cloudparts

Blackbox AI

Coding agents that run in cloud sandboxes and race each other (Blackbox, Claude Code, Codex, Gemini) on one task, from a company that now sells mainly enterprise model inference.

What it is
AI coding agents (cloud agents)
Used for
Model routers and gateways
Type
Commercial product
Made by
Blackbox AI Technologies Inc.
License
Proprietary
Facts checked
13 October 2026
Website

What it is

Blackbox’s Agents API starts a coding agent in a fresh cloud sandbox with a terminal, file system and Git access, connected to your GitHub, GitLab or Bitbucket repository. You create a run over HTTP or in the browser app, stream its terminal events, read and write files in the sandbox, send follow-up instructions, and get a pull request with a scoped branch and a generated summary. Its distinctive feature is multi-agent mode: the same task goes to Blackbox’s own agent, Claude Code, Codex and Gemini in parallel, and a judge model Blackbox calls Chairman LLM ranks the implementations and picks one, which you can override.

The same ideas come in a terminal agent, Blackbox CLI, which Blackbox describes as open source. It has a plan mode, reads AGENTS.md, hooks, skills and MCP servers, runs /multi-agent with each agent on its own Git worktree, and has a headless mode (-p) for CI. Blackbox also had VS Code extensions: the original BLACKBOX AI completion assistant, with over 5 million installs since 2022, and an autonomous Blackbox Agent extension. Both were last updated on 25 August 2026 and had been unpublished from the Visual Studio Marketplace by October 2026.

The company has changed what it sells. Its press kit now calls Blackbox “the high-trust inference layer”: a router to more than 300 models behind one OpenAI-compatible endpoint, with provider fallback, end-to-end encryption, zero data retention and dedicated single-tenant deployments of open-weight models. The only plan on its pricing page is Enterprise, an annual committed spend billed per token, and the coding agents draw on the same balance.

Why teams pick it

  • Multi-agent runs compare several vendors’ agents on the same task and keep the best branch
  • Fully programmable: create runs, stream events and manage pull requests over HTTP, or run the CLI headless in CI
  • Reads the conventions other agents use (AGENTS.md, MCP, hooks, skills), so setup carries over
  • Enterprise privacy controls: zero data retention, end-to-end encryption, SSO and single-tenant deployments

Watch out for

  • Enterprise only: access is through an annual committed-spend contract, with no published self-serve plan
  • Running four agents on every task multiplies token spend
  • Code and context go to Blackbox’s cloud, and closed models are reached through their providers
  • The company’s focus has shifted to enterprise inference: its docs now cover only the inference API, and its VS Code extensions are no longer listed on the Marketplace

Alternatives

What all cloud agents have in common →Compare side by side →
DevinCommercialCognition’s cloud coding agent: each session gets its own VM with a shell, editor and browser, takes a task from Slack, a ticket or the web app, and comes back with a pull request.

How it fits with other technologies

See it in a real stack

No architecture uses Blackbox AI yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2022–2026

Blackbox was founded by Robert, Richard and Roger Rizk, who lead it as CEO, CTO and COO. Its best-known developer product has been the BLACKBOX AI extension for VS Code, published in September 2022, which offered real-time code completion and documentation and passed 5 million installs. An autonomous agent extension followed in November 2024, able to create and edit files, run commands and use a browser with the user’s permission.

From there Blackbox moved to agents that run outside the editor: the CLI for the terminal and the Agents API for cloud sandboxes, both built around running several agents at once and choosing the best result. In June 2026 it described an orchestrator-executor design in which a strong model plans and verifies while a cheaper one implements, which lifted a Terminal-Bench 2.0 score from 58.4 to 69.7.

By mid 2026 the company was presenting itself as an enterprise inference provider. In July 2026 Artificial Analysis measured it as the fastest provider of NVIDIA’s Nemotron 3 Ultra, at 454 tokens per second, and its pricing page now lists a single Enterprise plan, an annual committed spend that covers both inference and the coding agents.

Why it’s built this way

Run several vendors’ agents on the same task
Instead of betting on one agent, Blackbox sends a task to its own agent and to Claude Code, Codex and Gemini, and lets a judge model pick the cleanest implementation, so teams compare results instead of waiting on one attempt.
Split the agent into orchestrator and executor
A solo model tends to ship plausible work that fails the real check. A strong orchestrator that explores, locks a plan and verifies, with a cheaper executor doing the edits, raised the ceiling on hard tasks.
Sell inference, not seats
Every surface, from the router to the CLI and Agents API, bills against one per-token balance, with no per-seat charge.

Timeline

  1. 2022

    BLACKBOX AI extension for VS Code published in September

  2. 2024

    Blackbox Agent, an autonomous VS Code extension, in November

  3. 2026

    Orchestrator-executor results on Terminal-Bench 2.0, in June

  4. 2026

    Fastest Nemotron 3 Ultra provider per Artificial Analysis, in July

  5. 2026

    Pricing page lists only an Enterprise plan on committed spend

Sources6 · facts checked 13 October 2026

Improve this page on GitHub