Premium Model Budget Governor

Astra does the work. You see the budget.

A local workbench, CLI, and MCP server for planning premium-model work, approving estimated spend, and inspecting usage afterward.

Explore the product tour   /   Install the beta   /   Connect MCP   /   Become a tester

Real development capture in a product-tour layout. Playback is optional on the website; no model runs in the tour or planner demo.

Public beta, not a stable release. The downloadable version is 0.4.0rc6. Execution is currently read-only and requires your explicit approval. This does not change the active model in existing Codex chats, make tokens cheaper, or guarantee the same quality at lower cost.

New Evidence, Same Astra

In a 12-call context pilot, both Astra profiles passed all frozen checks on three authored tasks. The focused catalog used 18.1% fewer input tokens and 26.8% fewer estimated credits, with unequal cache hits contributing to the credit difference. This is a small pilot, not a savings promise or proof of equal capability on every task.

A four-call extension checked read-only tool use and access to one named local skill. See the validation gate register for what passed and what still needs independent people and broader tasks.

Built for: Codex users who want Astra participation with explicit task budgets, and developers connecting governed tools through MCP. Not a fit yet: automatic control of every existing chat, write-enabled execution, or guaranteed weekly savings.

Idea, research guidance, and product management: Sulabh Dubey. Research synthesis and build execution: Codex by OpenAI, under his direction.

One weekly Codex allowance was consumed in roughly a day of Astra-heavy work. After a reset, only about 16% of the second allowance remained the following day. These were Sulabh Dubey's observed account readings, not a controlled benchmark.

The goal was not to stop using Astra. It was to keep its capabilities available without spending unnecessarily on repeated context, oversized evidence, and extra model handoffs.

That distinction shapes this project: direct Astra is the default. A cheaper model is an option, not the product's answer to every task. Hybrid workflows must account for all their stages, not just their cheapest step.

Idea, research guidance, and product management: Sulabh Dubey. Research synthesis, design, engineering, testing, documentation, and release execution: Codex by OpenAI, under his direction. Origin and credits.

What You Can Do

One Task, One Visible Decision

1. Choose a project and describe a small task. Add the relevant files or images.

2. Preview a direct Astra run, or explicitly select a supported multi-stage workflow.

3. Approve the proposed read-only execution after checking the estimate.

4. Inspect the result and usage receipt. Compare the complete cost, not a capsule alone.

The Workbench supports project management, evidence selection, stop/reconnect, and receipt inspection without writing JSON. Initial installation still uses a terminal. Human usability testing is open; we are not claiming anyone can install it effortlessly yet. Workbench guide.

Related MCP server: l6e-mcp

Install

You need: Python 3.10+ and, for real execution, an installed and authenticated Codex CLI with access to the requested model. Model calls use your account's capacity. The web demo is illustrative and does not run models.

1. Download The Beta

Get these files from the rc.6 release:

• Installer

• Python wheel

• SHA-256 checksums

Keep the installer and wheel in the same folder. Check the downloaded files against the release checksums before installing; only install code you trust.

2. Preview, Then Install

Open a terminal in that folder. The first command previews; the second installs:

python install_governor.py install --wheel premium_model_budget_governor-0.4.0rc6-py3-none-any.whl
python install_governor.py install --wheel premium_model_budget_governor-0.4.0rc6-py3-none-any.whl --yes

This creates a dedicated environment at ~/.pm-bg/runtime. It does not change your Codex settings, PATH, or other Python environments, and does not run a model. If that folder already exists, follow the upgrade/removal guidance instead of overwriting it. Keep projects and ledgers outside the runtime folder.

3. Check Setup And Open The Workbench

Windows PowerShell (replace the example project folder):

& "$HOME/.pm-bg/runtime/Scripts/pm-bg.exe" doctor
& "$HOME/.pm-bg/runtime/Scripts/pm-bg.exe" serve --project "C:\path\to\project"

Linux / macOS shell (replace the example project folder):

"$HOME/.pm-bg/runtime/bin/pm-bg" doctor
"$HOME/.pm-bg/runtime/bin/pm-bg" serve --project "/absolute/path/to/project"

The server opens a private local browser link. Do not share that session link. doctor checks setup without starting a model call; it does not prove model access or remaining capacity. The released wheel has local Windows and Ubuntu/WSL qualification; native macOS interactive usability is not established.

