Trust me I played
FarCry 4: 10+ times
FarCry 3: 5± times
FarCry 2: 3 times
FarCry 5: <2 times
FarCry: 1 time
FarCry Primal: 1 time
FarCry Blood Dragon: <1 time
FarCry 6: couldn't even complete
I think it might be unpopular opinion but by far FarCry 4 is the best one. Yours ?
Browsentic a chromium (Google Chrome, Edge or Brave) extension that loopback to a daemon (installed along with the extension) and connects to your AI harness you already use (claude, codex or agy at the moment). It is FREE*,* opensource*, no* account*, no* telemetry & MIT license.
https://github.com/imshaikot/browsentic
Browsentic is the last thing you would want to add your agentic browsing flow and human in-the-flow automation, I wouldn't want to claim that strongly but try it yourself - you'll see.
Key capabilities
Many more features and if there is any specific feature missing, just ask in the side-panel chat window to your AI agent to create one ;)
- Unlike Claude in Chrome it supports other harness and runs everything in your machine
- Unlike playwright extension, it is a human-in-the-loop bi-directional agentic browser extension (even for non tach-savvy users)
- Unlike browser-use, browsentic runs on your logged in browser (no new browser)
Docs: https://browsentic.com/docs
Install: https://browsentic.com/install
Two or three people, each in their own browser, playing the same arcade game's built-in co-op together. Every player runs their own emulator and only inputs are synced over WebRTC, so there is no server in between: all machines compute the same frames in lockstep at about 1.8 KB/s per peer, with no video stream and no display latency.
It ships no games and downloads none. You load a romset you are entitled to use, and a guest joining mid-game receives the host's session directly from the peers.
Play: https://imshaikot.github.io/retro-games-p2p-multiplayer-coop/
Source: https://github.com/imshaikot/retro-games-p2p-multiplayer-coop
I built a browser app where 2–3 people play CPS-1/CPS-2 and Neo Geo arcade games together. There's no video streaming: every peer runs the same emulator (FBNeo compiled to WASM) and we synchronise inputs only - deterministic lockstep. Frame N runs only after every player's input for N has arrived, so all machines compute an identical game for about 1.8 KB/s per peer. The broker (PeerJS cloud) does signalling and then leaves the path entirely.
The WebRTC lessons, in order of how much they cost me:
1. reliable: false does not mean unreliable. PeerJS only relaxes ordering - it never sets maxRetransmits/maxPacketLifeTime, so SCTP still retransmits until delivery. I log the negotiated channel properties at open to prove it rather than trusting the docs. If your design needs genuinely lossy delivery, measure what you actually got.
2. Design for reordering instead of retransmitting. The input channel is unordered, so every packet just carries the last 12 frames of input (24 bytes of masks). Nothing is ever retransmitted at the application layer - a lost or late packet is repaired by the next one ~16 ms later.
3. PeerJS cannot renegotiate, and an audio slot you never opened is gone forever. A call established with no track negotiates a recvonly m-line that can never carry audio afterwards. Worse: getSenders() on that connection hands you a phantom sender that accepts replaceTrack() and reports success while transmitting nothing. I establish every call with a silent placeholder track and verify the transceiver direction before trusting a sender. Mute/unmute is then replaceTrack on a sender that already exists - no renegotiation ever needed.
4. Glare is solved by convention. One media connection per pair, dialled by the lower peer ID. Both sides agree without talking; nobody double-dials.
5. Don't wait for ICE to tell you a peer died. In lockstep, one silent peer freezes the whole room. A bye control message drops a peer instantly on deliberate exit; a stall timeout (~2.5 s) is the backstop for the tab that got killed. Waiting for connection-state events costs several frozen seconds.
6. requestAnimationFrame stops in backgrounded tabs. Under lockstep, a backgrounded host stops publishing input and freezes everyone else. The frame clock runs off an AudioWorklet tick instead; rAF is just a second pump for when the tab is visible.
7. Data channels are fine for bulk. Savestates and the game file travel over the same channels in fingerprinted 16 KB chunks when someone joins mid-game - paced so the transfer never stalls a frame.
Honest limitations: default STUN only (symmetric NAT needs your own TURN - it's one env var), rooms cap at 3, and it ships/downloads no games - you load a romset you're entitled to use.
Demo (60 s, game pixels masked) and an architecture diagram are in the README.
Don't ask me how to get game ROMs ;)
Play: https://imshaikot.github.io/retro-games-p2p-multiplayer-coop/
Source: https://github.com/imshaikot/retro-games-p2p-multiplayer-coop
Happy to go as deep as anyone wants on the lockstep netcode or the channel behaviour.
