Blackbox AI

You will enjoy exactly one thing here nobody else offers: /multi-agent sends the same task to Blackbox, Claude Code, Codex and Gemini, each in its own worktree, and a Chairman LLM picks the winner, which is a fun way to settle an argument about which agent is better on your code. Everything else is a competent terminal agent that does what the others do a little later.

Pick it if your employer buys it and you want the race. Pick Claude Code if you are buying for yourself, because one good agent you understand beats four you referee.

The risk is the pivot. Blackbox now positions itself as an encrypted inference router over 300 plus models, and a router company maintains an agent the way a landlord maintains a lobby: enough to keep tenants, not enough to make it home. The CLI docs describe no sandbox and no permission framework for local tool calls, so the agent runs with whatever your shell has.

Run it in a container of your own, since the product will not. What it does right: Remote Agent sandboxes exist, so unattended runs happen somewhere disposable, and the mess stays in the cloud.

No benchmark is published. One principled property is documented: plan mode divides a task into ordered steps the user can approve, comment on or rewrite before any change applies, which makes the plan an editable artifact rather than a log. The same loop runs non-interactively behind the -p flag, so the interactive and headless paths share one implementation rather than two that drift.

The consequence is that verification is front-loaded into the plan, and nothing in the docs describes what checks the execution. The observation: a plan you can rewrite is worth more than a plan you can only approve.

Read the pricing page as a strategy memo: annual committed spend, metered per token against per-model rates, discounts of 5 to 10%, no seats, no platform fee. That is the unit economics of a reseller, and a reseller's margin is whatever the labs leave, which they can change with a price sheet. The coding agent is the storefront; the inference is the inventory.

Likely acquirer: a cloud that wants a model marketplace with a customer list, or nobody, since clouds already have one. Position: do not build on it; buy the inference if the rates beat your contract, and keep the agent replaceable.

The demo is a terminal agent. Procurement: no seat price, so sixty developers becomes a forecast of tokens against rates like Nemotron 3 Ultra at $0.32 in and $0.80 out per million, committed for a year, which is a contract finance signs blind. Security gets single-tenant deployment and AES-256 at rest, which is more than most. Headless mode covers CI. Identity and audit are not described anywhere I can cite.

Onboarding is an install script. Not yet: a usage forecast someone will sign, and SSO and audit logging in writing before a pilot.

Closed source, so I score what I can touch. MCP servers go in ~/.blackbox/mcp.json, skills load from .blackbox/skills/, and hooks fire on PreToolUse, PostToolUse and Stop, which is enough to wrap every tool call in my own script, log it, or veto it. I can route through my own OpenAI, Anthropic or Google account, so the meter can be mine. No local models, so nothing runs air-gapped.

Grudging respect for the hooks, which are the right three events. None for the box, which I cannot read and could not fork if the router business eats the agent.