skill
magenta-agents
Role-based AI agents for Claude Code: the rules become checks, and a task is done when the checks pass
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/devlnull/magenta-agents itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Nov 8, 2026Magenta gives your coding agent an engineering team, and a definition of "done" that a machine checks.
You type one command: /magenta <task>. Magenta reads the rules of your repository, turns your success criteria into checks (commands that a machine runs), and works until the checks pass. You approve the checks. Small tasks run fast in one agent. Risky tasks get the roles that they need: QA, Architect, Security, SRE, Database, and others.
It works best with Claude Code. OpenCode, Codex CLI, and Gemini CLI are also supported.
Benchmark
Measured on Magenta 0.16.1 (2026-11-04): 18 coding tasks with hidden tests, 4 Claude models, 3 runs of each task. The same model does the same task with a plain prompt and with /magenta.
| Plain prompt | Magenta | |
|---|---|---|
| Simple tasks | 99.0% | 98.6% |
| Complex tasks (a spec) | 99.0% | 97.2% |
| Hard tasks (hidden rules) | 87.9% | 97.6% |
| Held-out tasks | 93.7% | 91.4% |
| All tasks | 94.1% | 96.4% |
| Old tests still pass | 210 of 216 | 214 of 216 |
| Cost for each task | $0.114 | $0.694 |
Where Magenta helps: tasks where the repository has rules that the request does not say (hard tasks: 88% → 98%), and smaller models. Where it does not: small, clear tasks with a strong model. There, a plain prompt is as good and cheaper.
Best value: Sonnet 5.5 with Magenta passed 99% of the hidden tests (52 of 54 runs with all hidden tests passed) at $0.44 for each task.
The tasks, the method, the limits, and the raw results: evals/benchmark.
When to use Magenta
| Use Magenta when | A plain prompt is enough when |
|---|---|
| The repository has rules that the request does not say: security files, rounding rules, API conventions, stored formats | The change is small and clear, and you use a strong model |
| The change touches money, security, personal data, or a public API | You want the fastest and the cheapest answer |
| You want evidence: approved checks, tests that stay in the repository, and a record of each task | You will review each line yourself |
| You use a smaller model and want results near the results of a larger model |
Why Magenta
- Evidence, not claims. A task is done when the approved checks pass on the current code. In Claude Code, a hook does not let the agent stop before that.
- You own the definition of done. Only you approve a check. If an agent changes an approved check or a locked test, the check fails.
- It reads your rules first. A script finds the rule files of the repository and the other callers of the changed code before the work starts, with no model tokens.
- The right team for the task. Small, low-risk tasks run in one agent (direct mode). Other tasks get a lead and only the reviewers that the risks need. High-risk work gets the full team.
- Guardrails. The agents change the code freely, but not the checks, the approvals, the locked tests, or the secrets. Dangerous commands are blocked: no agent pushes, merges, or deploys.
- Auto mode.
/magenta --auto <task>runs to the end. You see only the questions that only you can answer, and the final result. - Measured at each release. Each version runs the benchmark before it is tagged.
How it works
flowchart LR
T["/magenta + your task"] --> S["Rule scan<br/>rule files, callers,<br/>risky areas"]
S --> M{"Select<br/>the mode"}
M -->|"small, low risk"| D["Direct<br/>one agent, test first,<br/>project checks"]
M -->|"most tasks"| C["Checks first<br/>you approve them"]
M -->|"high risk"| TM["Team<br/>all roles that<br/>the risks need"]
C --> L["Lead agent works<br/>until the checks pass"]
TM --> L
L --> R["Reviewers"]
R -->|"a finding becomes a check"| L
D --> F["Report with<br/>the evidence"]
R --> F
Quick start
Requirements: Python 3.9 or later, git, and one of the supported tools. On Windows, Claude Code needs Git for Windows.
1. Install. For Claude Code, use the plugin:
claude plugin marketplace add devlnull/magenta-agents
claude plugin install magenta@magenta-marketplace
Or use the installer (Claude Code at user level, OpenCode, Codex CLI, or Gemini CLI):
curl -fsSL https://raw.githubusercontent.com/devlnull/magenta-agents/main/install.sh | sh
On Windows, use PowerShell: irm https://raw.githubusercontent.com/devlnull/magenta-agents/main/install.ps1 | iex. The guide shows the other options.
2. Prepare your repository (one time):
python3 ~/.magenta-agents/magenta.py init /path/to/your/repository
With the plugin, add --no-hook: the plugin has its own hooks.
3. Run a task in Claude Code, in the root folder of the repository:
/magenta --auto add a status field (unpaid or paid) to the saved orders
/magenta find the performance issues of GET /v2/addresses/default. The average is 10 ms now.
I need p95 at or below 5 ms; p99 can be 10 ms.
To see the task and the checks at any time, type /magenta:task-status. To run the checks now, type /magenta:run-checks.
The agents
| Agent | Role |
|---|---|
| Product Manager | Reviews PRDs for value, evidence, and scope. Writes user stories. |
| Software Architect | Designs the structure, compares options, writes RFCs and ADRs, and writes the structural rules as checks. |
| Senior Principal Engineer | Plans, settles disagreements, checks production readiness, and does the final review. |
| QA Engineer | Writes the checks and the tests that define "done", finds edge cases, and gives the quality verdict. |
| Software Engineer | Implements server-side changes with TDD, in the patterns of your project. |
| Front-End Engineer | Implements client-side changes with performance, accessibility, and locale support. |
| Site Reliability Engineer | Defines SLOs and alerts, reviews resilience and rollout, leads incidents, and writes runbooks and postmortems. |
| Platform Engineer (DevOps) | Changes CI/CD pipelines, containers, Kubernetes and Helm, and Terraform. Never applies or deploys. |
| Security Engineer | Makes threat models, reviews code for security, triages CVEs, and keeps the scanners as checks. |
| Database Engineer | Plans zero-downtime migrations, tunes queries with evidence, and writes migration and query checks. |
| Technical Writer | Keeps the documentation, the changelog, and the release notes in agreement with the code. |
Most tasks use only a few of them. The team page shows when each role joins and which files it owns.
Documentation
| Document | Content |
|---|---|
| User guide | Install options, modes, task types, gates, checks, and daily use |
| Reference | Commands, configuration files, enforcement, and model profiles |
| The team | The roles, when each one joins a task, and the files that each one owns |
| Team guide | A 15-minute introduction for an engineer in a team |
| Benchmark | The tasks, the method, the results of each version, and how to run it |
| Architecture | How the source becomes the files of each tool |
| Design (RFC-0001) and decisions (ADRs) | The full design and the decision records |
| Changelog | The changes in each version |
Status
Magenta is new (version 0.x). All features are implemented and tested by CI on macOS, Linux, and Windows. The commands and the file formats can change before 1.0. The changelog records each change, and the benchmark records the results of each version.
Contributing
Contributions are welcome: new benchmark tasks, new roles, support for more tools, and bug reports. Read CONTRIBUTING.md first. To report a security problem, follow SECURITY.md.
If Magenta helps you, give the repository a star. It helps other engineers find it.