A coach for TryHackMe, powered by Claude Code. A consistent methodology (classify → enumerate → foothold → privesc), room-state tracking, ready-made commands, and specialist subagents for CVE research, privesc, and write-ups — all guiding you through each room step by step rather than solving it for you. A hands-on alternative to reading write-ups. Runs on Kali or any Debian/Ubuntu-based Linux (Kali just ships the tools preinstalled).
⚠️ Scope discipline. Everything here is built for authorized practice on TryHackMe boxes you've deployed. The config repeatedly tells Claude to stay on the assigned target IP only. Keep that guardrail. Never point this at a host you don't have explicit permission to test.
curl -fsSL https://raw.githubusercontent.com/pashki975/thm-claude-kit/main/install.sh | bash
That clones the kit into ./thm-claude-kit and offers to install the tooling (apt/pipx). Prefer to read
before you run (piping to bash executes a remote script blind)? Download it first:
curl -fsSLO https://raw.githubusercontent.com/pashki975/thm-claude-kit/main/install.sh
less install.sh # eyeball it
bash install.sh # then run — add --dir PATH, --tools, or --no-tools as needed
You still need Node.js 18+ and Claude Code (npm install -g @anthropic-ai/claude-code); the
installer warns if either is missing. Full options and the tooling list are in Install below.
thm-claude-kit/
├── README.md ← you are here
├── CLAUDE.md ← global context Claude reads automatically
├── WINDOWS-CHEATSHEET.md ← Linux-attacker's quick reference for Windows/AD
├── new-room.sh ← scaffolds a new room folder under rooms/
├── .gitignore ← ignores rooms/ (your scans, notes, creds, flags)
├── rooms/ ← ALL room folders live here (git-ignored)
│ └── .gitkeep ← keeps the empty dir in the repo
└── .claude/
├── commands/ ← slash commands (manual triggers you type)
│ ├── start.md → /start (room coach: setup + teach you through it)
│ ├── vpn-check.md → /vpn-check
│ ├── recon.md → /recon
│ ├── enum-udp.md → /enum-udp
│ ├── enum-web.md → /enum-web
│ ├── web-recon.md → /web-recon (vhosts / subdomains / cert SANs)
│ ├── wp.md → /wp (WordPress: wpscan + attack chain)
│ ├── enum-smb.md → /enum-smb
│ ├── db-enum.md → /db-enum
│ ├── steg.md → /steg (offline file/image analysis)
│ ├── pcap.md → /pcap (tshark traffic/capture analysis)
│ ├── listener.md → /listener
│ ├── linux-privesc.md→ /linux-privesc
│ ├── tunnel.md → /tunnel
│ ├── crack.md → /crack
│ ├── notes.md → /notes
│ │ ── Windows/AD ──
│ ├── enum-ad.md → /enum-ad
│ ├── kerberos.md → /kerberos
│ ├── win-shell.md → /win-shell
│ └── win-privesc.md → /win-privesc
├── agents/ ← subagents (specialists, run in their own context)
│ ├── recon-parser.md
│ ├── privesc-advisor.md (Linux)
│ ├── cve-researcher.md
│ ├── report-writer.md
│ ├── win-privesc-advisor.md (Windows)
│ └── ad-attack-advisor.md (Active Directory)
└── skills/
├── thm-trainer/SKILL.md ← the decision loop (runs first, every room)
├── thm-methodology/SKILL.md ← Linux game plan
├── network-forensics/SKILL.md ← packet-capture / traffic-analysis playbook
└── windows-ad-methodology/SKILL.md ← Windows/AD game plan
Commands = things you trigger by typing /name. Fast, predictable, one job each.
Agents = specialists Claude hands a focused task to ("use the cve-researcher agent"). They
run in a separate context so heavy output doesn't clutter your main session.
Skills = the connective tissue. They tell Claude how to think and in what order so when
you ask "what next?" you get a methodical answer instead of a guess.
