Skip to main content

Autonomous AI coding pipeline

Chat triggers the work. Magpie does the rest. From task description to pull request, fully automated with AI agents, CI validation, and intelligent blueprint execution.

@magpie fix login bug
Task classified: BugFix
Running diagnostic blueprint
Tests passing • Lint clean
PR opened: #247

How it works

Magpie orchestrates a complete autonomous coding pipeline from a single chat message

1

Trigger from chat

Mention @magpie in Discord, Teams, or run a CLI command with your task description. Magpie reads the full thread context to understand requirements.
2

Intelligent classification

The pipeline analyzes your task and classifies it as Simple (docs/typos), Standard (features), or BugFix (fixes). Each complexity level triggers a different blueprint optimized for that work.
Simple single-shot agent call
Standard TDD: plan tests implement
BugFix Diagnostic: investigate regression test fix
3

Agent execution with blueprints

Magpie uses a two-tier LLM architecture: Claude CLI for text generation (branch names, commit messages) and Goose agent for coding work. The blueprint engine orchestrates deterministic Shell steps and AI Agent steps in sequence.
scan-repo → plan → write-tests → verify-tests-fail 
→ implement → run-tests → lint-check
4

CI loop with auto-fix

After the agent completes coding, Magpie runs your configured lint and test commands. If CI fails, the fix blueprint receives the error output and attempts repairs (up to 2 rounds by default).
// Configured via environment variables
MAGPIE_LINT_CMD="cargo clippy"
MAGPIE_TEST_CMD="cargo test"
MAGPIE_MAX_CI_ROUNDS="2"
5

Commit, push, and PR

Once tests pass, Magpie generates a commit message, pushes to a new branch with collision handling, and opens a pull request via gh CLI. The result is posted back to your chat thread.
Magpie integrates with Plane for issue tracking and Daytona for sandboxed execution. All agent calls are traced to JSONL for observability.

Key features

Built for production autonomous coding workflows

Blueprint engine

Hybrid orchestration with deterministic Shell steps and AI Agent steps. Three built-in blueprints: Simple, TDD, and Diagnostic.

Two-tier LLM

Tier 1 (Claude CLI) for clean text generation. Tier 2 (Goose agent) for multi-turn coding with full tool access.

Multi-platform adapters

Discord bot, Teams webhook, and CLI. All implement the ChatPlatform trait for consistent integration.

Sandbox abstraction

Local execution or remote Daytona sandboxes. Swap implementations via trait without pipeline changes.

Org-scoped repos

Dynamic repo resolution from task messages. One Magpie instance works across multiple repositories in your GitHub org.

CI classification

Docs-only changes skip CI. Built-in test steps in blueprints eliminate redundant CI runs.

Observability

JSONL traces for every agent call. Verbose mode prints streaming events. Debug and audit all LLM interactions.

Git automation

Branch collision handling, conventional commit generation, and PR automation via GitHub CLI.

Architecture at a glance

Magpie is a Rust workspace with four crates

magpie-core

Core library. Pipeline orchestrator, blueprint engine, agent wrapper, git ops, Plane client, repo resolution, tracing. All other crates depend on this.

magpie-cli

CLI binary for running prompts, blueprints, or the full pipeline locally. Thin wrapper around magpie-core.

magpie-discord

Discord bot adapter (Serenity). Triggers pipeline when @magpie is mentioned. Creates a thread per task, archives on completion.

magpie-teams

Microsoft Teams webhook adapter (Axum). Receives Bot Framework events and triggers the pipeline.

Ready to automate your coding workflow?

Get started with Magpie in minutes. Install the CLI, configure your environment, and start issuing tasks from Discord, Teams, or the command line.

Get Started Now