LychD · The Dark Arts of LLMs

A daemon that keeps your agents alive

LychD is a persistent Linux service that runs AI agents on your own machine, with long-term memory, local model engines, an operator web UI and sandboxing. Open source, MPL-2.0, no CLA.

lychd --about
what
systemd-native agent daemon
runs_on
Linux · Python 3.12 – 3.15
license
MPL-2.0 · no CLA
status
pre-alpha · loopback · single user
inference
llama.cpp | vLLM | SGLang | ExLlamaV3
memory
PostgreSQL + pgvector · Btrfs
sandbox
rootless Podman Quadlets
Pre-alpha

An honest status note

LychD runs locally, on loopback only, for a single user. The install path does not work end to end yet, and we have not published benchmarks. If you want to help get it there, the code and the conversation are both public.

How Snayu uses LychD

An open core, custom compositions on top

LychD is the platform. What we deploy for each client is a composition: an application built on LychD and designed around that client's workflows, data and rules.

LychD itself is pre-alpha, so client work is delivered as a supervised engagement: Snayu engineers install, configure and run the composition for you, and the statement of work spells out what is in scope today and what is still on the roadmap.

LychD core

  • Open source under MPL-2.0, no CLA
  • Developed in public on GitHub
  • The same core under every deployment
  • Free to read, run and fork

Client compositions

  • Designed for one client’s workflows
  • Built and deployed by Snayu on your hardware
  • The custom work is yours, per your contract
  • Supported and improved as your needs change
Architecture

Five foundations, one Lich

Each part has a name from the grimoire and a plain job description. Here are both.

LychD architecture: the Vessel runtime connected to Ghouls, Altar, Animators, Phylactery, Coffin and Soulforgeyour machine · systemdGhoulsbackground jobsAltaroperator UI · logsAnimatorslocal model enginesPhylacterymemory · PostgreSQLCoffinsandboxed jobsSoulforgetrainingVesselruntime · web server
01

Vessel & Ghouls

Async runtime and background workers

The Vessel is the app runtime and web server that holds the Lich together. Ghouls are short-lived background jobs that do the heavy lifting and then rest.

LitestarGranianSAQ
02

Phylactery & Soulforge

Persistent memory and training

The Phylactery is the PostgreSQL state store where the Lich keeps its memory. The Soulforge is where training happens.

PostgreSQLpgvectorBtrfs
03

Animators Matrix

Local inference engines

Animators are model endpoints. A Soulstone runs on your own GPUs; a Portal reaches a remote one. Swap engines without rewriting agents.

llama.cppvLLMSGLangExLlamaV3
04

Altar

Operator web UI and logs

The web console where the Magus watches over the Lich. Its instruments are Atlas, Bridge, Loom, Nexus and Orb. Structured logs are recorded today; OpenTelemetry tracing is planned.

SvelteKitStructlogOpenTelemetry (planned)
05

Coffin

Sandboxing for lower-trust jobs

Agents you trust less are sealed in a Coffin: a rootless Podman Quadlet supervised by systemd, holding only the tools you grant.

PodmanQuadletsystemd
Anatomy of a run

Follow one task through the Lich

How the parts are meant to work together on a single run, from the operator's request to the trace it leaves behind.

anatomy of a run intended design · pre-alpha
Altaroperator UIVesselruntimePhylacterymemoryAnimatormodel engineCoffinsandboxtask · Sigil1recall Karma2precedent3prompt + tools4tool call5run tool (sealed)6result7trace + Karma8answer · trace9
  1. 1 Altar → Vessel task · Sigil
  2. 2 Vessel → Phylactery recall Karma
  3. 3 Phylactery → Vessel precedent
  4. 4 Vessel → Animator prompt + tools
  5. 5 Animator → Vessel tool call
  6. 6 Vessel → Coffin run tool (sealed)
  7. 7 Coffin → Vessel result
  8. 8 Vessel → Phylactery trace + Karma
  9. 9 Vessel → Altar answer · trace
9/9

The answer shows up in the Altar, with the full trace behind it.

The stack

Boring, well-known parts. On purpose.

Top to bottom, the layers a request passes through on its way to your hardware and back.

Altar & docs Interface
SvelteKitZensical
Vessel Runtime
LitestarGranianPostgreSQL + pgvectorSQLAlchemyAdvanced AlchemyAlembicPydantic AIPydantic GraphSAQ
Animators Inference
llama.cppvLLMSGLangExLlamaV3TabbyAPI
Iron Host
LinuxsystemdPodman + QuadletBtrfs

Requirements

  • Linux
  • Python 3.12 to 3.15
  • NVIDIA GPU (assumed by the install guide)
  • A GGUF model that supports tool calls

Hardware profiles

  • Workstation1–2 × RTX 3090 / 4090 / 5090Qwen3.8 27B · Gemma 4 31B
  • Private cluster4–8 × RTX 4090 / 5090DeepSeek V4 Flash · gpt-oss-120b
  • Data centreDGX H200 / B200 / GB200DeepSeek V4 Pro · GLM-5.3 · Kimi K3

Lit squares show the minimum GPU count; outlined ones are optional. Models are current open-weight examples that fit each tier. No LychD benchmarks are published yet.

Roadmap

Metamorphosis: five books into 2027

The roadmap is written as a story in five books. The first chapter, Genesis, is next.

Book I · Chapter 1 Current

Genesis

  • A shared agentic lab
  • A phone and voice experience
  • A Minecraft showcase
  • Visual work
Enthusiast release Planned for 21 Dec 2026
Book II Sealed
Book III Sealed
Book IV Sealed
Book V Sealed
Glossary

A short grimoire

The docs speak in their own tongue. This is the translation. Hover or tap a term to see what it connects to.

The whole running system, over time.
The human operator.
The app runtime and web server.
A short-lived background job.
The PostgreSQL state store.
Governed memory: traces, Karma and provenance.
An attributable record of past actions, corrections and consent, used as precedent.
A model endpoint: a Soulstone (local) or a Portal (remote).
A validated TOML configuration file.
The configuration tree.
The identity and permission context carried through a request or run.
The sandbox profile for lower-trust agent jobs.
Training.
The web UI, with instruments Atlas, Bridge, Loom, Nexus and Orb.
A workflow definition.
An accepted architecture decision record.
An app built on LychD.
The docs.
Get started

Want a composition built for your team?

We run fixed-scope pilot projects that put LychD to work on one real workflow, with our maintainers alongside your team.