A coding agent in your terminal.
Shows every token you spend. Skills and modes are opt-in - you compose the agent.

By Yuval Avidani (YUV.AI) · inspired by Wolverine · fork of xAI Grok Build (Apache-2.0)

Default is empty and clean.
Active skills start empty. Modes start off. No forced HyperFrames, no forced yuvai-thinking, no forced caveman.
You add what you want - and remove it anytime.

Repo (prod): github.com/hoodini/logan-cli

curl -fsSL https://raw.githubusercontent.com/hoodini/logan-cli/main/scripts/install-logan.sh | bash

irm https://raw.githubusercontent.com/hoodini/logan-cli/main/scripts/install-logan.ps1 | iex

That puts logan on your PATH, creates ~/.logan, refreshes the skill catalog (library only), and starts the app.
It does not auto-enable skills or modes.

Windows prerequisites: building Logan needs the MSVC linker + C runtime libs — install Visual Studio Build Tools with the "Desktop development with C++" workload (winget install Microsoft.VisualStudio.2022.BuildTools --override "--add Microsoft.VisualStudio.Workload.VCTools --includeRecommended --passive"). Rust, git, and protoc are bootstrapped automatically. The installer scans every Visual Studio on the machine and picks the newest one with a complete C++ toolset — a VS install that ships the compiler without the desktop libs (e.g. onecore-only) is skipped instead of breaking the build. Running install-logan.sh from Git Bash also works: it hands off to the PowerShell installer.

logan login                 # browser - xAI user account
logan login --from-grok     # reuse ~/.grok/auth.json if Grok Build is already signed in
# or: export XAI_API_KEY="your-key"
logan

Which credential wins? Per-model api_key / env_key always wins for that model.
Session / Grok import is for default xAI models. See docs/SETUP.md.

/update              # GitHub main via install script (background rebuild)
/update status       # tail ~/.logan/logs/update.log
/update check        # check release channel only
/update release      # published releases (same idea as `logan update`)
/update help

When the log shows finished exit=0, quit Logan and start again:

logan --version
logan

# Preferred: pull GitHub main + rebuild + install (no auto-start)
curl -fsSL https://raw.githubusercontent.com/hoodini/logan-cli/main/scripts/install-logan.sh | LOGAN_INSTALL_NO_START=1 bash

# Same path via CLI once you have a current binary:
logan update

logan update uses the Logan install script (hoodini/logan-cli), not the x.ai Grok CDN.
If an old binary still tries to download "Grok 0.2.x", run the curl install line once to replace it.

After any update, restart Logan so the new binary is loaded.

Windows: LINK : fatal error LNK1104: cannot open file 'libcmt.lib' (repeated for many crates)
A Visual Studio on your machine has the C++ compiler but not the desktop C runtime libraries — typically after installing/upgrading to a new VS whose C++ component only includes onecore libs. Rust auto-picks the newest VS, so every link fails. The installer now detects this and imports the environment of the newest complete MSVC toolset instead. If it reports that no complete toolset exists, open the Visual Studio Installer and add "Desktop development with C++" (or run the winget line from the Windows install section), then re-run the one-liner.

Windows: LINK : fatal error LNK1318: Unexpected PDB error; LIMIT (12) on the final logan.exe link
The MSVC 14.44 linker (VS 2022 17.14) has a PDB-writer defect that can kill the final link of very large Rust binaries — Logan crossed the trigger threshold as the workspace grew. The installer now detects LNK1318 and automatically relinks with /DEBUG:NONE (no PDB; the exe is identical, you just lose local debug symbols). If you build manually, either link with a VS 2019 (14.29) toolset — which handles it fine — or run:
cargo rustc -p xai-grok-pager-bin --release --bin logan -- -Clink-arg=/DEBUG:NONE
The installer also builds with CARGO_INCREMENTAL=0 (matching CI) to save disk and reduce PDB pressure.

