Agents: A Self-Hosted, Observable Multi-Agent Workflow Orchestrator

This article introduces the Agents project—a self-hosted, observable agent orchestrator that supports running multi-agent workflows on code repositories. Its core features include skill composition, memory management, event-triggered collaborative scheduling, multi-backend model support (including local LLMs), fine-grained cost control, and reproducibility design. The project aims to enable developers to build and own their own agent universe without SaaS dependencies, ensuring data security and control.

The Agents project is maintained by eloylp, with source code hosted on GitHub (link: https://github.com/eloylp/agents, last updated: 2026-05-31). Its core vision is to enable developers to build their own agent universe, with three core philosophies:

1. Self-hosted: Code and prompts remain on the user's infrastructure, no SaaS dependencies, data never leaves the controlled environment;
2. Observable: Full event chain tracking, with real-time monitoring of the entire process from trigger to execution;
3. Composable: Modular design via Skills, Memory, and Triggers, supporting execution triggered by repository events, tags, cron schedules, or inter-agent scheduling.

Agents uses a daemon architecture, with the following execution pipeline:

1. Prompt Composition: Workspace guardrails + Skills + selected prompts + runtime context + Memory;
2. Containerized Execution: Launch containers from the agents-runner image, configure git identity, run AI CLI (Claude Code/Codex/local LLM), enforce JSON output, and enable repository tools;
3. Response Parsing: Extract outputs, scheduling requests, and update Memory;
4. Persistence and Distribution: Record tracking information, distribute scheduling requests, and write back to Memory. The design ensures isolated execution, parsability, and traceability.

The project offers three interaction methods:

1. Web Dashboard: A graphical interface supporting workflow design (agent, prompt, skill management, etc.), real-time event streams, tracking, and memory viewing;
2. MCP Server: Terminal-based conversational interaction, controlling agents via MCP clients like Claude Code/Cursor;
3. REST API: A programmatic interface supporting script calls, on which the dashboard is built. The multi-mode design adapts to visual operations, automated scripts, and conversational development scenarios.

Agent Collaboration: Supports reactive inter-agent scheduling with safety limits including depth (anti-recursion), fan-out (concurrency control), and deduplication (avoid repetition), enabling complex task division (e.g., code review + test generation). Security Design:

• Built-in prompt guardrails: Resist indirect injection, cautious public operations, memory access restrictions, GitHub tool specifications;
• Layered authentication: Local admin creation, HttpOnly sessions (UI), revocable Bearer Tokens (MCP/API).

Multi-Backend: Each agent can choose Claude, Codex, or a custom backend; the same fleet can mix different providers. Local Models: Supports llama.cpp, Ollama, vLLM, or OpenAI-compatible endpoints. An experimental built-in Anthropic-to-OpenAI translation proxy allows the claude CLI to connect to local LLMs, lowering the barrier to use.

Cost Management: Fine-grained Token budgets (daily/weekly/monthly UTC limits) supporting isolation scopes like global, cross-workspace (repo/agent/backend), and workspace+X; threshold alerts, Token consumption leaderboards, and multi-channel CRUD operations. Versioning: Prompts/Skills/Guardrails are versioned by scope (global/workspace/repository). Edited versions are published as immutable, and exact used versions are tracked to ensure reproducibility, facilitating debugging and auditing.