Need help? Installation and recovery / FAQ

Use a dedicated environment for development:

git clone --branch v0.4.0-rc.6 https://github.com/sulabhdubey/premium-model-budget-governor.git
cd premium-model-budget-governor
python -m venv .venv

Activate it with .venv\Scripts\Activate.ps1 on Windows PowerShell, or source .venv/bin/activate on Linux/macOS, then:

python -m pip install -e ".[dev]"
pm-bg plan --input examples/astra_preferred.json
python -m pytest

The example is illustrative: it plans but does not execute a model. Replace its estimates and approval state with real task inputs before using it for a decision. See the planning contract and CLI quickstart.

Use It With Codex Or MCP

The Workbench is one entry point. Compatible agents can also call local tools such as plan_model_workflow, manage_task_budget, and build_capsule_from_files.

Adding MCP exposes tools; it does not force an agent to use them or automatically govern every existing chat. The integration guide uses an explicit runtime path so the client does not accidentally launch a different Python installation.

What The Tests Actually Show

We publish the cases that did not save money, too.

Read the evidence: Release qualification / Real execution receipt / Pilot, methods, and negative findings / Focused-catalog tradeoffs.

Token-rate-estimated credits are not provider billing records or weekly-limit percentages. Independent held-out task evaluation and technical/nontechnical onboarding sessions remain open. We will not convert these early observations into a blanket "same Astra quality for less" claim.

Controls, Not Magic

• Whole-task budgets: persistent reservations, expiring leases, and replay protection for cooperating runners.

• Evidence preparation: explicit file selection, capsule scoring, and graph summaries; required evidence must not silently disappear.

• Usage recovery: reconcile a recorded terminal receipt without rerunning a paid task; missing usage remains unknown.

• Reviewed preferences: inspect evidence-based proposals, activate manually, and roll back. No automatic policy promotion is claimed.

• Publication checks: bounded pattern scanning for secrets and private terms, with hash-bound reports and manual-review states.

See capability controls, architecture, and publication privacy for the contracts.

Know The Boundaries

• Read-only execution does not revoke inherited connector permissions.

• Estimates and local admission checks are not a provider-enforced in-flight spending cap.

• Secret and prompt-injection scans are supplementary checks, not antivirus or a security guarantee.

• Desktop-only actions and unsupported capabilities must be reported, not silently replaced.

• Smaller context can remove useful guidance; focused discovery stays opt-in.

• Nothing here bypasses provider limits or guarantees identical output quality.

Report sensitive findings through SECURITY.md, not a public issue. Cloned before the privacy cleanup? Read history migration before contributing; do not merge the old history back.

Help Test The Beta

We are looking for technical and nontechnical testers, and creators who want to evaluate it independently. No endorsement or positive result is expected.

1. Try installation and one small, non-sensitive, read-only task.

2. Record where you got stuck, what worked, and whether the answer was useful.

3. Submit a beta trial report.

Never post credentials, private project files, prompts, or local session links. Review any attachments before sharing. A failed installation or a more expensive workflow is useful feedback, not a result to hide.

For a cost comparison, use the independent validation protocol: freeze the task and quality criteria first, assess answers before revealing costs, and count every stage, failure, and retry. Model calls consume your own capacity; no paid comparison is required just to report onboarding feedback.

Onboarding protocol / Contributing / Creator testing brief

Free, Open Source, Independently Built

The published code is available under Apache-2.0. There is no paid unlock required for the features in this repository; your model-provider usage remains separate. Possible future paid additions are a roadmap discussion, not a currently shipping Pro product.

This is an independent project, not an OpenAI product or endorsed by OpenAI.

Try the beta. Share what happened. Help us measure where it really helps.