One install. Shared memory, smart routing, disciplined workflow, multi-AI audits -- across every AI coding agent. No BS.
# Add to your Claude Code skills
git clone https://github.com/TheRealSeanDonahoe/ijfwYour AI is brilliant. It's also forgetful, undisciplined, alone, and quietly burning tokens you never needed to spend. One install fixes all four. Richest on Claude Code, connected through Codex, Gemini, Cursor, Windsurf, Copilot, Hermes, Wayland, OpenCode, Qwen Code, Kimi Code, OpenClaw, and Aider.
Thirteen AI coding agents sharing one local memory that survives every session, every project, every restart. One disciplined workflow drives anything you'd open an AI for: code, books, launches, research, businesses. From idea to ship.
An on-demand three-AI critique puts a second lineage in the room so one model's blind spot never reaches production. Tasks dispatched to the right model for the job, output rules that strip the padding, prompt-cache discipline that compounds every session. Your bill goes down while your quality goes up.
A portable design contract keeps every AI on-brand: drop a DESIGN.md in your project root, or pick from 12 curated templates (Claude Design compatible), and the same palette, typography, and component rules carry across every agent on your stack. No more "make it look nice" as a prompt. No more cross-agent drift.
Whatever AI tools you use today, whatever you switch to tomorrow, IJFW makes them work as one brain. You bring the AI accounts you already use; IJFW adds nothing to pay for, runs entirely on your machine, and keeps every byte local.
One command to install. Then your AI starts behaving like a senior partner that's already read the brief. And you start wondering how anyone still works the old way.
No comments yet. Be the first to share your thoughts!
npm install -g @ijfw/install && ijfw-install
One command. Thirteen AI coding agents configured. Nothing to log into.
Windows (PowerShell 5.1 or 7+):
iwr https://raw.githubusercontent.com/TheRealSeanDonahoe/ijfw/main/installer/src/install.ps1 -OutFile install.ps1
.\install.ps1
Preflight (all platforms): Node 18 or newer, Git, and a bash shell (Git for Windows ships one). Windows users do not need WSL. If execution policy blocks the script, run PowerShell as powershell -ExecutionPolicy Bypass -File .\install.ps1.
What you see:
+----------------------------------------+
| |
| IJFW It just f*cking works. |
| |
+----------------------------------------+
Installed at ~/.ijfw
==> LIVE NOW (6)
o Claude Code
o Codex
o Gemini
o Windsurf
o Hermes
o Wayland
==> STANDING BY (2) auto-activate on install
o Cursor
o Copilot
==> ONE MORE STEP inside Claude Code
/plugin marketplace add ~/.ijfw/claude
/plugin install ijfw
Full log ~/.ijfw/install.log
Live now is the platforms on your machine, configured immediately. Standing by is every platform you do not yet have, pre-staged. Install Cursor next month and IJFW activates automatically. No rerun, no setup, no thought.
Every existing config gets a .bak.<timestamp> backup. Your existing MCP servers, model preferences, and per-project trust settings are preserved.
Every install ships the complete guide. Open it in your terminal or your browser -- no network required, no page to bookmark, no docs site to remember.
ijfw help # terminal, paged through less -R
ijfw help --browser # rendered in your default browser, dark theme
90 seconds to your first win. 10 minutes to the full tour. 45 minutes to master it.
Copy this whole block, paste it into any AI coding agent you use (Claude Code, Codex, Cursor, Windsurf, Copilot), and hit enter. The agent will install IJFW, verify the MCP server is running, and prompt you to restart.
Install IJFW (It Just F*cking Works) on my machine with the one-command installer, then verify it is working.
Run this exactly:
npm install -g @ijfw/install && ijfw-install
After it finishes, do these verifications and paste the results back to me:
1. Run `ijfw doctor` -- expect a platform table showing which AI agents are live
2. Run `ijfw --help` -- expect the full command list
3. Check `~/.ijfw/claude/.mcp.json` exists and contains an absolute node path
4. Run a quick MCP handshake test:
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{}}}' | node ~/.ijfw/mcp-server/src/server.js | head -1
expect a JSON response containing "result" and "ijfw-memory"
If anything fails, read the error from the installer log at ~/.ijfw/install.log and tell me what is broken.