The most important skill is thm-trainer — the decision loop that runs on every room
before any tool does. It exists because the #1 failure mode is an agent confidently doing
things that lead nowhere: parallel scans, sshing into a box that was never about shells, an
hour of motion with no progress. The loop fixes that by forcing five beats:
1. CLASSIFY → what is this room teaching? what's the goal (flag string vs shell)?
2. OBSERVE → the lightest thing that fits the class — don't over-scan
3. HYPOTHESIZE→ "the way in is probably X, because Y" (one sentence, out loud)
4. TEST → one focused action whose result decides the next move
5. DECIDE → new info → loop; goal met → STOP; no new info → reclassify
The rule underneath it all: every action should change what you do next. If a scan wouldn't change your plan, don't run it. Being stuck = repeating actions with no new information — the signal to reclassify, not to scan harder.
You kick a room off with /start (below). It's a playing coach, not an autosolver: it
does the mechanical setup (VPN check, /etc/hosts, notes), classifies the room and names the
goal, then teaches you through it one step at a time — explaining why each step, pointing you
at the right kit command, and letting you run it. It reads your level: light and terse when
you're flowing, and when you're stuck it drops into drive mode, reasons through the loop out
loud, and walks you through the next move in detail. The thm-trainer skill supplies the loop
it coaches from.
On a one-technique easy box the coach just runs the loop. Complex rooms are different — they're
a chain (web foothold → creds in a config → pivot to an internal host → AD → privesc), where a
finding three steps back is the key that unlocks the next move. To handle that, the coach treats
notes.md as a living state file — the "room brain" — that it re-reads before every
recommendation and rewrites after every result. It tracks: the goal, the kill-chain position
(recon → foothold → lateral/pivot → privesc → domain), confirmed facts, every credential/key
held and where it works, hosts discovered (reachable or pivot-only), open leads, and dead ends.
That's what lets it reason about logistics — "we hold web-admin creds, we're at foothold, and
there's a 3306 we haven't touched; those creds are probably reused on the DB" — instead of just
reacting to the last command. /notes scaffolds the brain's structure; the coach keeps it
current. On complex rooms, the quality of the brain is the quality of the coaching.
Enumeration commands split their steps into cheap (always worth it — read the page source,
fingerprint the stack, list shares) and heavy (directory brute-force, nikto, full UDP sweep,
mass table dumps). Before any heavy step, the command — and the coach, when choosing what to
recommend — applies the observe→hypothesize gate: given what we already know, would this scan
find anything new, or just churn? If you've already read the full source of a bespoke
single-page app that names its own /api endpoints, a 200k-word ffuf brute-force and a nikto
sweep are low-yield — so they're skipped with a one-line reason and offered as opt-in, and the
command points you at the move that actually pays off instead. Heavy scans are opt-in for when
cheaper observation left a real gap, never the default. (This is the fix for the classic failure
of firing every checklist step "for completeness" against a target you already understand.)
The kit ships with a small installer, install.sh, that clones this repo into a target
directory (the repo is the single source of truth — nothing is bundled inside the script), then
offers to install the tooling (apt + pipx — works on Kali, Ubuntu, or Debian).
curl -fsSLO https://raw.githubusercontent.com/pashki975/thm-claude-kit/main/install.sh
bash install.sh # installs into ./thm-claude-kit by default, prompts about tooling
Options:
bash install.sh --dir PATH # choose the install directory (default: ./thm-claude-kit)
bash install.sh --tools # install the apt/pipx tooling without prompting
bash install.sh --no-tools # skip the tooling prompt entirely (kit files only)
bash install.sh --help # show all options
It's safe to re-run: if the target dir is already a checkout of this repo, it does a git pull to bring
you up to date instead of re-cloning. Your rooms/ folder and its contents are git-ignored, so
your scans, notes, and flags are never touched.
