openstation

Start Here

Project status

Current capabilities, runtime limits, and validation gaps.

OpenStation is pre-v1. The core runtime and configuration surfaces are implemented, while multi-agent routing, cost aggregation, and live connector acceptance still have gaps.

Available today

Area Available behavior
Agents Manifest declarations, Claude-native charters, profiles, and explicit executor selection
Conversations Local REPL, Slack, Telegram, and Email connectors; identity and space policy
Permissions Settings-file enforcement on Claude executors, scoped roles, generated gates, and drift checks
Workspaces Git-backed artifacts, sessions, events, and inspection commands
Automation Markdown interval jobs, manifest schedules, event triggers, and run history
Operations serve, per-turn budgets, approval requests, and connector supervision

See getting started for installation from source. The CLI is not yet a published package in the documented installation flow.

Limits to plan around

Agent routing is selected at startup. One process serves one agent. Declaring several channel claims does not enable per-message dispatch among those agents. Use explicit agent selection in separate processes where needed.

Connection names are not routing namespaces. Selectors and automation delivery use connector types. Two connections of the same type cannot be independently selected by the conversation key prefix.

Budgets apply per turn. There is no aggregate workspace or daily spend cap, nor per-tool cost accounting. claude-cli does not support maxTurns; supported budget behavior varies by executor. See permissions.

The Pi executor is experimental. Its permission translation is not equivalent to Claude Code's enforcement. It does not support Claude hooks, sandbox settings, or MCP tool translation.

Background automation requires serve. The REPL does not run jobs. Scheduling is interval based and has no cron or timezone semantics.

Validation still needed

The internal roadmap tracks live acceptance work for the Slack acceptance agent, the Telegram transport, and the Email transport. Fake-backed transport tests do not establish production reliability. Verify your credentials, inbound messages, replies, and any attachment or automation delivery path before relying on a connector.

This page summarizes the current repository documentation; the roadmap remains the detailed engineering record. For deployment procedures, see deploying.

View Markdown source on GitHub ↗