Thanks to visit codestin.com
Credit goes to magpie.apache.org

Apache Top Level Project

Get back your time, maintainer.

Apache Magpie™ is a carefully curated set of AI agent recipes, specially designed to handle the repetitive parts of running an open-source project — triaging issues and PRs, mentoring contributors, drafting fixes, managing security reports — so maintainers can spend their time on design, relationships, and the work that needs a human. The agent proposes; the human decides.
Free & open-source
ASF governed
See it in actionOne security report, start to finish: installed from the marketplace, triaged against the project's history, fixed, released, and published as a CVE. Every message that leaves the project is drafted by the agent and sent by a person.
Add the marketplace, take the families you need
Read the whole thing as text

Install

Add the marketplace, take the families you need
$ /plugin marketplace add apache/magpie
Added marketplace apache-magpie (12 plugins)
$ /plugin install magpie-setup@apache-magpie
$ /plugin install magpie-security@apache-magpie
$ /plugin install magpie-agent-guard@apache-magpie
Installed: magpie-setup, magpie-security, magpie-agent-guard
Nothing committed — the install is yours, on this machine.

Set up

The one command worth memorising
$ /magpie-setup
Detected example-org/example-app
Tracker example-org/example-app-security · list [email protected]
Write this to .magpie-local/ (gitignored)? — answered "y"
Recommended next, and each is its own step:
  · isolate this agent      · install the action guard
  · prepare the security model

Isolate & guard

Two layers about what the agent can reach and what it can run
Magpie did that: Wrote the sandbox profile and wired the guard hook — 6 h of setup, across the team last month.
You: set up isolation and the action guard
Sandbox: Seatbelt — Bash subprocesses see only the paths you allow
Apply 3 settings changes and 1 sudo step? — answered "y"
~/.ssh, ~/.aws and your tokens are out of reach
magpie-agent-guard: PreToolUse hook now inspects every shell command
✗ denied: force-push to a protected branch. The guard refuses — no prompt to click through.

Privacy

The layer that is about other people's data, not your machine
Magpie did that: Pinned the approved models and wired redaction before any fetch — 3 h of privacy review last month.
You: set up privacy for this project
Approved-LLM gate: a private-list fetch refuses unless every model in the stack is approved
PII redaction: names and addresses → N-a3f9d2, E-7c1b04 before a model sees them
The mapping that reverses it stays on this machine and is never sent anywhere.
Record the approved model registry for this project? — answered "y"
Sandboxes stop the wrong command. This stops the right command exporting someone else's mail.

Security

Tell the security family what this project is
Magpie did that: Drafted the security model and wired the tracker conventions — 5 h of policy-wrangling last month.
You: prepare the security model for this project
Read the code, the docs and 2 years of advisories
Drafted: trust boundaries · what is in scope · what is not
Open a PR adding SECURITY.md and the tracker conventions? — answered "y"
A written model is what lets the next report be routed without a meeting.

Import

Ask for exactly the window you care about
Magpie did that: Read the list and opened both issues — 6 h of copy-paste last month.
You: sync the security issues that arrived in the last 2 days
[email protected] — 2 unimported reports
  1. "path traversal in archive extractor"  — [email protected]
  2. "login page is slow"                   — [email protected]
Open 2 issues in example-org/example-app-security? — answered "y"
Opened #41 and #42. Identifiers are public; contents stay private.

Triage

Classify, check the history, agree a disposition
Magpie did that: Checked 148 past issues for duplicates — 9 h of searching last month.
You: triage the two new reports
#41 archive extractor — path traversal on member names
Severity: high · CWE-22 · affects 3.0.0 – 3.2.0
Related to #12 (fixed 2024) — different code path, not a duplicate.
Team thread: 3 replies, disposition agreed — accept, fix in 3.2.1
#42 login page is slow — no security impact
Close #42 as not-a-vulnerability, with a reply? — answered "y"
Reply to the reporter — written in an instant, sent by a human
Magpie did that: Drafted the reply, with the history it just checked — 4 h of writing last month.
To: [email protected]
From: [email protected]
Subject: Re: path traversal in archive extractor
Hello,

Thank you for the report. We have confirmed the issue and are
tracking it privately; it is not related to the 2024 report you
may have seen, which touched a different code path.

We will request a CVE and let you know before the advisory is
published, so you can check the credit line reads the way you
want it to.

Please keep the details private until then.

-- the Example App security team
Plain text, real links, no tracking. Magpie wrote it; you send it.

