Bello plugin
A harness around a coding agent: supervisor, reviewer, adversary and local log distiller, each on its own model
- 226 ★ on the repository
- Last commit
Broad access
Why this level
- The coder runs commands and edits files, and the supervisor only partly judges risky commands
- Subscription logins and API keys give runs access to your limits and billing
- The log distiller downloads a 599 MB model and a Codex helper on first run
Install
- Works in
Manual install
pipx install belloInstalls the CLI. Then bello doctor, a model-source login and the plugin: codex plugin marketplace add AlexeyKulaev/Bello-codex-marketplace --ref main.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The coder runs commands and edits files, and the supervisor only partly judges risky commands
- Subscription logins and API keys give runs access to your limits and billing
- The log distiller downloads a 599 MB model and a Codex helper on first run
This is third-party code. Review the repository files before installing.
What it does
Bello runs a task from task.md through a coder and four optional roles. The coder works in a disposable workspace and runs its own checks, then the finished patch goes back to the project. A supervisor follows the run, judges risky commands before they execute and restarts a failing run without losing the workspace. A reviewer with a fresh context compares the result with the task, while an adversary receives the outcome without the development history and tries to break it with edge cases and invalid input. Each role runs on its own model with its own reasoning effort: a Codex or Claude Code subscription, or an OpenAI, Anthropic or OpenRouter API. The local log distiller is a fine-tuned ModernBERT model that shortens tool output before the coder reads it. An advisor in the Codex and Claude Code plugin reads the task and the repository and recommends a configuration for the priority you name; the run itself goes through the bello command.
Who it is for. Developers who run long tasks through Codex or Claude Code and want to cut limit usage without losing quality.
Good fit when
- The task is written in task.md and you want a reviewer and an adversary to check the result separately
- The weekly Codex limit drains too fast and some roles should go to cheaper models
- You want to mix models from different vendors in one chain: coder on one, reviewer on another
- Tool output bloats the context and you want to compress it locally
Not a fit when
- You need a quick small fix: a multi-role run costs time and model calls
- You have no Codex or Claude Code subscription and no API key: roles have nothing to run on
- You want to run on Windows without sandbox preparation: it needs the one-time setup from docs/windows.md
Example request
Recommend the Bello configuration with the best price-quality balance for task.md, then run the taskLimitations
Requires Python 3.11 or newer and git; on Linux also the bubblewrap package, and on Windows 11 or Server 2022/2025 a one-time sandbox setup. API providers need Node.js 22.19 or newer and bello runtime install. The distiller model weighs 599 MB and downloads on first run. The savings figures in the README come from the author's tests and will differ on your tasks.
How to disable. Remove the plugin from Codex or Claude Code and uninstall the package with pipx uninstall bello. Workspaces and reports live inside the project, for example .supervisor/FINAL_REPORT.md, and can be deleted by hand.
FAQ
Which models can it use?
Codex and Claude Code subscriptions through bello runtime login, plus APIs from OpenAI, Anthropic, OpenRouter and other Pi providers. Roles are configured independently and any can be switched off.
What does Goal mode do?
A mode for long goal-driven tasks. By the author's data, in NetHack Bello in Goal mode set a best score for the lawful human female Valkyrie, 42.18% higher than Codex Goal on the same model.
How do I check that everything is ready?
bello doctor shows which model sources and components are set up. bello update upgrades it and every setting is in bello config.
Where is the run result?
In .supervisor/FINAL_REPORT.md in the project: the result, changed files, checks and remaining issues. The patch returns to the project after the coder works in a disposable copy.
Is Claude Code required?
No, a Codex subscription or an API key is enough. For a Claude Code subscription install bello[claude] and log in with bello runtime login claude-code; the extra bundles the official Agent SDK.
Discussion
No questions or reviews yet.
Related
Pluginsmulti-agents runs a team of coding agents on a task and measures whether it beat a single agent and what it cost.
Argus runs long engineering and research tasks with four roles, where a reviewer reads results but edits nothing.
Team Mode splits a Codex task between an explorer, an executor and a reviewer while the main agent keeps final acceptance
A terminal app that builds a team of Codex, Claude and Ollama agents and runs tasks in parallel under a lead agent
A bridge between local CLI agents: Codex, Cursor or Claude Code delegate work to Antigravity, Grok, Kimi and seven more.
Viden is a terminal cockpit for a coding agent: chat with a model, approve edits and supervise Codex and Claude in one window.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Bello (this card) | no | 226 | ||
| Ralph Orchestrator | no | not required | 3,203 | runs an agent loop against a plan, while Bello has supervisor, reviewer and adversary roles |
| CLI Agent Orchestrator | no | 1,548 | orchestrates several CLI agents, while Bello backs one coder with reviewing roles |
README in short
The README opens with a short summary, then covers install, a quick start with an agent dialogue, the roles and benchmark results: NetHack, four models with and without Bello, the efficient budget mode and the log distiller. It ends with configuration and the MIT license.