You need Node.js 18+ and Claude Code first (the installer checks and warns if missing):
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo bash -
sudo apt install -y nodejs
npm install -g @anthropic-ai/claude-code
If you let it (--tools, or answer yes at the prompt), it installs:
# Core + Linux
seclists enum4linux smbclient hashid john hashcat rlwrap exploitdb
# pcap / WordPress
tshark wpscan
# UDP + database + pivoting
snmp onesixtyone tftp-hpa proxychains4 chisel mariadb-client postgresql-client redis-tools
# Windows / AD
netexec impacket-scripts evil-winrm enum4linux-ng freerdp2-x11 ldap-utils
# via pipx
certipy-ad bloodhound sshuttle
Package names vary a little by distro/version. On Kali these are all in the default repos. On Ubuntu/Debian most are too, but a few security tools (netexec, impacket-scripts, evil-winrm, seclists, exploitdb) may need the universe repo,
pipx, or a manual install — the installer warns and continues rather than failing on any that aren't found. Ifnetexecisn't found it may becrackmapexec(cme), orpipx install netexec; the commands usenxc.ligolo-ngandmongosharen't apt packages — grab them from their releases. Some tools you download onto the target at exploit time (winPEAS, linpeas, pspy, PrintSpoofer, GodPotato, chisel/ligolo agents) — keep them in atools/folder to serve over HTTP.
cd thm-claude-kit
claude # authenticate on first launch
The .claude/ folder lives at the install root, so every room folder under rooms/ inherits
the commands, agents, and skills automatically. Configure once, use everywhere.
cd thm-claude-kit
./new-room.sh blue # creates rooms/blue/ with scans/ + notes.md
cd rooms/blue
claude
All room folders live under rooms/, which is git-ignored — so your scans, notes,
cracked creds, and flags never get committed. The shared .claude/ config at the repo
root is tracked, and rooms still inherit it because rooms/ sits underneath it.
Point Claude at the room. /start is your playing coach for the whole room:
/start <paste the room's description + task text here, plus the target IP>
Paste the room text, don't pass a URL. TryHackMe rooms are behind auth, so Claude can't fetch the page — give it the room's description/task text and the target IP directly.
It does the setup (VPN check, /etc/hosts if there's a domain, seeds notes.md), classifies
the room, names the goal (flag vs shell), then lays out the game plan and coaches you into the
first step — explaining why it's first and which command to run. From there it guides you one
step per turn: you run each command and paste the output, it teaches you to read it and points
you at the next move. It stays light when you're flowing and drives harder when you're stuck.
You can still run the individual commands below yourself at any time; /start is the guided
way through.
The coach tags its guidance so it stands out from command output in the chat: 🎯 COACH (what
to do now), 💡 WHY (the reasoning), 🧭 READ (how to interpret output you pasted),
/vpn-check is folded into /start's setup, but you can also run it standalone:
/vpn-check 10.10.123.45
Checks tun0 is up, shows your VPN IP (you'll need it for reverse shells), and pings the target.
/recon 10.10.123.45
/enum-udp 10.10.123.45 # UDP hides SNMP/TFTP/DNS — run it alongside the TCP scan
Full port sweep → targeted service scan on open ports → summary table → notes.md updated → suggested next steps. Then let a specialist parse it:
use the recon-parser agent on the latest scans
Pick the command matching what's open (and what the room is about):
/enum-web http://10.10.123.45
/web-recon futurevera.thm # vhosts/subdomains/cert SANs — when a domain is in play
/enum-smb 10.10.123.45
/db-enum 10.10.123.45 mysql # or mssql/postgres/mongo/redis
/steg ./downloaded-image.png # when the room hands you a file to analyze
Check versions against known vulns:
use the cve-researcher agent on the services in notes.md
/listener 4444
Gives you LHOST, copy-paste payloads (bash/python/nc/PowerShell), the listener command to run in a separate terminal, and TTY-upgrade steps for after you land the shell.
/linux-privesc # walks the checklist: sudo -l, SUID, caps, cron/pspy, cred hunting
Run the enum it lists, paste the output back, then:
use the privesc-advisor agent on this output: <paste>
If a second internal host appears, tunnel to it and keep going:
/tunnel 172.16.0.0/24 # ligolo-ng / chisel / sshuttle / SSH forwards
Record each internal host in notes.md as a new target and re-run the methodology against it.
