Built on the wrong stack. Shipped anyway.
WrongStack is a terminal-native AI coding agent and autonomous multi-agent orchestration platform built for developers who want more than another chat window.
Give it a goal, and WrongStack can assemble a fleet of specialized agents, assign each one its own provider, model, tools, permissions, and budget, coordinate their work, and keep going until the task is verifiably complete.
You stay in control the entire time. Watch the fleet, steer it, pause it, resume it, or walk away and let it work autonomously.
At its core, WrongStack is a standalone TypeScript system with its own runtime, tools, permissions, codebase intelligence, event system, long-term project memory, and multi-agent coordination layer. It is not just a thin wrapper around another coding agent.
You can sign in with your ChatGPT Codex account, securely add the API keys you already have, and use providers and subscriptions such as DeepSeek, MiniMax, Kimi, Zai GLM, Alibaba Qwen, Anthropic, OpenAI, and OpenCode Go — or run WrongStack entirely with local models.
Models can be switched at runtime. Different tasks can use different providers and models, and fallback profiles can automatically take over when a quota is exhausted, an API becomes unavailable, or a provider has an outage. The goal is simple: your workflow should keep moving instead of stopping because one model or provider failed.
This also means you are not forced into a single-provider setup. One agent can run on one model, another agent can use a completely different provider, and the Director can coordinate them as part of the same fleet.
WrongStack exposes the same engine through multiple surfaces — including REPL, TUI, WebUI, desktop-style interfaces, and higher-level orchestration views — while sharing the same session state, project memory, tools, and runtime underneath.
Its autonomy loop is straightforward:
Plan → Execute → Reflect → Repeat
The Director can fan work out to specialists for coding, debugging, research, review, testing, architecture, and other roles. Agents can collaborate on the same objective while keeping their own provider, model, context, budget, and permission boundaries.
WrongStack also includes durable project memory through SAGE. Instead of treating every session like a fresh chat, solved problems and useful knowledge about your files, symbols, commands, and commits can become reusable context for future work.
And autonomy does not mean unrestricted access.
WrongStack includes explicit permission boundaries, deny rules, capability gating, protected destructive actions, secret scrubbing, SSRF protections, and restricted subagent capabilities by default.
Use cloud models, subscription-backed providers, local inference, or all of them together.
Switch models while you work.
Assign different models to different jobs.
Build fallback chains.
Keep your fleet moving.
One engine. Multiple surfaces. A fleet, not a lone agent.
Built in TypeScript. Open source. MIT licensed.
wrongstack.com
github.com/WrongStack/WrongS…