Board

The same issue, on the tracker board
Columns: Reported → Triaged → Fix in progress → Fix ready → Released
#41 Path traversal in archive extractor — Reported to Fix in progress
#42 Login page is slow — closed, not a vulnerability — Reported to Reported
Triaged — severity high, assigned
Picked up for a fix
…and the fix opens itself
Opened the fix PR on a private branch, with the test that proves it — 7 h of fix-shepherding last month.
Pull request example-org/example-app#8317: Normalise archive member paths before extraction
+ 24  −3   src/extract.ts
+ 61  −0   test/extract_traversal_test.ts
private branch · linked to #41 · awaiting review
Reviewed, merged, released
Columns: Reported → Triaged → Fix in progress → Fix ready → Released
#41 Path traversal in archive extractor — Fix in progress to Released
#42 Login page is slow — closed, not a vulnerability — Reported to Reported
Reviewed and merged
Example App 3.2.1 released

Learn

The model learns from what you just decided
Magpie did that: Turned two dispositions into a sharper security model — 4 h of model upkeep nobody ever gets to last month.
You: update the security model with what we learned from #41 and #42
Reading both dispositions and the fix that landed
#41 accepted → archive extraction is now a named trust boundary
#42 rejected → page latency is stated out of scope, so the next one routes itself
Open a PR updating SECURITY.md with both? — answered "y"
Every report makes the model sharper instead of leaving it staler.

Publish

The CVE record
Magpie did that: Filled the CVE record from the issue and the fix — 3 h of form-filling last month.
CVE ID: CVE-2026-XXXXX
Product: Example App (example-org)
Affected: 3.0.0 – 3.2.0
Fixed in: 3.2.1
CWE: CWE-22 — path traversal
CVSS: 7.5 (high)
Credit: R. Chen
References: example-org/example-app#8317
Submit the record to the CVE program?
Announce it — written in an instant, sent by a human
Magpie did that: Drafted the advisory from the record and the fix — 3 h of writing last month.
To: [email protected]
From: [email protected]
Subject: [CVE-2026-XXXXX] Path traversal in Example App archive extractor
CVE-2026-XXXXX: path traversal in the Example App archive extractor

Severity: high (CVSS 7.5, CWE-22)
Affected:  3.0.0 through 3.2.0
Fixed in:  3.2.1

An archive member name was joined to the extraction directory
without normalisation, so a crafted archive could write outside
it. Upgrade to 3.2.1. There is no workaround.

Credit: R. Chen, who reported it privately and followed the
embargo throughout.

https://example.org/security/CVE-2026-XXXXX
The last gate in the story — and a person is still the one pressing it.
Closed — and off the board
Magpie did that: Closed #41 — advisory out, nothing left to chase — 31 h across 12 reports last month.
Columns: Reported → Triaged → Fix in progress → Fix ready → Released
#41 Path traversal in archive extractor — closed — Released to Released
#42 Login page is slow — closed, not a vulnerability — Reported to Reported
CVE-2026-XXXXX published — closing #41
Board clear. Nothing left open.
Published
ID: CVE-2026-XXXXX
Published: 2026-09-13
Severity: 7.5 HIGH
Weakness: CWE-22
Status: Analyzed
One report, one fix, one release, one advisory — and every message that left the project was written by Magpie and sent by a person.
How the security family actually works

Example App is a fictional project, and the CVE identifier is a placeholder. The workflow is real; the vulnerability is not.

