OpenWolf: Shared Local Project Memory for Claude Code, Codex, and OpenCode

OpenWolf is an AGPL-licensed npm tool that keeps one project memory folder shared across Claude Code, Codex CLI, and OpenCode, and trims oversized terminal output before it lands in the agent's context. The video is a 35-second walkthrough of the GitHub repo and the project site, with the voiceover reading the README pitch almost word for word. The repo has about 2.3k stars and shipped v2.5.2 the day after this video was posted.

What the video shows

The frames alternate between two screens. The first is the GitHub page for cytostack/openwolf, showing 28 open issues, 28 pull requests, a main branch, and a file tree with bin, docs, scripts, src, and tests folders. A commit message visible in the tree reads fix(2.5.1): multi-writer s..., which dates the recording to the v2.5.1 era. Badges in the README show npm v2.5.1, 4.2k downloads per month, 2.3k stars, AGPL-3.0, and Node 20 or newer.

The second screen is openwolf.com. The hero reads "Your agents change. Your project memory shouldn't." A terminal panel shows openwolf init detecting agents, registering Codex hooks, installing skills, creating a .wolf/ folder, and indexing 247 files. Further down are stat callouts (48%, 10x, 0) and a bar chart comparing a 41,283-token command output against 1,850 tokens of "entered context."

What it actually does

The README and site describe five jobs:

A .wolf/ directory in your project stores task checkpoints, a project map, notes, and known fixes. Claude Code, Codex CLI, and OpenCode read from and write to the same folder, so switching tools does not lose what the last agent learned.
A Bash output governor catches large command outputs (grep floods, git show dumps, test logs, whole files printed with cat) and condenses them before they enter the context window. The full output stays on disk with a pointer.
Repeated file reads get flagged so the agent does not re-read something already in context.
Token usage is read from the harness transcript per session and per agent, priced at list rates, with cache rebuilds attributed to a trigger such as a model switch, compaction, or version change.
A local dashboard shows usage and memory state.

Integration depth varies. Claude Code gets full lifecycle hooks. Codex CLI and OpenCode get core hooks (OpenCode via a native plugin). Cursor, Gemini CLI, and Antigravity only receive project instructions, with no interception.

The privacy claim

The video says OpenWolf works "without sending telemetry or making external API calls." The README states the same: pure local file I/O, no API calls, no telemetry, no added latency. This is consistent across the GitHub repo, the npm package description, and the site. I did not audit the source, but the claim is the project's stated design.

Claims I could not verify

The site copy visible in the frames says the team audited 16 live projects and 6,869 real API calls before building 2.x, and that 48% of tool-result tokens flow through Bash. When I fetched openwolf.com, those figures and the "Bash Output Governor" feature card were not on the page, so either the site was redesigned between the recording and now or the numbers moved to a subpage. Treat the 48% and 10x numbers as the project's own marketing, not independent measurement.

Who it is for

The video's pitch is developers running AI coding CLIs who are tired of burning context on raw command output and who move between agents. That matches the README's "Without / With" comparison table: each agent starting cold versus one shared brain, and a monthly token invoice with no line items versus per-session accounting. If you use one agent and rarely run noisy commands, the token accounting is most of the value.

Key Takeaways

OpenWolf stores project memory in a local .wolf/ folder that Claude Code, Codex CLI, and OpenCode all share.
Its main token saver is condensing large Bash output before it enters context, with the full output kept on disk.
Everything runs as local file I/O; the project claims no API calls and no telemetry.
Install is npm install -g openwolf then openwolf init in the project; Node 20 or newer is required.
The repo is AGPL-3.0, has about 2.3k stars, and released v2.5.2 on September 15, 2026.
The 48% and 10x efficiency figures come from the project's own audit and were not on the live site when checked.

Resources

cytostack/openwolf on GitHub -- source, README, and releases (AGPL-3.0, ~2.3k stars)
openwolf.com -- project site with guide, commands, and config docs
openwolf on npm -- package page; v2.5.2 confirmed via npm view