/ docs
Menu
Hubu

Agent spend control plane

Governed spend
for AI agents.

Hubu governs AI-agent spend. Gongbu executes only Hubu-authorized provider work. Agents get a budget, never your keys.

Start with the local stack →Explore the architectureWatch a demo
One governed flow

CONTROL PLANE

Hubu

  • Policy + budgets
  • Authorizations
  • Ledger + reconciliation
versioned executor contract

EXECUTION PLANE

Gongbu

  • Provider credentials
  • Temporal workflows
  • Artifacts + retries

Two supported modes: sandbox and live. Both start Gongbu by default. Decisions-only is a future advanced concept and is not supported. Advanced: custom executors →

Example

One image request, end to end

Illustrative amounts.

  1. 01Agent asks for an image

    The request arrives through unified MCP. The agent never holds a provider key.

  2. 02Hubu checks policy and reserves $0.05

    Policy allows it, so Hubu holds an illustrative $0.05 maximum against the budget.

  3. 03Gongbu calls the provider

    Gongbu runs only the Hubu-authorized work and reports a $0.03 receipt.

  4. 04Hubu settles $0.03 and releases $0.02

    The ledger records the actual cost and returns the rest to the budget.

Hubu's cream-and-jade seal mascot on a cinnabar base
Hubu's mascotWatch the agent create variations →

Watch the introduction

Meet Hubu in 3 min

Give agents room to work, with spending boundaries you control.

Watch on YouTube ↗

See Hubu in use: watch the sandbox demo and follow the walkthrough →

!
Experimental and local-first.

Suitable for development, evaluation, and controlled live-provider experiments—not yet for money-grade production workloads.

Read the production warning →

Works today

  • macOS source install from a release tag
  • Codex setup helper; other MCP clients by manual config
  • Gemini and FLUX.2 Pro live providers
  • Sandbox with no provider credentials
See what works today →

Managed stack experience

From install to a running stack

Follow the managed lifecycle from operator-owned configuration to a ready stack while preserving separate Hubu and Gongbu ownership.

00

Install from a release tag

Install guide →

Build the four binaries from an exact release tag on macOS.

01

Initialize a profile

hubu stack init --mode sandbox --profile "$HOME/hubu-sandbox"

Create and register operator-owned starter files without starting services.

02

Select and configure

hubu stack select --profile "$HOME/hubu-sandbox"

Select the profile for later commands, then edit stack.toml, credentials.toml, and providers.toml.

03

Start in one shot

hubu stack start

Validate, render, and start missing managed components in dependency order.

04

Check readiness

hubu stack status

Inspect the whole stack through one stable, redacted readiness view.

05

Connect your favorite agent harness

hubu init codex --stack-profile "$HOME/hubu-sandbox"

Use the Codex helper shown here, or connect another MCP-capable harness to the unified MCP process.

06

Run governed work

authorize → execute → settle, release, or reconcile

Keep policy and money state in Hubu; provider execution and artifacts in Gongbu.

Find your path

Documentation that follows the system

QUICK START

Quick start

From install to your first governed request.

Start here →GOVERNANCE

Write a policy

Human-authored rules every spend request is evaluated against.

Read the policy guide →AGENT SURFACE

Connect your agent

Codex setup helper, or configure another MCP client by hand.

Set up MCP →EXECUTION

Use your own executor

Run Hubu-authorized work on any service that implements the executor contract.

Read the executor guide →