When everything is green, tell me to fully exit my AI agent (not just a new tab -- full process quit) and restart it so IJFW loads on session start. Once restarted, run /mcp inside Claude Code to confirm plugin:ijfw:ijfw-memory shows as connected.
One command reverses the install across every configured platform. Memory is preserved by default.
ijfw uninstall # keeps ~/.ijfw/memory/ (recoverable on reinstall)
ijfw uninstall --purge # also removes memory (destructive)
Equivalent aliases: ijfw off (same as ijfw uninstall), or ijfw-uninstall from the @ijfw/install npm package.
What gets removed: IJFW entries from ~/.claude/settings.json (mcpServers.ijfw-memory, enabledPlugins['ijfw@ijfw'], extraKnownMarketplaces.ijfw), Codex / Gemini / Cursor / Windsurf / Copilot / Hermes / Wayland platform configs, ~/.ijfw/claude, ~/.ijfw/mcp-server, the ijfw* binaries from ~/.local/bin/, and plugin cache under ~/.claude/plugins/cache/ijfw/. Every modified file gets a .bak.<timestamp> first. Your other plugins, MCP servers, and per-project trust settings stay untouched.
Run this before any publish or production deploy. ijfw preflight executes an 11-gate quality pipeline in under 90 seconds on a warm cache and exits 0 only when every blocking gate passes.
ijfw preflight
| # | Gate | Blocks on fail? | What it catches |
|---|------|-----------------|-----------------|
| 1 | shellcheck | yes | Unbound variables, POSIX violations in hook scripts |
| 2 | oxlint | yes | Unused imports, dead variables in JS/TS |
| 3 | eslint-security | advisory | Security anti-patterns (non-literal fs paths, injection sinks) |
| 4 | psscriptanalyzer | advisory on macOS | PowerShell lint (blocking in Windows CI) |
| 5 | publint | yes | package.json bin/exports integrity |
| 6 | gitleaks | yes | Plaintext secrets and credentials |
| 7 | audit-ci | yes | npm audit: high and critical vulnerabilities |
| 8 | knip | advisory | Unused exports and dead code |
| 9 | license-check | advisory | Production dep license compatibility |
| 10 | pack-smoke | yes | npm pack -> temp install -> ijfw --help exits 0 |
| 11 | upgrade-smoke | yes | Plugin key wiring after upgrade from floor version |
Each gate uses npx --yes <tool>@<pinned-version> with versions tracked in preflight-versions.json. Missing tools are reported as "skipped" with a positive install hint. Preflight is optional -- it is not on the install path -- so the three-second install promise holds.
SLO: warm cache <=90s, cold cache <=240s. Both are printed in the summary line.
PASS 8 WARN 2 SKIP 1 FAIL 0
Time: 7s within warm-cache SLO (<=90s)
All blocking gates passed.
A live window into every AI session across Claude, Codex, and Gemini. Every PostToolUse event appends one JSONL line to ~/.ijfw/observations.jsonl. The dashboard reads that ledger and streams new events to the browser in real time via SSE.
Dogfood receipt -- one machine (the author's), 30-day window: $5.1k today across 5 active sessions, 97% cache efficiency, 2.71B tokens served from cache, $9.4k actual 30-day burn. These are one user's real numbers, not an averaged benchmark. Your own mileage materialises the moment you run ijfw dashboard start against your own traffic.
Dollar-saved ledger. The dashboard's headline tile shows actual spend, estimated spend without IJFW, and savings amount + percentage computed from three levers applied to your real session data: cache hit rate vs a 25% baseline (Anthropic cache reads at ~10% of input pricing), model routing (Haiku fraction vs all-Sonnet baseline), and output discipline (30% measured midpoint of the 20-40% range). Composite capped at 5x for defensibility. Methodology toggle inline -- every number traced to its source. Your numbers, defensible math, zero marketing.
ijfw dashboard start # bind 127.0.0.1:37891, open browser
ijfw dashboard status # show port + observation count
ijfw dashboard stop # graceful shutdown
The web dashboard at `http://loca