Skip to main content

ctf-flightdeck

Play CTFs with a fleet of AI agents, and stay the pilot.
One terminal console: every challenge, every agent, every flag. You step in with one key.

PyPI version Python versions Tests MIT license Sponsor Pull requests welcome

A one-minute recording of the demo, uvx ctf-flightdeck mock: the console opens on an event in progress with six agent lanes, three of which need you; one key submits a flag a lane found; a lane that ignores the hint in its challenge's description is opened and redirected with one key, and finds the flag; a lane brute-forcing a server whose source it already has is killed and started fresh; a wave of 43 challenges drops and 16 lanes run at once; the help screen lists every key

The fleet view: a board of eight challenges on the left; on the right, six agent lanes sorted by what needs you, with chips for a found flag, a lane that ignores the hint, a web lane heading the wrong way and a live box a teammate holds; under the list, the found flag with 1 to submit it and 0 not to

Try it in ten seconds

uvx ctf-flightdeck mock

That opens flightdeck's screen, the console, on scripted data: no install, no API key, no network, nothing spent. uvx comes with uv, which brings its own Python. (No uv? pipx run ctf-flightdeck mock does the same.)

What it is

flightdeck is a terminal program for playing capture-the-flag (CTF) events with AI agents. It pulls the board (the event's list of challenges), runs your own solver scripts first (the ones that need no model), starts an AI agent on each challenge in parallel, and submits the flags you approve. Each agent works in its own lane, and you watch them all on one screen. When a lane loops, stalls, overspends or ignores the hint in its challenge's description, the console tells you, and one key sends a hint, redirects it, starts a second lane beside it, kills it, moves it to a stronger model or hands you its shell.

  • Bring what you have. An API key (Anthropic, OpenRouter, OpenAI), a free model on your own machine (Ollama, MLX), a coding-agent CLI (Claude Code, Codex, or an agent that speaks ACP), a folder of solver scripts.
  • Built for CTFs. It knows points, solves, categories, 40 challenges released at once, teammates, the servers an event runs for a challenge (live instances), flag formats and counted tries. It plays events on CTFd, rCTF, noCTF and GZCTF (common CTF platforms), on another through a plugin of your own, and on plain folders of challenges.
  • Plays by the event's rules. Every stage (pull, solvers, agents, submit, live instances) is auto, confirm, manual or off, checked before anything runs.
  • Safe by default. Each lane runs in its own Docker container with no network, or, when its challenge has a live instance, a network that reaches that instance and nothing else. Challenge text is treated as hostile data, never as instructions. Each lane has a cap on the model tokens and on the dollars it may spend.
  • Hands-on or hands-off. Steer every lane yourself, or engage Autopilot: it runs the event by rule while you are away, inside a budget you set, never looser than the event's rules, and hands back when something needs a human. Runs can go headless, with a status file and a webhook that pings you.
  • Learns from each run. A report of where lanes stalled and which interventions helped, writeup drafts, and the list of challenges your solvers missed.

Status: v0.7.0 (alpha) is the newest release. On GitHub this page follows the code, which can be ahead of that release: the changelog lists what is "Unreleased", and the roadmap what comes next.

Quickstart

You need uv on macOS or Linux (Windows through WSL), and something for the agents to think with: an API key for a model API (Anthropic, OpenRouter or OpenAI) with credit on it, or a model on your own machine (Ollama, MLX). Docker, running, is recommended: each agent then works in a container of its own.

uvx ctf-flightdeck mock          # the demo from above: the console, before you set anything up
uv tool install ctf-flightdeck   # installs the flightdeck command (uv brings the Python it needs)
flightdeck                       # open the home screen, and choose "Get started"

Get started is a form on the home screen: your model API, its API key (typed out of sight and kept in a file only you can read), the models (it offers some for every API) and Docker or host. ctrl+s saves it. The home screen then offers what comes next, first choice first:

  1. Build the lane image, when your lanes run in Docker and it is not built yet: once, about 830 MB. It is what each agent's container has inside.
  2. Practice: 14 small practice challenges on your model. This spends money on your model API unless the model runs on your machine; each lane stops and asks you at $2 or 400,000 model tokens by default.

The setup form came after 0.4.0 (flightdeck --version says which you have). On 0.4.0, "Get started" asks the same at the prompt and goes on to set up an event: for the practice set, run the four commands under "The same by command" below instead.

The console opens on a welcome screen that says what you are looking at and names the three keys to begin with: Enter to open a lane to watch it and talk to it, a to approve a flag an agent found, and ? for every key, any time.

From then on flightdeck is all you type. Its home screen resumes an event you played before, starts a new one (a form: the event's name, its site, your access token for it, its VPN file), and has the practice set and the demo. Stuck? flightdeck doctor checks your setup and says how to fix each problem. The home screen's last choice, "Settings and health", shows the same checks beside your settings.

Other ways to install. pipx install ctf-flightdeck works as well (it needs Python 3.12+), and Debian and Ubuntu have a .deb: Install and first run. Already installed? uv tool upgrade ctf-flightdeck (or pipx upgrade ctf-flightdeck) gets the newest release; flightdeck --version says which you have. Your config and API keys stay as they are.

