Agent budgets are guesses, and nobody checks them. We measured ours across 49 governed runs — then fixed what the measurement found.
No key, no signup — one click runs the real budget, permission and audit gates, refusals and all.
Across work whose real cost moved 3.6×, the estimate moved 1.17× — it read the goal's string length, not the work. 97% of that error is systematic, not run-to-run noise. Estimating from the planner's task graph instead cuts the scatter 47% and lands 94% of predictions within 2×.
| Task size | Predicted | Actual | n |
|---|---|---|---|
| Small | 3.71¢ | 1.96¢ | 17 |
| Medium | 4.14¢ | 3.69¢ | 14 |
| Large | 4.33¢ | 7.05¢ | 18 |
A declarative charter defines the mission, the money, and what counts as done. Everything after flows through four gates — no task consumes a dollar or a permission it hasn't earned.
A goal becomes a dependency-aware task graph, each task mapped to a declared competency and a token budget.
Each task clears the budget, daily-burn, and margin checks, and runs only under a just-in-time permission that expires when it's done.
Outputs are graded against a category-tuned rubric; code isn't accepted until its tests pass, and a failed task self-repairs before it ships.
Every report is sealed with a SHA-256 proof hash in an append-only ledger; settlement runs over Stripe or x402/USDC.
Per-task ceilings, a daily burn cap, a runway floor, and a session circuit breaker that fast-fails a runaway loop before it drains the account.
Agents earn capabilities through a track record, and when one hands work to another — a sub-agent, an MCP tool, Claude Code — the grant can only narrow. A child spends its parent's budget, never a fresh one, and revoking a parent revokes everything carved from it.
Every deliverable is verified against the charter's success criteria and sealed with a SHA-256 proof hash in an append-only ledger you can verify offline.
One entry in your MCP config and Claude Code or Claude Desktop runs under a budget check, scoped authority, and a verifiable receipt. No tool on that surface can raise a budget or grant a capability — the levers stay in the charter, where an agent cannot reach them.
Sovereign-OS runs on your own machine and your own LLM key, and settles — when you choose to — through your own Stripe account or wallet. It is the control plane — budgets, permissions, audit — not a custodian. Your spend bills your key; your revenue lands in your account.
Self-hosted and open source. Your keys live in your own environment — nothing leaves your box. Start the governed workspace, then open it in the browser.
The whole governed control plane is MIT-licensed and yours to run — your machine, your keys, your data. A managed edition is live for teams who'd rather not operate it — bring your key, keep your data, start free.
Open source, MIT-licensed, and running in under five minutes — on your machine, on your keys. Put your AI workforce under a charter.