/notes # keep notes.md current anytime
use the report-writer agent # generates writeup.md when solved
Just ask:
what next?
The thm-methodology skill kicks in and gives you the next methodical step — usually "you
skipped an enumeration path" (often UDP), which is true more often than not.
Windows boxes are a different game (see WINDOWS-CHEATSHEET.md for the Linux→Windows
mental-model table). Claude auto-loads the windows-ad-methodology skill when it sees a
Windows/AD target, so what next? gives Windows-appropriate steps. Recognize an AD box by
open ports 88 (Kerberos) + 389/636 (LDAP) + 445 (SMB).
/vpn-check 10.10.123.45
/recon 10.10.123.45
Note the domain name and DC hostname. Add the DC's FQDN to /etc/hosts pointing at the
target — Kerberos tooling fails without name resolution.
/enum-ad 10.10.123.45
Null/guest sessions, shares, RPC, LDAP, and a harvested user list (users.txt).
/kerberos 10.10.123.45 SPOOKY.THM users.txt
AS-REP roasting needs no creds — it finds users with pre-auth disabled and hands you a hash
to crack. Also try password spraying, reading anonymous shares, and /db-enum if MSSQL is up.
Build the picture and let the AD specialist plan the path:
use the ad-attack-advisor agent — run BloodHound with user:pass and plan the route to DA
Then Kerberoast service accounts:
/kerberos 10.10.123.45 SPOOKY.THM svc_user:Password123
/win-shell 10.10.123.45 svc_user Password123
Picks evil-winrm / psexec / RDP based on what the cred allows.
/win-privesc
Run the enum it lists, paste output back, then:
use the win-privesc-advisor agent on this output: <paste whoami /priv + winPEAS>
SeImpersonatePrivilege → a Potato attack is the most common THM Windows privesc — the
advisor gives you the exact PrintSpoofer64.exe line.
Domain rooms pivot constantly — from the compromised host, /tunnel to internal targets and
reuse domain creds across the range (one cred often unlocks many hosts). Then DCSync krbtgt
if you reach it (full domain compromise), record flags in notes.md, and
use the report-writer agent.
| Command | Argument | Does |
|---|---|---|
/start |
pasted room text + IP | Playing coach: setup + teach you through the room step by step |
/vpn-check |
target IP | Verify VPN up + target reachable, show your tun0 IP |
/recon |
target IP | Staged nmap (all ports → service scan), summarized |
/enum-udp |
target IP | UDP top-ports + follow-up; full sweep gated to when it'll pay off |
/enum-web |
URL | Reads source first; ffuf/nikto gated — skipped when low-yield |
/web-recon |
domain or IP | Vhost/subdomain fuzzing + cert SANs + takeover check |
/wp |
WordPress URL | wpscan enum → plugin CVEs / creds / wp-config loot |
/enum-smb |
target IP | Null/guest share listing first; heavy enum gated |
/db-enum |
IP type [creds] | Connect + read schema first; brute/RCE steps gated |
/steg |
file or dir | Offline file/image analysis (metadata, embedded, LSB) |
/pcap |
.pcap file or "live " | Network forensics: creds/files/attacks/exfil; gated digs |
/listener |
port (opt) | Reverse shell payloads + listener + TTY upgrade steps |
/linux-privesc |
— | Linux local privesc checklist + analysis |
/tunnel |
subnet/target (opt) | Pivot in: ligolo-ng / chisel / sshuttle / SSH forward |
/crack |
hash or file | Identify hash, pick john/hashcat, crack it |
/notes |
— | Update notes.md from scans + conversation |
/enum-ad |
target IP | Windows/AD: SMB, LDAP, RPC, user harvesting |
/kerberos |
DC-IP DOMAIN [users/creds] | AS-REP roast + Kerberoast |
/win-shell |
IP user pass/hash | Get a session: evil-winrm / psexec / RDP |
/win-privesc |
— | Windows local privesc checklist + analysis |
| Agent | Invoke with | Does |
|---|---|---|
| recon-parser | "use the recon-parser agent" | Structures raw scan output into findings |
| cve-researcher | "use the cve-researcher agent" | Maps versions → CVEs, checks searchsploit |
