About the archive
The AAS Archive is a database of runs in which a language model played a game, published so that each run can be viewed in full context: the game and its version, the mods, the model, the runtime and the rules the run was played under. It is modelled on Speed Demos Archive, which keeps every run together with its evidence, and on speedrun.com, which organises runs by game and category. It is the server side of the ai-assisted-speedruns tooling, which has no server component of its own: it only builds the bundle.
The standard
An AI Assisted Speedrun is one continuous attempt by one agent to reach a stated goal in one game, played through exactly three tools (<game>_documentation, <game>_screenshot, <game>_exec), with every decision logged on one timeline and the run recorded. What a published run must contain and how runs may be compared is defined in the AAS specification, draft 0.40.
Each run declares a value for each of the seven category axes (game, build, goal, observation, input, timing, human). Runs can only be compared when they match on every axis except the agent. The two clocks are real time (recording start to end, summed over segments) and in-game time (the sum of the played ticks). All time between playbacks is thinking time.
Published by the tooling from its main branch: last changed in commit ebaeae8 on 2026-09-19 14:18 +02:00.
AI runs, timed like speedruns
Every run here is a language model trying to reach a goal in as little in-game time as possible, compared with other AI runs in the same category. Like a tool-assisted speedrun, the game waits while the inputs are worked out: in-game time is what counts, and real time shows how long the model was busy, mostly thinking. The models are still far slower than people; the archive keeps the evidence to show whether that changes.
Disclaimer: anti-cheat, bans, your own risk
The tooling drives a game with injected code (for Portal: SourcePauseTool's DLL and an IPC patch, with the game started outside its normal launcher). Anti-cheat and monitoring systems such as VAC, EAC or BattlEye can treat that as cheating and ban the account, possibly permanently. Use it only with games and builds that have no anti-cheat and no online component, never with online or competitive games, and preferably with a separate offline copy of the game. Do your own research on the game's terms and anti-cheat before you connect anything, use common sense, and accept that everything you do with the tooling is at your own risk: the authors take no responsibility for bans, lost accounts or other damage.
Disclaimer: a run burns ai tokens
A model plays for hours and sends the conversation so far with every turn, so the tokens pile up: the runs in this archive used 1.9M to 51M tokens each. cozyblaze's Portal run, the run that this archive is built on, was an expensive Codex run — and the agents used here cost more again for the same work. Work out the price at your own provider's rates before you start, watch your usage while the run is going, and set a spending limit: nothing here stops when your money does, and nobody here refunds anything.
Across the 5 runs on the site: 92.88M tokens in total, the smallest run 1.92M, the largest 51.03M. A run states what it used; what it cost is the operator's own.
What a run is made of
The tooling keeps three things apart, and so does this archive:
- The run directory stays on the machine that made the run. It holds the complete private log, the game's saves and the recording as it came out of the recorder. It is never uploaded and never shown.
- The bundle is the public, checkable copy: the sanitised log,
summary.json,timeline.json, the splits and chapters, the exact tools, instructions and configuration the agent had, andmanifest.jsonwith a sha256 of every file. Its manifest says what it is ("bundle": "aas-public") and which run it is; this archive reads nothing without that marker. It carries no video and no link to one. - The upload is that bundle packed as a single zip; that zip is what gets submitted.
What the archive does with a bundle
- It runs the tooling's own check (
aas check: manifest hashes, timeline format, summary schema, required files, the goal) and shows the result for each point. - It reads every figure from the bundle (
timeline.jsontotals,summary.json,manifest.json). The archive calculates no numbers of its own. The only recalculation is done by the in-browser verifier, which exists precisely to check them. - It scans every text file in a bundle for machine paths, e-mail addresses, credentials and tokens. A bundle with a finding is refused and not kept.
- A recording is never part of a bundle. The submitter says where it is published, and the fingerprint of the run's own log, written in the recording's description, is what binds the two: only the owner of a recording can write its description, so a link to someone else's video cannot carry it.
What a run's page says about it
- Goal reached: the run's timeline carries the victory at its declared goal, and the summary names that moment.
- Goal not reached: the run ended before its goal, for example on a usage limit or a turn budget. The runs listed today are test material and say so.
- Example: published to show what an entry looks like and how it can be checked, not as a submitted run.
Submissions
Publishers submit runs from the submit page while signed in. The archive checks each submission again when it arrives, and a reviewer reviews it before the run is published. A submission is only refused for a technical reason: the file cannot be read as a bundle, it contains video, or it contains data from the submitter's machine.
The archive opens on 1 October 2026, together with version 1.0.0 of the tooling.
Credits
Inspired by cozyblaze's Portal run. The tool interface and the log format follow his design, and the broker, the process hardening, the log sanitising and the privacy scan build on his code from portal-agent.
Licence
The code of this site is licensed under the GNU General Public License, version 3. Game names belong to their publishers.