About this project
Downcity is an agent infrastructure project for developers, indie builders and teams who ship many agent-powered products and workflows. Its premise is one reusable runtime layer serving many agents and products, so model routing, tools, memory, tasks, authentication, usage, billing and operations do not need to be rebuilt for every new AI product.
The repository is a pnpm workspace monorepo. The public downcity package exposes two command surfaces: city/downcity, the local City container that manages and runs agents, and fed/downfed, a Federation Server manager. Supporting packages include @downcity/workspace (workspace resources, rooted file and search tools, environment, private storage and built-in shell), @downcity/agent (agent runtime and City container with AgentWorkspace, Session, Plugin SDK, multi-agent ownership and HTTP/RPC transport), @downcity/federation (Federation runtime and Embassy SDK for services, auth, env, Bureau, user and admin access), @downcity/type (shared protocol types), @downcity/services (accounts, balance, usage, payment and Stripe flows) and @downcity/ui (React + Tailwind components for console and host applications). Starter templates cover agent, edge federation, local federation and React UI scenarios, and a homepage directory holds the website and end-user documentation.
Typical local usage: install the CLI with npm or pnpm, connect to a Federation via downcity federation use and federation status, register a repository as a workspace with downcity agent create ., then start the local City with downcity on and inspect it with downcity agent list or downcity status. Agent identity and config live in a user-level ~/.downcity directory (agents/, plugins/) while the workspace is entered at execution time. The README notes agents have no independent started/stopped state, and that chat still works while the CLI City is off because a temporary local City is created and disposed for the interaction. Listed built-in agent capabilities are chat, task, memory, shell, contact, skill, web, sound and workboard.
Integration is offered at several levels. The SDK example builds a Workspace with a Shell sandbox, creates an Agent and Session, injects a model created by the host application (the README states @downcity/agent does not resolve provider or model IDs itself), sends a prompt and awaits the finished turn. A RemoteAgent class connects to a running City over HTTP by base URL. Product surfaces are the Downcity CLI, Agent SDK, City SDK and UI SDK, plus plugins, services and React UI components.
Platform support and safety: macOS uses Seatbelt, Linux uses Bubblewrap, and Windows 11 24H2 and newer (build 26100+) is described as development/unstable via Microsoft MXC processcontainer with native cmd.exe execution. The README says preflight failures never fall back to unrestricted execution, and that MXC is in Public Preview rather than a production security boundary. Shell and script commands run in the agent sandbox by default, with a writable project, open network and sandbox home/cache under .downcity/sandbox. Advice includes auditing changes with git status and git diff, keeping secrets in local environment variables or downcity env, using tokens for console, HTTP and chat boundaries, and noting that host-level installs and system-directory writes fall outside the sandbox.
Documentation is published at downcity.ai/docs with separate City SDK, Agent SDK and UI SDK pages, package-level READMEs, and a Simplified Chinese translation of the main README.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.