openstation

Reference

Project overview

OpenStation's scope and design in one page.

The agent you built on your laptop already works. OpenStation makes it reachable, safe, and managed — without rewriting it into a framework.

The problem

Every agent framework asks you to rebuild your agent inside its abstractions: its prompt templates, its tool registry, its context pipeline. But the best agents today already exist as local coding-agent projects — a folder with instructions, skills, and tools that works when you run it by hand. What's missing isn't a smarter framework. It's everything around the agent: channels, identity, permissions, memory, scheduling.

What OpenStation is

The station your agents operate from — where they're reachable, provisioned, and accountable:

Export your local agent. An agent that runs as a Claude Code project becomes reachable over Slack, Telegram, and Email through pluggable executors (Claude CLI and SDK today; Pi experimental; Codex on the roadmap). The platform never owns the reasoning — it routes work to the runtime that does.

Everything around the agent, nothing inside it. Out of the box: messaging connectors, identity and deny-by-default permissions enforced by the runtime itself (no bypass path), per-conversation serialization and global concurrency caps, human-approval gates, session continuity, and an event bus with full observability. A context layer that gives every executor the same memory is on the roadmap — it's what makes the multi-runtime story real.

A workspace that works. Each agent lives in a persistent, git-backed workspace: markdown knowledge the agent reads itself, automations wired to triggers, file watchers, and schedules. Every change the agent makes is a commit — diffable, reviewable, revertible.

Agentic IaC. The whole station is declared in YAML — agents, channels, people, permissions, budgets, triggers, automations. Agent behavior stays in plain Claude-native files. Version it, review it, reproduce it.

Why it's different

Typical agent framework OpenStation
Rebuild your agent in our SDK Bring the agent folder you already have
Framework assembles context The agent's runtime materializes context
Permissions as app logic Deny-by-default, enforced by the runtime the agent runs in
State in a database you can't read State in a git repo you can diff

Status

Every layer is built and tested (464 tests): four connectors, the Bridge and TurnRunner, both executors with settings-based enforcement, the YAML config plane, git workspaces with commit automation, triggers, scheduling, approvals, and the full CLI. v1 completes when two real agents run end-to-end on real channels — in progress.

View Markdown source on GitHub ↗