apra-fleet

Product Introduction

1. Definition: Apra-fleet is an open-source, Apache-2.0 licensed MCP (Model Context Protocol) server and CLI tool. Technically, it is a multi-agent orchestration and operations platform designed to schedule, credentialize, isolate, and observe AI agents across a heterogeneous fleet of machines.
2. Core Value Proposition: It exists to solve the operational complexity of scaling from a single AI agent demo to a production-ready "agentic workforce." Its primary value is enabling developers to run durable, multi-hour AI workflows across their own hardware (local machines, VMs, GPU boxes) while seamlessly mixing multiple LLM providers (Claude, Gemini, Codex, Copilot, Antigravity, local models) based on cost and capability.

Main Features

1. Multi-Provider, Multi-Machine Fleet Management: Apra-fleet's fundamental unit is a "member"—a pairing of a physical machine (local or remote via SSH) with a specific LLM provider's CLI. The control plane registers these members, manages their health, and routes tasks. This allows a single workflow to use Claude Sonnet for coding, Claude Opus for review, and Antigravity's Gemini for testing, all on different machines.
2. Durable, Observable Workflow Engine: Workflows are not just prompts; they are resumable programs with execution journals. The system includes stall detection, crash watchdogs, member reservations, and automatic orphaned process recovery. This provides the operational reliability required for unattended, multi-hour autonomous runs, with progress visible via live dashboards.
3. Secure, Out-of-Band Credential Management: A core security tenet is that no LLM ever sees raw credentials. Secrets are entered in a separate terminal, encrypted at rest, and resolved server-side. Credentials are scoped to specific members with time-to-live (TTL) expiry and composed, allow-listed permissions, moving beyond "god-mode" access.
4. Cost-Optimized, Deterministic Execution: The platform distinguishes between exploratory (token-heavy) and hardened workflows. Mechanical steps like shell commands, git operations, and file manipulations run at zero LLM cost. Models are invoked only at defined "judgment nodes" (e.g., code review, planning), dramatically reducing operating expenses after a workflow is discovered and solidified.

Problems Solved

1. Pain Point: The "operations gap" in AI agent deployment. While frameworks exist to compose agent logic, they lack production-grade features for scheduling across real hardware, credential security, cross-vendor model routing, and ensuring long-running workflows actually complete without manual babysitting.
2. Target Audience: Software engineering teams and DevOps professionals already using multiple AI subscriptions (e.g., Claude, GitHub Copilot, Google Gemini); developers with idle on-premise or cloud hardware (GPUs, servers); organizations with data privacy or cost control requirements that preclude solely using cloud-hosted agent services.
3. Use Cases: Autonomous software development sprints ("fleet-sprint" workflow for plan-develop-review-test-deploy cycles); cross-provider quality assurance (using a different LLM for review than for development); cost-tiered task routing (cheap models for mechanical tasks, premium models for complex reasoning); executing sensitive workflows on air-gapped or private infrastructure.

Unique Advantages

1. Differentiation: Unlike single-process agent frameworks (LangGraph, CrewAI), apra-fleet orchestrates across isolated, real machines. Unlike CI/CD runners (GitHub Actions self-hosted runners), it maintains conversational state and context between agents across phases. Unlike multi-cloud job schedulers (SkyPilot), it manages the full agent lifecycle, credentials, and LLM context, not just compute tasks.
2. Key Innovation: The architecture cleanly separates the generic, domain-agnostic workflow engine (Layer 4) from the vertical application (e.g., software engineering in Layer 5). This "seam" allows the same durable orchestration layer to be applied to different domains. Furthermore, the product serves as its own existence proof, autonomously building and maintaining its own codebase through the fleet-sprint workflow.

Frequently Asked Questions (FAQ)

1. What is apra-fleet and how does it relate to MCP? Apra-fleet is an MCP (Model Context Protocol) server that exposes fleet management capabilities—like registering machines and launching workflows—to any MCP-compatible AI agent (Claude Code, Cursor, etc.). The agent becomes the orchestrator, using natural language to command the fleet via this protocol.
2. Can apra-fleet use local LLMs like Llama or Mistral? Yes, through the OpenCode provider integration. This allows you to mix local, private models with commercial APIs like Claude and Gemini within the same fleet, routing tasks based on privacy needs and cost.
3. How does apra-fleet handle security and API keys? It uses a strict out-of-band (OOB) authentication model. API keys and SSH passwords are typed into a dedicated, secure terminal session, never pasted into an AI chat. Credentials are encrypted, scoped to specific fleet members, and have configurable TTLs, following the principle of least privilege.
4. What does "durable workflow" mean in apra-fleet? It means workflows are designed to survive interruptions. The engine maintains an execution journal, can detect and resume stalled tasks, automatically recovers from crashed agent processes, and re-authenticates expired credentials. This is essential for running autonomous workflows that last hours or days without supervision.
5. Is apra-fleet a replacement for CI/CD like GitHub Actions? No, it serves a different, complementary purpose. While it can execute shell and git commands, its strength is orchestrating stateful, multi-step AI agent collaborations with context passing between phases. It is conversational and adaptive, whereas traditional CI/CD is trigger-based and pipeline-defined. They can be used together.