Self-hosted by default
Commander installs on your own VPS in one command. We never see your runs, your prompts, or your integration tokens.
AGENT INFRASTRUCTURE
AI agents you actually control.
Commander is a self-hosted orchestrator for AI agents. Your VPS, your provider keys (Claude, Gemini, GPT) — never ours.
Join the waitlist
No spam. One email when early access opens.
commander · run #1042
▶task: refactor auth middleware
↳model: claude-opus-4 · provider: anthropic
⚙tool_call: write_filepending
◈approval gate
canUseTool("write_file") → awaiting approval
→ notified via Telegram
✓
Built for operators, not platforms.
Provider-agnostic by design
Anthropic, Google, OpenAI — switch in a dropdown, run them side by side, never lock to one. Commander is the orchestrator; the model is your choice.
Your keys, your bill
Bring your provider keys. Costs are yours, scale is yours, rate limits are yours. We don't proxy a single token.
Approval gates, not autopilot
Every tool call is gated by a `canUseTool` surface. Approvals route to Telegram or email — watch what the agent does before it does it.
Updates as the moat
Monthly minor releases, weekly patches, security CVEs in 24h. Subscription includes everything we ship.
COUNCIL · MULTI-LLM COLLABORATIVE
Three minds on every problem.
The first consumer coding agent where models actually collaborate. Claude and Gemini take turns on a single run, share a scratchpad, and review each other's work — with a third frontier model joining the rotation.
- One run, multiple models, one scratchpad — not parallel chatter, real turn-taking
- Per-turn audit: provider, tokens, cost, status — for every move every model makes
- Strict or lenient fail-policy — you decide what happens when one model trips
BETA · ROLLING OUT
Verifiable runs, not promises.
Commander is moving from charge-then-hope to delivery-vs-payment. Every run will produce a signed receipt anyone can verify offline.
Delivery-vs-payment settlement
Rolling outStripe authorizes the charge up front; capture only fires after the worker confirms the run completed. Failures auto-reverse — no support ticket.
Signed receipts
Rolling outEvery run emits an Ed25519-signed claim — run id, model, tools used, outcome — verifiable offline against our published key.
Public verifier
SoonPaste a receipt at verify.dekimu.com to confirm authenticity and inclusion in the daily anchor.
verify.dekimu.com ↗TIER PREVIEW
Council
At early-access launchThree frontier models on every coding run. They take turns, share a scratchpad, and review each other's work.
What's included
- Multi-LLM collaborative loop — Claude, Gemini, and a third frontier model take turns on a single run
- Shared scratchpad — each turn reads what the previous model wrote, including tool output
- Per-turn audit trail — provider, tokens, cost, status logged for every turn
- Cost drill-down by provider in the dashboard
- Strict and lenient fail-policies — abort on first error or push through
- All Pro features included
Council ships with Commander — self-hosted, your VPS, your provider keys. Pricing and availability announced when early access opens.
Join the Council waitlist →WHY COUNCIL
Cursor runs Claude. Aider picks one model per task. Devin runs one model end-to-end. None of them put two frontier models on the same problem at the same time.
Council does. On a single run, Claude and Gemini take turns — with a third frontier model joining the rotation — sharing one scratchpad, reading each other's tool output, and reviewing each other's work. Every turn is audited: provider, tokens, cost, status.
It's the first consumer coding agent where models actually collaborate, not just hand off.
Pricing at launch.
Early access — Q3 2026.
Commander goes external when the foundation is solid. Get on the waitlist now and we'll mail you once the door opens.