches main. Vector does all of it in one window, over the repository already on your disk — with a browser it can drive, a tokenomics engine that prices its own agent's runs, 134 one-click plugins, and any MCP server you write yourself.

• As many agents as you want

  No cap. Each one works in its own copy of the repository, and you merge the ones that landed.

• Subagents for the parts

  A very hard task splits itself, the pieces run in batches of up to eight, and a card shows what each one is doing.

• Ships to your own cloud

  Provision Supabase, sync environment values, publish to your Vercel account.

• Reviewed before main

  A summary and line comments on each pull request once automatic review is on, or run the same pass on your branch first.

• A browser it drives

  It loads your app, clicks through it, reads the console, and tells you what actually happened.

• A tokenomics engine

  Runs are priced from the tokens the provider reported, at catalog rates where the model has a price. In the desktop app, each request is sized first, so "hi" gets one small, direct reply, and a Run Guard stops a run within one step of the cap you set.

Low cost, measured: Vector's offline efficiency benchmark, archived at docs/benchmarks/vector-efficiency-2026-10, ran Vector's runtime at an 11 October 2026 revision (commit 6f7f3c2ea) and Claude Code 2.1.215 through identical scripted tool steps against a local fake model server, with the same number of model requests, on three fix tasks. In a first session, Vector's modelled spend was 67–75% below Claude Code's at the same list rates. The benchmark measures what each agent sends, not the quality of answers or your provider's bill, and does not show that Vector is cheaper than other agents in general.

• GitHub
• Linear
• Sentry
• Supabase
• Vercel
• Figma
• Notion
• Stripe
• Postgres
• Slack
• Playwright
• Your own MCP server

Self-evolving · new in desktop 1.99.110

An agent that learns you and your company.

Vector Brain: every session starts from what the last one learned.

Vector Brain is Vector's own technology. Give it your specs, PDFs and docs once. It reads them on your computer, links them to the code that implements them, and keeps what proved useful in runs that passed. Then, offline, with a code graph and retrieval of its own, it gives your model only the few lines of them that bear on each request.

Your prompts, your data, your Vector: the brain lives on your computer, you choose who runs the model for each request, and you can see, undo or erase what it learned. Free models labelled May train on your prompts are the exception.

• A brain for each repository

  Rules, memory, what the agent learns and the documents you give it, in one local graph. It adds to it as it works, and one click undoes anything it learned.

• Linked to your code

  Specs, PDFs and docs are read into requirements, limits and decisions, linked to the code they name, and flagged when the code drifts from them.

• Only what the task needs

  The tokenomics engine sends the few cards a task needs and stops at its budget. "hi" carries no memory at all, so it costs nothing extra.

On Vector's knowledge eval of 19 October 2026, recall through code links brought in the governing requirement for 40 of 40 tasks, against 1 of 40 by keywords alone, at 131 tokens a request on average. The eval is five short specs and 58 requests written alongside the feature, not an independent benchmark.

Privacy

Your code lives on your computer.

Vector's servers keep no copy of your prompts, Vector never trains on your code, and you pick who runs the model.

• Local first

  Sessions, history and checkpoints are saved on your computer, and desktop dictation is transcribed there too.

• Straight to your provider

  With your own key, prompts go straight to that provider under its terms, never through Vector, and a model on your own computer keeps them there.

• No copy on our servers

  Vector's servers never keep your prompts or code, except what you choose to send: a share link you publish, public for 30 days at most, or a bug report.

• Never trained on

  Vector collects nothing for training, and Velocity, the 550B-parameter model we're building, won't be trained on customers' code.

• On Vector's own models

  Codium, coming soon, routes only to providers OpenRouter records as zero data retention. Requests to free models whose providers don't train ask them not to train on your prompts, and free models whose providers may train are labelled May train on your prompts and sent only after you accept that once. Providers may still keep prompts.

• Counts, not content

  Usage counts are totals like sessions and tokens, never your prompts, code, file names or output. You can switch them off, and crash reports are off by default.