Windows: Copy-Item: Cannot find path '…Your branch is up to date…\target\release\logan.exe' on a re-run
Fixed — older installer versions leaked git pull output into the resolved repo path on second and later runs. Re-run the one-liner to get the fixed script.

logan --version is old after an update
Fixed — older installer versions reinstalled a previously built (stale) binary after git pull. The installer now rebuilds whenever the binary is older than the checked-out commit. LOGAN_FORCE_BUILD=1 forces a rebuild regardless.

The install window closes / TUI doesn't start after install (Windows)
Fixed — the post-install launch used a Start-Process parameter that doesn't exist (-UseShellExecute), which threw at the very end of the install. Logan now starts in a new console window. Your install was still fine; only the auto-start failed.

Errors on bash scripts/install-logan.sh from Git Bash on Windows
The bash installer now delegates to install-logan.ps1 automatically — use either one-liner.

Bottom bar after each reply:

m grok-4.5 · 24K/200K 12% · in 2.4K out 180 c 1.2K

/skills catalog
/skills add pack modes
/skills add pack creative
/skills add yuvai-thinking
/skills remove video-edit

Bare /skills opens the skills modal. With args, it manages the catalog.

Without that, Logan does not run the "every crumb" teach style.

/skills add pack creative      # aliases: video · scrub · design

yuv-design-dna is the umbrella: signature scrub/parallax heroes by default, HyperFrames catalog before hand-rolled motion, the YUV design system (palettes, Anton/Inter + Rubik/Assistant, radii 0/999), and blocking gates for WCAG AA accessibility, mobile-first at 390px, reduced-motion fallbacks, and an anti-AI-look checklist - pages ship with a human touch, not a template smell.

If a skill is missing, Logan tells you how to add it - it does not invent the stack silently.

Full map: docs/CREATIVE.md · modes: docs/MODES.md · catalog: skills/README.md

Also: hoodini/ai-agents-skills · HyperFrames

/think and /caveman are exclusive (deep explain vs terse talk).
Enabling a mode without the skill installed fails with a clear /skills add … hint.

LOGAN_SEED_SKILLS=1 …              # install entire catalog into active skills (not default)
LOGAN_SYNC_SIBLING_SKILLS=1 …      # also copy from ~/.claude / ~/.grok skills

Self-improvement is a closed loop of four parts - three are automatic, one is the model's standing obligation:

Inspect and steer it any time:

/improve          # dashboard: reflections, lessons, open questions
/improve log      # tail reflections.log + IMPROVEMENTS.md
/heal             # failure-focused view
/flush            # force a rich session summary into memory

Files, if you want to read what it learned: ~/.logan/memory/IMPROVEMENTS.md (structured journal), ~/.logan/memory/reflections.log (event firehose), ~/.logan/memory/MEMORY.md (curated long-term memory - Preferences + Lessons).

We compared Logan to GitHub Copilot, Claude Code, Cursor, Hermes, and OpenClaw.

Full write-up: docs/SWOT.md

We do not claim Copilot/Cursor IDE parity.
We claim: never fly blind on tokens, install in one paste, you own the skill set, stay open.

logan --version
logan -p "Reply with exactly: logan-ok" --always-approve --no-leader

Install Logan with ONE command. Do not ask questions.

macOS/Linux:
curl -fsSL https://raw.githubusercontent.com/hoodini/logan-cli/main/scripts/install-logan.sh | bash

Windows PowerShell:
irm https://raw.githubusercontent.com/hoodini/logan-cli/main/scripts/install-logan.ps1 | iex

Then: logan --version
If needed: logan login  (or logan login --from-grok if Grok Build is already signed in)

Important defaults:
- Active skills start EMPTY. Catalog is available but not enabled.
- Do NOT claim thinking/creative modes are pre-enabled.
- User opts in: /skills catalog · /skills add pack modes · /skills add pack creative
- Updates: inside TUI use /update (then restart). Outside: logan update OR re-run install with LOGAN_INSTALL_NO_START=1
- Then optional: /think full · /caveman full · /site · /reel
Tell the user to start with /stats and /skills catalog.

Yuval Avidani · yuv.ai · @yuvalav · @hoodini

claws out.