A self-aware personal agent.

ARCHON

Most AI products fail not on the model, but on orchestration.

View on GitHub → (opens in new tab)

Unifies multiple AI models and coding agents into one orchestration layer for real development workflows.

3models orchestrated
3coding agents delegated
Opensource, self-hostable

What Archon is

Archon is an evolving, self-aware personal agent: a control plane that plans, routes, executes, and recovers across providers and tools, with persistent memory, context compression, usage accounting, and human approval built in, rather than a single chat wrapper.

Key decisions

01

Context

Single-prompt chat wrappers break on real, multi-step work. Reliability comes from orchestration, memory, and the ability to inspect and approve what the agent did, not from a bigger model.

02

Decision

Build a lightweight control plane in Python: route across Claude, GPT, and Gemini, expose tools over the filesystem, web, and MCP, persist compressed memory, and delegate heavy tasks to coding-agent workers (Claude Code, Codex, OpenCode) through a worker router, with policy and human approval as first-class.

03

Result

Work flows through sessions, jobs, and turns with reasoning traces, redaction, and usage tracking. Archon reaches the operator on the terminal, Telegram, voice, and live phone calls, and runs its own news and deep-research pipelines.

A foundation for autonomous but inspectable personal systems, where orchestration, memory, and recovery absorb the complexity that would otherwise leak into every task.

How it is built

  • Model layerAnthropic Claude, OpenAI, and Google Gemini behind one agent
  • Control planeOrchestrator, policy, sessions, and jobs over execution turns
  • Worker delegationRoutes heavy tasks to coding agents: Claude Code, Codex, OpenCode
  • Tools and MCPFilesystem, web read and search, content, and Model Context Protocol tools
  • Memory and contextPersistent memory with context compression, distillation, and usage accounting
  • ChannelsTerminal REPL, Telegram with approvals, voice (STT and TTS), and Twilio calls
  • SafetyRedaction, policy guardrails, and human approval gates
Intelligence is not the model. It is the system around it.
agent protocol