Monolith

Local AI Workbench for testing, comparing, and evaluating local LLMs on real hardware.

Monolith is a local-first dashboard and evaluation workbench for practical local AI testing. It is built around real workstation constraints: model profiles, GGUF paths, llama.cpp launchers, context scaling, prompt suites, generation speed, VRAM behavior, and repeatable model comparisons.

Real hardware. Real testing. No hype.

Project status

Monolith is alpha software.

It is currently useful as a local AI workbench for technical users, but it is not yet packaged as a general-purpose application. Expect rough edges, manual setup, and active iteration.

Current version:

alpha v0.11.12

What Monolith is for

Monolith is intended to help answer practical local AI questions:

• Which local model is actually useful on this hardware?
• How does a model behave across context sizes?
• How fast does it run with a given quant, launcher, and cache setup?
• Does it follow operational safety constraints?
• Does it hallucinate under uncertainty?
• Can it handle Linux, Docker, ZFS, config review, and coding prompts reliably?
• Which model should be used for daily work, long-context work, or agent backend experiments?

Current capabilities

Monolith currently includes:

• Local model profile tracking
• Chat/test run logging
• Prompt and response metadata capture
• Token count and speed tracking
• VRAM/performance fields
• Core local-eval prompt suites
• Context-scaling evaluation scaffolding
• Agent Lab planning/review scaffolding
• Agent backend evaluation prompt suites
• FastAPI-based local dashboard UI

Current safety posture

Monolith is designed to run locally on a trusted machine.

The repository intentionally excludes:

• .env files
• API keys and tokens
• passwords
• private SSH keys
• SQLite databases
• raw logs
• local model binaries
• screenshots with private data
• real local model configuration files

Use the sanitized examples under configs/ as templates for local configuration.

See docs/PUBLIC_ALPHA.md for current public-alpha status, limitations, and expected setup path.

Local configuration

Real local configuration files are ignored by Git.

Use:

configs/models.example.yaml

as starting points, then copy them locally to:

configs/models.yaml

Do not commit your real local config files.

Development rules

This project follows a conservative local-first workflow:

• Prefer small, reversible changes.
• Keep generated data out of Git.
• Keep real model paths, private machine details, and local benchmark data out of public files.
• Run scans before public-facing commits.
• Do not add automatic command execution without review boundaries.
• Treat agent workflows as proposal/review-first until proven safe.
• Keep llama.cpp and production launchers outside this repo unless explicitly planned.

Versioning

Monolith uses alpha milestone versioning.

Recent examples:

• alpha v0.11.12 — repo-local bootstrap helper and clean-clone validation
• alpha v0.11.11 — setup doctor hardening and runtime diagnostics
• alpha v0.11.10 — clean-clone install validation and release hardening

Next planned milestone:

• alpha v0.11.13 — guided setup wizard

Canonical local WebUI

Monolith's local development WebUI uses one fixed default address:

http://127.0.0.1:8765/

Start it with:

python scripts/run_webui.py

The default can be overridden through MONOLITH_WEB_HOST and MONOLITH_WEB_PORT, but project docs and local validation should use port 8765.

Current version is tracked in:

• VERSION
• dashboard_fastapi/app.py as APP_VERSION
• CHANGELOG.md

Before committing release metadata changes, run:

python scripts/check_release_metadata.py

This checks that version references match, the changelog title stays at the top, and the latest changelog entry matches VERSION.

Security

See SECURITY.md.

License

No open-source license has been selected yet.

Until a license is added, this repository is public for visibility and review, but reuse rights are not granted beyond GitHub's normal viewing/forking terms.