Progressive Agentic ModesAgentic automations that adapt to your project. Start simple — each mode builds on the previous one, evolving as trust grows.
1
Agentic Triage Mode
Stable
Auto-label, categorize, and route incoming PRs and issues to the right reviewers instantly.
Auto-labeling
Priority scoring
Docs
2
Agentic Drafting Mode
Stable (security)
Draft a fix for a well-scoped problem and open a PR — every PR reviewed and merged by a human committer, never the agent.
Fix PRs
Human-merged
Docs
3
Agentic Mentoring Mode
Experimental
Joins issue and PR threads in a teaching register — clarifying questions, pointers to conventions, examples from prior PRs.
Conventions
Prior PRs
Docs
4
Agentic Pairing Mode
Experimental
Developer-side skills you run in your own dev loop — multi-agent review pipelines and self-review before you open a PR.
Self-review
Pre-flight
Docs
5
Agentic Autonomous Mode
Off
Limited fix-and-merge for objectively boring changes — lint, allow-listed dependency bumps, formatting. Deliberately off until the earlier modes have proven out, and never for security-class changes.
Auto-approve
Audit trail
Rollback
Docs
Pick the skill families you needModes describe how agents are used; skill families are how those capabilities ship. Adopt only the families that fit — symlinks for the picked families land in your skill directory.Most families work on any project, inside or outside the Apache Software Foundation. The ones marked with the Apache oak leaf () are ASF-specific — they encode Foundation processes like the release lifecycle and the contributor-to-committer path. We'd love to support other maintainership and governance models too; contributions are welcome.
10 skills
setup
Infrastructure
Isolated agent setup, framework adoption & maintenance, shared-config sync. The prerequisite every adopter starts from.Read the overview
15 skills
security
Triage · Drafting
The 16-step security-issue lifecycle — from security@ import through CVE allocation and publication, with state sync. Maintainer-only.Read the overview
8 skills
pr-management
Triage
Maintainer-facing PR-queue management — triage, queue stats, and deep code review.Read the overview
8 skills
issue
Triage · Drafting
Issue lifecycle — triage, bug reproduction, fix drafting, and backlog re-assessment against the current branch.Read the overview
7 skills · experimental
repo-health
Triage
Read-only maintenance audits — CI runner obsolescence, workflow security, stale or vulnerable dependencies, license/NOTICE drift, and flaky tests. Each proposes remedies for the maintainer to apply.Read the overview
10 skills
release-management
Triage · Drafting
The 14-step ASF release lifecycle — planning, RC cut & sign, [VOTE], tally, promote, [ANNOUNCE], archive, audit. The agent never holds the signing key or publishes.Read the overview
4 skills · experimental
mentoring
Mentoring
Contributor mentoring — spec and tone guide in place; first skill (pr-management-mentor) shipping.Read the overview
6 skills
contributor-growth
Mentoring · Triage
The contributor-to-committer path — welcome first-timers, keep the backlog newcomer-ready, track activity, assemble nomination evidence, and run post-vote onboarding.Read the overview
2 skills · experimental
pairing
Pairing
Developer-side pre-flight review that runs in your own dev loop — self-review or three independent axis-focused passes (correctness, security, conventions). Read-only, opens no PR, mutates nothing.Read the overview
5 skills
utilities
Meta
Framework meta-skills — author or update skills (write-skill) and print a live index of every skill (list-skills).Read the overview
New to agents? Learn hereBuilding software with an AI agent is a new skill, even for people who have written code for years. It isn't harder than other coding — it's different: behaviour is probabilistic, prompts and skills are code, and you test with evals rather than single checks.Magpie's maintainer-education stream teaches you that craft one page at a time — no AI background needed. Work through the three groups in order, or jump to the one you need; two hands-on references sit alongside them all. Every release ships the learning material for the skills it contains (PRINCIPLE 18).
Vendor NeutralityMagpie's workflows never name a vendor. Skills declare a capability; a tool fulfils it; every vendor binding lives in a tool behind a contract — across six independent axes: LLM backend, agentic runtime, forge/tracker, communication channels, source control, and project governance. Swapping a backend is a config change, never a rewrite of the workflows.
10/10Capability contracts100% work across 2+ vendors
75/75Skillscarry no vendor lock-in
24/24Agent harness100% harness-neutral
Live figures from the framework repo — every axis already works across more than one vendor or is neutral by construction; we keep adding backends and organisations so the numbers stay honest.
Swap LLM providers freelyAnthropic, OpenAI, Bedrock, or local Ollama/vLLM — skills target a capability floor, not a provider SDK. No lock-in.
Works across agent CLIs & ecosystemsSkills follow the open AGENTS.md standard, so they run across agent CLIs and ecosystems — Claude Code, Codex, and others — rather than being tied to one vendor's tool.
Forge, tracker & VCS behind contractsGitHub and Jira today; GitLab, Gitea, Bugzilla, and non-Git VCS (SVN, Mercurial, …) are documented, labelled extension points — each one adapter, not a fork.
All organisations are welcomeMagpie isn't hardwired to the Apache Software Foundation. A governing body — a foundation, a company, or an informal maintainer collective — is an organization that bundles its governance vocabulary, backend defaults, and identity (logo included). A project declares organization: ASF (or independent, or your own) once and inherits the rest — so the same skill runs unchanged for an ASF project and a non-ASF one.
Add your organizationRunning Magpie under a different foundation or company? Propose an organization bundle so every project under it inherits your defaults — start the conversation on the developer mailing list.Email the dev list
Built in the open, the Apache WayApache Magpie is a project of The Apache Software Foundation, built in public and run by its community following the Apache Way — consensus-driven, vendor-neutral, and earned on merit.
Decisions happen on open channels, and everyone is welcome to take part.
Day-to-day work happens on GitHub (issues & pull requests) and in Slack for quick questions — but every significant discussion and decision also lands on the developer mailing list. Prefer not to use GitHub or Slack? You can follow along and take part fully over email alone: the dev list is the one channel that reaches everyone.
Developer mailing listDesign discussion, proposals, and the formal [VOTE]s that reach consensus all happen on [email protected]. If it didn't happen on the list, it didn't happen.
Slack — #magpieReal-time chat with maintainers and contributors in the #magpie channel on the ASF Slack — questions, pairing, and quick design back-and-forth.
Start contributingNew here? The contributing guide walks through setup, conventions, and your first pull request. The agent proposes; humans review and merge — and we mentor newcomers along the way.
The people behind Apache Magpie™25 committers & PMC members steering the project.
Amogh DesaiAndré AhlertAndrew MusselmanArnavCalvin KirsCoty SutherlandCraig L RussellElad KalifEvan RusackasGPKIsmaël MejíaJames FredleyJarek PotiukJB OnofréJustin McleanMatt TopolMike DrobPaul KingPiotr P. KarwaszRémy MaucheratRich BowenRichard ZowallaRussell SpitzertisonYeonguk Choo
Hey maintainer …

Want to get your time back?
Join the growing community of open-source projects using Apache Magpie to scale their review processes without burning out their maintainers.
Suggest a change