x3

Peak Intelligence

A Discord LLM roleplay and utility bot

Note

This is not meant to be selfhosted, please add the bot instead: https://discord.com/oauth2/authorize?client_id=1307635432632094740

Or on Matrix (E2EE supported): @x3_bot:matrix.org

Features

• SillyTavern character cards /persona card:url
• Character card creator and editor /personamaker new, /personamaker edit
• Turn impersonate akin to SillyTavern /impersonate
• Server and DM quotes /quote [...], x3 quote
• Image generation (powered by stablehorde) /generate
• Impromptu image generation with personas (akin to SillyTavern) /persona images:true
• Image editing utilities x3 say
• Additional Markov Chain, Eliza and ALICE algobots /persona model:Markov Chain
• Message regeneration & prefill /regenerate
• Per-channel image and text blacklist /blacklist
• In-chat context control (forget last n messages, all previous messages) /lobotomy
• Web-based model.json editor GUI
• Furry
• Works in DM, app and server contexts
• Extensive LLM model selection (Kimi K3, GLM-5.2, Gemini 3.7, DeepSeek V4, etcetc)
• Supports vision models, and automatically generates text descriptions for text-only models
• It has an extremely overcomplicated internal model router - it constantly juggles between vision and text models in a conversation (balancing between vision quality and response style consistency), in a very configurable manner
• It is optimized for the lowest running cost possible, aiming for maximum cache hits and routing to the cheapest providers
• Has a self-maintained toolcalling format that works even on providers that don't support the OpenAI one
• Aims to be compatible with every OpenAI API, even those with obscure nonstandard parameters (looking at you DeepSeek)
• Subpar $\LaTeX$ rendering
• Models can run websearches (grounding) and discord server searches
• /antiscam to patch up Discord's incompetent spammer problem (if you're a mod)
• Automatically preserves longer context history by summarizing previous messages
• Ability to export and import conversations /chatlog export, /chatlog import, /lobotomy
• Can render & embed HTML/SVG blocks (like SillyTavern's frontend does, but in Discord) with Gotenberg
• Has a small local embeddings model that determines whether the model should respond after an interaction for some time (10min by default)
• No need to write complex system prompts, just steer the behavior to your liking with /context commands (add, edit, remove)
• /site: a real-time infinite website generator experiment. As you click on links, the LLM continues to generate new pages
• An integrated code interpeter tool and discord utility. LLMs can run Python code, create and attach images and files in their temporary filesystem
• Free and Discord based, add it
• Also is a Matrix bot: @x3_bot:matrix.org

Matrix bot

Matrix support uses mautrix's pure-Go E2EE backend via the goolm build tag.

go build -tags goolm -o x3

For all local terminal builds without passing -tags each time, set Go's user-level default:

go env -w GOFLAGS=-tags=goolm

Set X3_MATRIX_ENABLED=true and the X3_MATRIX_* values from .env.example. If X3_MATRIX_ACCESS_TOKEN is empty, x3 logs in with X3_MATRIX_USERNAME/X3_MATRIX_PASSWORD and creates or reuses a dedicated Matrix device from X3_MATRIX_CRYPTO_DB. The Matrix UX uses text commands such as !x3 persona, !x3 chat, !x3 context, !x3 lobotomy, !x3 regenerate, and !x3 chatlog export.

For encrypted rooms, keep X3_MATRIX_DEVICE_ID and X3_MATRIX_CRYPTO_DB stable across restarts. To verify the headless bot device, set X3_MATRIX_RECOVERY_KEY to your Matrix recovery key before starting x3. The bot will import your cross-signing keys and sign its own device on startup.

Note

Currently LaTeX rendering in Matrix is only supported by a couple clients; develop.element.io, SchildiChat and Nheko should work with the experimental feature enabled.

Smart continuation triggers

x3 can optionally use MiniLM locally to determine whether a message in a server is a continuation of the previous topic. MiniLM is ~20MB in size and can run practically realtime on my RPi 4's CPU.

The ONNX model is loaded from disk but not included in the repo. Download a Sentence Transformers ONNX export, for example:

cd models/minilm
curl -L -o models/minilm/all-MiniLM-L6-v2.onnx https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/onnx/model_qint8_arm64.onnx

Set X3_MINILM_MODEL_PATH if you use another filename. ONNX Runtime must also be available as a shared library; set X3_MINILM_ONNX_RUNTIME_LIB or ONNXRUNTIME_LIB_PATH to its full path when it is not discoverable (But if you use the Alpine docker image, it already installs it for you, so you only need to fetch the model).

Run in Docker (or Podman)

podman build -t x3 -f Dockerfile .
podman run -d -v /path/to/your/containers/x3:/bot x3

To build a Discord-only container image without Matrix support:

podman build --build-arg GO_BUILD_TAGS= -t x3 -f Dockerfile .

Sandboxed code interpreter

Install gVisor and verify that Docker can use the runsc runtime:

docker run --rm --runtime=runsc hello-world

Build the interpreter image:

docker build --pull --no-cache -t x3-code-sandbox:latest codeinterp/sandbox

Enable it in .env:

X3_CODE_INTERPRETER_ENABLED=true
X3_CODE_INTERPRETER_EXECUTABLE=docker
X3_CODE_INTERPRETER_IMAGE=x3-code-sandbox:latest
X3_CODE_INTERPRETER_RUNTIME=runsc

Resource-limit settings are listed in .env.example.

Treat the runtime installation as part of the sandbox boundary:

• Keep gVisor on its latest production release and rebuild the sandbox image regularly. The interpreter deliberately uses --pull=never; builds and image updates must be performed by an administrator, never during an untrusted run.
• Configure runsc with its default systrap platform (or a separately reviewed runsc-* Docker runtime profile). Do not map the configured runtime name to runc, enable debug/strace logging, use host networking, or expose host mounts.
• Prefer a dedicated, rootless Docker daemon or a separate worker host. Anyone who compromises a process with access to a rootful Docker socket effectively controls that host, independently of gVisor.
• Pin X3_CODE_INTERPRETER_IMAGE to an image digest in production. Confirm the runtime registration with docker info, and smoke-test the rebuilt image before enabling the feature.