five steps

Getting started.

Install, restart, verify. Warden handles the rest — pruning, compression, memory, verification. No config files, no settings, no tuning. Five commands total, and three are optional.

What warden init does

✓
Registers MCP server— in every detected agent config (30+ paths checked)
✓
Writes rules files— CLAUDE.md, .cursorrules, AGENTS.md, .devin/rules, etc.
✓
Builds code index— tree-sitter parsing — functions, classes, imports, call sites
✓
Compresses memory files— strips filler from existing CLAUDE.md, AGENTS.md, etc.
✓
Creates project database— SQLite — memories, CCR cache, audit trail, outcomes

Everything is active after init. No manual commands needed. Idempotent — safe to re-run anytime.

1

Install

One command installs Warden globally and runs init — registers in every detected agent, writes rules files, builds code index, compresses memory files.

$ npm install -g warden-ai && warden init

Or: curl -fsSL https://raw.githubusercontent.com/rynald0cst0ltziam/Warden-AI/main/install.sh | bash

warden init auto-detects every MCP-compatible agent on your machine (Claude Code, Cursor, Windsurf, Codex, Cline, Zed, JetBrains, Amazon Q, Gemini CLI, Aider, Goose, OpenHands, and 20+ more) and registers Warden in each. It writes rules files (CLAUDE.md, .cursorrules, AGENTS.md, .devin/rules) that tell your agent to use Warden's wrapper tools. It builds the code index for structural queries. It compresses memory files to save tokens every future session. Idempotent — safe to re-run.

2

Restart your agent

Close and reopen your IDE. The agent reads the rules files on startup and starts using Warden tools automatically.

New projects and spaces work automatically. When you open a new project in any agent, Warden auto-initializes on first serve — writing rules files and creating the project database. No need to run warden init again.

3

See your savings

Three ways to watch tokens accumulate as Warden prunes every tool call.

In your agent

Every tool call shows ‹warden› savings inline. Session start prints: "Warden active — X tokens saved (Y% reduction), Z rules live."

Terminal HUD

Live, real-time display of token savings, rule status, and confidence bars. Updates as the agent works.

$ warden hud
Web dashboard

Browser UI at localhost:7878 — rule status, confidence bars, token savings, recent decisions. Auto-refreshes every 3 seconds.

$ warden dashboard
4

Verify it works

Run a full health check — 10 checks, clear pass/fail report.

$ warden doctor

Checks MCP registration, rules files, pruning engine, memory system, FTS5 index, code index, CCR cache, outcome tracking, budget tracker, and database integrity. If anything fails, it tells you exactly what to fix.

5

Run benchmarks

See actual compression numbers on your own files — pruning, file compression, and response rules.

$ warden benchmark

Runs all pruning modules on real files from your project. Shows before/after tokens, reduction percentage, and trust guard pass rate for each. This is real evidence, not marketing claims.

After init, everything is active

✓4/4 pruning rules live
✓Output compression (max, always on)
✓Code intelligence (call graph, impact, architecture)
✓Memory system (FTS5 recall, conflict detection)
✓Trust guard (every line verified verbatim)
✓CCR (every cut reversible, 7-day TTL)
✓Outcome tracking (regression detection)
✓Session handoff (read at start, generate at end)

Making it automatic

The rules files written by warden init tell your agent to:

✓Call warden_handoff({ read: true }) at session start — read previous session's handoff
✓Call warden_status at session start and print savings
✓Call warden_memory_recall at session start for past decisions
✓Call warden_context_select before starting work
✓Use warden_grep instead of built-in grep
✓Use warden_file_read instead of built-in file read
✓Use warden_run_tests instead of running tests directly
✓Use warden_run_command instead of running shell commands
✓Call warden_memory_save after durable decisions
✓Call warden_record_outcome after completing tasks
✓Call warden_handoff({}) at session end or before context compaction

After running warden init, restart your agent. The agent reads the rules files on startup and follows them. You don't need to invoke any Warden commands yourself — the agent handles it all.