| privesc-advisor | "use the privesc-advisor agent" | Ranks Linux privesc vectors from enum output |
| report-writer | "use the report-writer agent" | Builds writeup.md from notes + scans |
| win-privesc-advisor | "use the win-privesc-advisor agent" | Ranks Windows privesc (tokens, services, kernel) |
| ad-attack-advisor | "use the ad-attack-advisor agent" | Plans AD attack path from BloodHound/enum |
Skills load automatically. thm-trainer is the decision loop that runs first on every
room (classify → observe → hypothesize → test → decide). Then a methodology skill supplies the
ordered steps for the room type — thm-methodology (Linux), windows-ad-methodology (Windows/AD),
or network-forensics (packet-capture rooms) — routing to the UDP, web-recon,
database, steg, privesc-checklist, and pivoting commands as needed. You don't invoke skills
manually — Claude picks them and they drive what what next? recommends. When Claude starts
wandering, the phrase that snaps it back is: "you're churning — reclassify." /start is the
coach that walks you through this loop out loud; the skill is what it coaches from, and it also
applies when you ask what next? without going through /start.
- Edit any command — they're just markdown prompt templates.
$ARGUMENTSis replaced by whatever you type after the slash command. Change wordlist paths, add flags, whatever fits your style. - Add a command — drop a new
.mdin.claude/commands/. The filename becomes the command name. Thedescription:in the frontmatter shows in the/menu. - Add an agent — drop a new
.mdin.claude/agents/. Keep thetools:list minimal (only what it needs) so it stays focused. - Per-room overrides — a
.claude/folder inside a specific room folder overrides the shared one for that room only.
The kit is laid out so you can safely put it in a git repo and share/back up your config without ever leaking target data.
- Tracked (safe to commit/push):
README.md,CLAUDE.md,WINDOWS-CHEATSHEET.md,new-room.sh, and the whole.claude/folder — your commands, agents, and skills. - Ignored (never committed): everything under
rooms/. That's where every box's scans, notes, cracked creds, flags, and writeups live..gitignoreexcludesrooms/*but keeps the emptyrooms/directory itself (via.gitkeep) so it survives a fresh clone.tools/and*.logare ignored too.
To start tracking it:
cd thm-claude-kit
git init
git add .
git commit -m "THM Claude Code kit"
# verify no room data is staged:
git status --ignored # rooms/<yourbox>/ should appear under "Ignored files"
Because rooms live under the repo root where .claude/ sits, they still inherit all the
config — git-ignoring them changes nothing about how the tools work, only what gets committed.
- Wordlist paths assume
seclistsis installed at/usr/share/seclists/. Adjust if yours differ. - Permission prompts: Claude Code asks before running commands by default. That's a good
thing here — it keeps scope in check. If you run with
--dangerously-skip-permissions, only do it in a throwaway VM and understand you're auto-approving arbitrary commands. - GUI tools (Burp, Wireshark) aren't driven by Claude Code — it's CLI-only. It can use
tshark,mitmproxy, or Burp's REST API instead. - The listener runs in its own terminal, not inside Claude Code —
nc -lvnpblocks, so keep a spare terminal open for the catch. - UDP + pivoting are the two most-missed paths — if a Linux box has you stumped, check
/enum-udpoutput, and if the network has more than one host,/tunnelin. - Reuse creds everywhere — SSH/SMB/web/db password reuse is rampant on THM.
/db-enumfindings in particular tend to unlock other services. - Keep notes.md honest — the agents (especially report-writer and privesc-advisor) read it, so the better your notes, the better their output.
Happy hacking — stay on scope.