Hubu / 户部#

Hubu is an open-source spending control plane for AI agents.

It gives agents permission to use paid services without giving them direct control of payment keys, provider credentials, or unrestricted budgets.

Project status#

Experimental and local-first

Hubu provides a complete managed local stack for governed agent workloads, including experimental third-party provider integrations that can incur real charges. Use live mode cautiously, with reviewed targets, conservative policies, and strict budgets. No production direct payment rail is supported yet. The built-in direct payment path remains a mock, and Hubu is not approved as money-grade production financial infrastructure.

What works today#

Advanced integrations#

Custom executors can implement the versioned executor contract. This is an advanced integration point, rather than a starting stack mode. Decisions-only is a future advanced concept and is not currently supported. Both supported modes retain reservation-and-settlement accounting through Gongbu.

Why Hubu and Gongbu?#

The names come from two ministries in imperial China.

Hubu (户部) was the Ministry of Revenue. It managed state finances, taxation, budgets, and household records.

Gongbu (工部) was the Ministry of Works. It organized construction, infrastructure, equipment, and the execution of public projects.

That historical division captures the spirit of the system:

Hubu governs resources. Gongbu performs the work.

Hubu decides whether work may consume money. Gongbu carries out the authorized work with the provider. The two cooperate, but retain separate responsibilities, credentials, storage, and failure boundaries.

Explore the interactive architecture →

The core idea#

Agents can have budgets, but they should not hold private keys.

A human establishes the operating boundary: who the agent is, what it may purchase, how much it may spend, and when human approval is required.

When an agent requests paid work, Hubu evaluates the request against that boundary. An allowed request reserves budget before execution begins. A denied request stops without consuming funds. A request that needs approval waits for an explicit human decision.

Gongbu receives authorized work, communicates with the configured provider, manages retries and artifacts, and reports the outcome. Hubu then settles the actual cost, releases unused budget, or preserves the operation for reconciliation when the provider's billing result is uncertain.

One governed workload#

The complete idea can be followed as one continuous lifecycle:

Agent request
    ↓
Identity, policy, and budget evaluation
    ↓
Spend authorization and budget reservation
    ↓
Provider execution through Gongbu
    ↓
Settlement, release, or reconciliation

Every logical operation has a stable identity so retries recover the same financial state instead of creating accidental duplicate spend. Decisions and successful money movement are recorded for later inspection.

The result is a separation of authority:

• The agent proposes work.
• Hubu governs whether resources may be used.
• Gongbu performs the authorized work.
• Humans retain policy and approval authority.

One agent-facing surface#

Agents interact with Hubu and Gongbu through the unified MCP server.

The unified surface makes both governance and execution tools discoverable in one place, while preserving the two backend boundaries. It routes each request to its actual owner rather than combining financial authority and provider execution inside one process.

Codex is one supported MCP client, but the architecture is designed around the protocol rather than a single agent harness.

Start exploring#

• Why Hubu, compared with other approaches →
• Frequently asked questions →
• Start the local Hubu stack →
• Operate supported live providers →
• Understand policy, budgets, and spend lifecycle →
• See how Gongbu executes authorized work →
• Use Hubu with your own executor →
• Integrate through unified MCP →

See the roadmap for what comes next.

For operational procedures, protocol details, and troubleshooting, continue through the documentation navigation.