A second brain for Claude Code. Now for every AI coding assistant.
Coding agents re-read files they already saw, scan whole directories to find one function, and forget your conventions between sessions. OpenWolf gives Codex, OpenCode, Claude Code, and other agentic systems one persistent project brain: improved context management, optimized architecture scaffolding, and smarter token utilization. No workflow changes.
Coding agents are powerful. But they work blind.
An agent does not know what a file contains until it opens it. It cannot tell a 50-token config from a 2,000-token module. It reads the same file three times in one session without noticing. It has no map of your project, no memory of past corrections, no awareness of what it already tried.
OpenWolf gives every agent a project index, a learning memory, and a token-aware read layer. And in 2.0 it measures real usage from harness transcripts, so the savings are verified rather than estimated.
Same project. Same prompts. Different setups.
Everything works invisibly
OpenWolf hooks into your agent's lifecycle. No commands to remember. It just makes every session smarter.
Every Agent, One Brain
Codex, OpenCode, and Claude Code share one project memory, with Cursor and Antigravity in beta. init auto-detects what you have installed and wires it. Learn once, apply everywhere.
Measured Tokens
Real usage is read from harness transcripts at session end: input, output, cache, per agent. openwolf report shows measured numbers next to the estimates. Verify, don't trust.
Context That Survives
A token-budgeted digest of your project's state is injected at every session start. Compaction is snapshotted and restored. Sessions stop starting from zero.
Symbol-Level Reads
Big files index their functions and classes with line ranges. Agents fetch one function with offset/limit instead of the whole file. Stale ranges are suppressed automatically.
Self-Learning
Every correction gets logged. Every bug fix gets remembered. Every preference gets enforced. Your agents get smarter with every session, without you repeating yourself.
Secure by Default
Loopback-only dashboard with token auth, zero shell interpolation, path traversal guards, and secret files excluded from every index. Plus bundled /security-audit and /reframe skills.
Three steps. Then invisible.
Initialize
Run one command in any project. Creates .wolf/ directory, registers hooks, scans all files.
Work Normally
Just use claude as you always do. Hooks fire invisibly, tracking, learning, enforcing. You don't interact with any of it.
Get Smarter
Every session, OpenWolf learns preferences, logs bugs, prevents repeated mistakes. View everything on the real-time dashboard.
The .wolf/ directory
Every project gets a .wolf/ folder containing state, learning memory, and configuration. Markdown is the source of truth.
anatomy-index.json
Durable project index: descriptions, token estimates, content hashes, and per-file symbols. Rendered to anatomy.md.
cerebrum.md
Learned preferences, conventions, Do-Not-Repeat mistakes. Gets smarter every session.
STATUS.md
Session handoff. The next session reaches productive context in one small read.
buglog.json
Bug encounter and resolution memory. Searchable. Prevents re-discovering the same fix.
hooks/
7 Node.js lifecycle hooks shared by all wired agents. Pure file I/O, no network, no AI calls.
token-ledger.json
Estimated and measured usage per session and per agent, straight from harness transcripts.
The enforcement layer
Seven hooks fire on every agent action. They warn but never block. Pure Node.js. No network, no AI, no extra cost.