The same by command

For a script or a habit, each step has a command of its own:

flightdeck init          # the setup form's questions, at the prompt; your API key is pasted hidden
flightdeck image build   # once, for Docker lanes (about 830 MB)
flightdeck doctor        # checks everything; its last line says what to fix, or what to run next
flightdeck practice      # the 14 practice challenges, on your model

And for an event:

flightdeck start                                                    # set up an event and play it
flightdeck play ~/ctf/example-ctf                                   # an event you set up before
flightdeck play ~/ctf/example-ctf --autopilot --budget-dollars 20   # Autopilot, inside a run budget
flightdeck play ~/ctf/some-board                                    # or any folder of challenges

--autopilot needs Docker lanes and a run budget. A model on your own machine has no price, so give its budget in tokens (--budget-tokens). Autopilot loosens no rule: where event.toml says confirm, flags and live instances still wait for you (Autopilot).

Optional, once you have played: flightdeck pack install adds flightdeck's solver pack: some 800 solvers that need no model and run first on every challenge. Installing it takes uv and git.

Everything else is in the player's guide: the words flightdeck uses, models and API keys, every key the console takes, boards, your own solvers, events and teams, coding-agent lanes, Autopilot, headless runs, reports.

The console

An alert on a lane that ignores the description's hint, with numbered one-key answers under the lane list

Lanes that need you sort first, and the line under the list says why and numbers the answers. h hint, r redirect, f fork, k kill, e escalate to the next model, t take over the lane's shell, a approve a flag, Enter for one lane full screen, ? for every key. The console opens in a simple layout: one list of lanes and the keys that apply now. /layout deck shows the whole dashboard, as pictured here, and /theme changes the colours. More pictures in the screenshot gallery.

Your data stays yours

flightdeck has no telemetry, no analytics, no account, no update check and no server of its own. Nothing about you or your runs is sent to this project, ever.

It talks only to what you point it at:

  • The model API in your config. A lane's challenge and its work go to the model you chose, as with any AI tool. Choose a model on your own machine (Ollama, MLX) and nothing leaves it.
  • The CTF platform in your event profile, to pull the board and submit the flags you approve.
  • A webhook of your own, only if you set one up for notifications. It gets a short message (a lane needs you, a flag was accepted) and never a flag.

Your API keys and access tokens stay in your environment or a file only you can read, never in the config, and logs, transcripts and reports replace anything shaped like a credential. Runs are written to a folder on your machine. Sharing a result is something you do by hand: flightdeck share prints a summary and you decide where to paste it. A coding-agent CLI you plug in (Claude Code, Codex) follows its own maker's policy. Details: Security.

Leaderboard

Events flightdeck has flown in, where the team finished, and how hands-on the players were: the leaderboard. Played an event with it? Add your flight and your GitHub name goes on the board.

Help make it better

Three ways, and the first needs no code:

  1. Play and share what you ran. This is the most useful thing you can do. After any run (the practice set counts), flightdeck share latest prints a flight card: your models, your setup, solves, time, cost and how often you stepped in, with no flags and no challenge text. Post it. Enough cards tell everyone which model is fastest, cheapest and best for which kind of challenge, on which hardware.
  2. Build. Fix a bug, add a feature, improve a doc. Good first issues are small and well described, and CONTRIBUTING.md gets you set up in five commands.
  3. Sponsor. flightdeck is free and MIT-licensed. Measuring models costs real API money, and sponsorship pays for it: sponsor on GitHub.

A star helps other players find it.

For AI agents

If you are an AI agent reading this for a user:

  • What this is: a Python CLI (flightdeck, PyPI package ctf-flightdeck) that runs AI agents on CTF challenges under a human's control.
  • Safe to run without asking: uvx ctf-flightdeck mock --screenshot DIR (headless, writes SVGs), flightdeck --help, flightdeck doctor, flightdeck config --example, flightdeck models.
  • Ask your user first: flightdeck practice and flightdeck play spend money on a model API, and play --event talks to a live CTF platform under that event's rules.
  • Machine-readable map of the docs: llms.txt.
  • Working on the code: read AGENTS.md first. It is short and its rules are strict.

Docs

Using flightdeck: Player's guide · Model APIs · Models · Boards · Solvers · Plugins · Practice set · Screenshots · Leaderboard · Changelog · Security

How it is built: Vision · Architecture · Console spec · Decisions · Roadmap · Contributing

License

MIT. See LICENSE.

Metadata

Release files for ctf-flightdeck 0.7.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for ctf-flightdeck 0.7.0
File Size Uploaded Provenance Actions
ctf_flightdeck-0.7.0.tar.gz 1.2 MB With provenance Details for ctf_flightdeck-0.7.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for ctf-flightdeck 0.7.0
File Interpreter ABI Platform Actions
ctf_flightdeck-0.7.0-py3-none-any.whl Python 3 none any Details for ctf_flightdeck-0.7.0-py3-none-any.whl

Total release size: 2.5 MB

Release files / ctf_flightdeck-0.7.0.tar.gz

Download URL ctf_flightdeck-0.7.0.tar.gz
Size 1.2 MB
Tags Source
SHA-256 checksum
How to use checksums
187381712011cd5de6db46b7dcb423fd63f629fdc1395702be05a1216a28104c
BLAKE2b-256 checksum
How to use checksums
29d9f93636cc43ba40bdceef646e946d9f0d4dd6e03f9787b523d5e1abe10f02
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

Release files / ctf_flightdeck-0.7.0-py3-none-any.whl

Download URL ctf_flightdeck-0.7.0-py3-none-any.whl
Size 1.3 MB
Tags Python 3
SHA-256 checksum
How to use checksums
2c4251692f56b0bb02382f104a51b7459120e86c60dd9e2ecdc0ac0c874fb0a1
BLAKE2b-256 checksum
How to use checksums
7f901d79bc61e0d7fed5962966451f0e8c179bc02352e733629224e3edee9320
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

Security for ctf-flightdeck 0.7.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for ctf_flightdeck-0.7.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page