AI Horde, multiple users and sharing your KoboldCpp
AI Horde is a volunteer network where people with GPUs generate text and images for others and earn kudos that buy priority for their own requests. KoboldCpp has a built-in Horde worker: enter your Horde API key, a worker name and a model name in the launcher's Horde Worker tab, or use --hordekey, --hordeworkername and --hordemodelname. To share with friends instead, use a tunnel with a password; KoboldCpp queues their requests one at a time unless you add parallel slots. We tested the sharing options locally but did not run a worker.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
On this page
What AI Horde is, and the KoboldCpp Horde worker
The AI Horde is a volunteer network run by the Haidra project (AGPL software): people with GPUs, the workers, generate text and images for others, the requesters, and earn kudos that give their own requests priority. KoboldAI Lite and SillyTavern can be requesters; KoboldCpp can be a worker. If you only want to use it without a GPU, open Lite online (KoboldAI, Lite). For accounts and rules, use the Horde’s own site, aihorde.net.
On 14 November 2026 its public status page showed 46 text workers, 44 queued text requests and about 32,000 tokens generated in the past minute, with no maintenance mode F96. That is a snapshot; it changes by the minute.
Run a Horde worker from KoboldCpp
Register on the Horde
Create an account and an API key on the Horde’s site. Keep the key private.Fill in the worker settings
The launcher’s Horde Worker tab has the model name, generation length, maximum context, API key and worker name. The flags are Your AI Horde API key; setting it (with names) starts the embedded Horde worker.Default Launcher Horde Worker tab: API Key (If Embedded Worker)Since v1.65All details, Name of your AI Horde worker.Default Launcher Horde Worker tab: Horde Worker NameSince v1.65All details, Model name your AI Horde worker advertises.Default Launcher Horde Worker tab: Horde Model NameSince v1.65All details, Largest context your Horde worker accepts (0 = same as --contextsize).Default 0Launcher Horde Worker tab: Max ContextSince v1.65All details and Most tokens your Horde worker will generate per job.Default 0Launcher Horde Worker tab: Gen. LengthSince v1.65All details (since 1.65; the embedded worker is older).Start it and watch the console
Jobs appear in the console while it runs. Hides prompts and generated text from the console (auto-on for Horde workers).Default falseLauncher Quick Launch tab: Quiet ModeSince v1.51All details hides prompts from it.Check your worker on the Horde's site
Its pages show your worker and kudos.
Since 1.124 a worker can serve chat-completions jobs (release notes). Disables SSL certificate verification for outgoing connections (e.g. the Horde worker). Insecure.Default falseLauncher Network tab: NoCertify Mode (Insecure)Since v1.59All details exists for SSL problems but turns certificate checks off: avoid it. We did not run a worker: it needs an account and would serve strangers’ prompts, so we have no job or kudos numbers.
Kudos and what to expect
Kudos are the Horde’s points: you earn them by generating for others and spend them for priority. We measured none, and we promise nothing; the Horde’s pages say what they are worth today.
Privacy: strangers’ prompts on your PC
Sharing your KoboldCpp with friends
The recipe: a tunnel and a password (tunnel, password), then limits: Per-IP rate limit: each IP may start a new generation only every N seconds (0 = off).Default 0Launcher Network tab: IP Rate Limiter (s)Since v1.99All details (seconds between generations per address) and Maximum number of requests allowed to wait in the queue (default 10).Default 10Launcher Network tab: Queue LimitSince v1.123All details (how many may wait). Cap image size with Shared-server safety: caps image steps and size (optional max side length, min 512).Default 0Launcher Image Gen tab: Clamp Resolution (Hard) / (Soft)Since v1.65All details. For remote model switching there is admin mode (Enables admin mode: models/configs can be unloaded and swapped at runtime from the UI or admin API.Default falseLauncher Admin tab: Enable Model AdministrationSince v1.83All details, Folder of .kcpps/.kcppt/.gguf files that admin mode may switch between.Default Launcher Admin tab: Config Directory (Required)Since v1.83All details, Password protecting admin functions (strongly advised on public instances). Also via KCPP_ADMINPASSWORD.Default noneLauncher Admin tab: Admin PasswordSince v1.83All details, 1.83) and router mode (Reverse-proxy router that swaps to the model named in each OpenAI-API request, then answers it. Requires --admin.Default falseLauncher Admin tab: Router ModeSince v1.110All details, 1.110), which loads the model named in each OpenAI request (not tested by us).
How KoboldCpp handles several users
There is always a queue (the old --multiuser flag is deprecated, upstream). By default the second request waits for the first F26. We tested the limits on our CPU F96:
| 1 | 200: ran at once | 1.9 |
| 2 | 200: waited for the first | 3.4 |
| 3 | 503: Server is busy; please try again later. | 0.0 |
| 4 | 503: Server is busy; please try again later. | 0.0 |
With Number of generation slots that can run at once (experimental continuous batching). Default 1.Default 1Launcher Network tab: Parallel Requests / Parallel Total ContextSince v1.113All details 2, two requests run together; on our CPU that raised total throughput from 18.8 to 31.1 tokens per second with automatic threads F75. The server keeps one shared context pool: with a 4096 context and 2 slots, the log showed one pool of 4352 cells, and Total KV budget in tokens shared by all parallel slots; defaults to --contextsize.Default 0Launcher Network tab: Parallel Requests / Parallel Total ContextSince v1.124All details 8192 changed it to 8448 F96. How a slot’s share of the pool is divided we did not test. We have no GPU numbers.
Multiplayer stories in Lite
Hosts a shared Kobold Lite session that several users can join and edit together.Default falseLauncher Network tab: Shared MultiplayerSince v1.79All details (1.79) lets several people edit one shared Lite session. We tested the server side: the status route returned a turn counter, a story saved with setstory moved the turn on, and getstory returned it F96. We did not test the Lite screen. The routes are under /api/extra/multiplayer/ (API).
Sharing GPUs between your own machines (RPC)
RPC GPU sharing over the network: disabled, connect (use remote GPUs) or host (share yours).Default disabledLauncher Network tab: RPC Mode (disabled/connect/host)Since v1.114All details (1.114) lets one KoboldCpp use another machine’s GPU over your network. That is not a multi-user feature; see multi-GPU.
Frequently asked questions
What is AI Horde?
A crowd-sourced network, run by volunteers, where people with GPUs generate text and images for others. KoboldAI Lite can use it without a GPU of your own.
How do I run a Horde worker with KoboldCpp?
Create a Horde account and API key on the Horde's site, then start KoboldCpp with --hordekey, --hordeworkername and --hordemodelname, or fill the launcher's Horde Worker tab. We did not run a worker, so we cannot give job counts.
What are kudos?
The Horde's points: workers earn them by generating for others, and requesters spend them for priority. The Horde's own pages are the place for current rules.
Is it safe to run a Horde worker?
Your PC then generates text for people you do not know, so the prompts may contain content you would not choose, and you are bound by the Horde's rules. Turn the worker off when you need the GPU.
Can several people use one KoboldCpp at the same time?
Yes, through a queue. By default one request runs and the others wait; with --parallelrequests 2 two run together. With --queuelimit 1, extra requests got a 503 busy answer in our test.
What does --parallelrequests do?
It lets that many requests run together, sharing one context pool. On our CPU it raised total throughput: see the speed page.
Is the Horde free to use?
Using it from Lite is free; kudos only decide priority. The Horde is volunteer-run, so there can be a queue: on 14 November 2026 its status page showed 44 queued text requests.