Where Warden fits the agent stack
Warden is the part of the AI dev stack that actually acts — here's how it sits alongside the other four tools.
Every other piece of an AI dev stack prepares for a moment that Warden owns: the moment an agent does the thing. Warden is a lean, multi-provider, terminal-native coding agent — a single binary that edits your code, runs your tools, and lives where you already work.
It doesn’t try to be the whole stack. It orchestrates nothing — Arbiter does that. It doesn’t invent ground truth about your code — TheAuditor does that. It doesn’t remember how your standards have shifted — Curator does that. Warden is the part that acts, with everyone else’s output already in context.
The four it works with
- TheAuditor feeds Warden deterministic code context, so it edits against facts instead of guesses.
- Arbiter dispatches Warden across providers and brings runs back after a crash.
- Curator hands Warden memory about your evolving standards, reachable over MCP.
- BenchProctor is the independent benchmark for the SAST findings Warden acts on.
What Warden brings to the table
It speaks to multiple providers from one binary, with full MCP client and server mode — it connects to any MCP server and re-exposes its own tools to any MCP-speaking client. It ships a real permission engine with fail-closed defaults and composable hook events, so a single deny blocks a tool call. It tracks cost per model with a hard spend cap, and it resumes a session by id or working directory. When the context window fills, it compacts surgically — preserving the load-bearing context rather than summarizing it away.
Where it stands
Warden is pre-alpha (v0.1.0) and open source (MIT). Three providers are fully wired today and three more are scaffolded; native Windows is best-effort, with WSL proven. No semver promises until v1.0. If that’s your kind of tool, get launch updates — or read how it pairs with TheAuditor.
Was this useful?