llm-cockpit 0.5.7

Multi-user web cockpit for Ollama: dashboard + chat + admin, pip-installable in one process

llm-cockpit

A local-first, multi-user web interface for Ollama: a dashboard for what's loaded and how it's behaving, plus a Claude-shaped chat / code UI you can pip install and have running in five minutes.

The cockpit assumes you already have Ollama running. It does not install, manage, or supervise Ollama — it talks to it.

Status

v0.5.7 beta — pre-1.0 release hardening. The core local cockpit is implemented and the current sprint is closing UI/release-readiness gaps before the first public PyPI v1.0.0. Do not treat the package as public-1.0 until the UC-11 release checklist, TestPyPI/dry-run, and Neuroforge install smoke have passed. See docs/process/SPRINT_STATE.md and docs/specs/functional/UC-11-pypi-publish.md.

What it does

• Dashboard with placement board. Kanban-style zones — GPU 0, GPU 1, …, Cross GPU, On Demand. Admin drag-drops model cards to shape what's warm where; non-admin sees the board read-only. Each card is intentionally compact: 30-day calls, cold-load time, single-GPU and tensor/multi-GPU tokens/s, single-GPU and tensor/multi-GPU context, and a temperature-backed heat signal. "Load model" searches the Ollama registry and downloads without leaving the page. GPU panel is optional (nvidia-smi).
• Chat. Pick any chat-tagged model from your Ollama install and have a streaming conversation. Per-user history, per-conversation system prompt, code-block highlighting.
• Code. Same shell as Chat, filtered to code-tagged models, with a coder-default system prompt and diff rendering.
• Admin (user management). Add / delete users, set roles on a chat < code < admin ladder, reset passwords. Force first-login password change for any seeded or admin-created account.
• Admin (Ollama configuration). Sortable model-management table with tag, placement, keep-alive, performance metrics, per-model test/delete, and sequential "Test all models" progress/ETA. The page also contains the tagging-heuristic editor, code-mode default system prompt, per-model metrics drill-down, and full audit log.
• LAN access. Installer asks whether to bind to 127.0.0.1 only or 0.0.0.0, so phones / tablets / other laptops on the same LAN can use the cockpit without a reverse proxy. HTTPS is out of scope for v0.1; for off-LAN access use a VPN (Tailscale / WireGuard) or a TLS terminator.

Quick start

# 1. Have Ollama running (https://ollama.com/download)
ollama serve   # or: systemctl --user start ollama

# 2. Install the cockpit
# Public PyPI install is gated on v1.0.0. Until then, use a local wheel
# or a tagged GitHub install from the private/release repo.
pipx install dist/llm_cockpit-*.whl

# 3. Bootstrap (probes Ollama, creates admin / ollama, sets must_change_password)
cockpit-admin init

# 4. Run
cockpit-admin serve

# 5. Open http://localhost:8080  → log in as admin / ollama → change password → use.

Other planned shapes:

• pipx install llm-cockpit after public PyPI v1.0.0.
• cockpit-admin systemd-install on Linux once UC-08 Slice E is re-verified.

Roles (ADR-004)

Each user has one role on a ladder. Higher roles include lower-rung capabilities.

Role	What it can do
chat	Log in, chat with chat-tagged models, see own conversations, change own password.
code	Above + code with code-tagged models, see own code conversations.
admin	Above + manage users, configure Ollama (tags, pull/delete, defaults), see system-wide metrics + audit log.

Bootstrap seeds one user: admin / ollama with a forced password change on first login.

Repo layout

src/cockpit/                 Python package (planned shape per ADR-002 v1.1)
├── cli.py                   cockpit-admin entry point
├── main.py                  FastAPI app
├── routers/                 auth, dashboard, chat, code, admin_users, admin_ollama
├── services/                users, model_tags, metrics, audit, settings
├── ports/                   LLMChat, Telemetry        (hexagonal)
├── adapters/                ollama_chat, telemetry, fake_chat, fake_telemetry
├── models.py / schemas.py
├── migrations/              alembic
├── frontend_dist/           built Next.js static export, bundled at wheel-build time
└── default_config/          model_tag_heuristics.yaml, code_default_system_prompt.md
docs/                        mirror of the vault subset (synced at sprint review)
├── PROCESS.md, SPRINT_STATE.md
├── decisions/               ADR-001..004
├── design-principles/       DP-INDEX (inherits from AgenticBlox)
├── specs/{user,functional,test}/  UC-01..UC-12
├── architecture/COMPONENTS.md
└── STATUS.md
scripts/sync-docs-from-vault.sh

Documentation

Where	What
docs/PROCESS.md	Spec-First + 1-week-sprint discipline.
docs/architecture/COMPONENTS.md	Component map + the two ports (LLMChat, Telemetry).
docs/decisions/	ADRs. ADR-001 process; ADR-002 stack; ADR-003 public framing; ADR-004 role ladder.
docs/design-principles/DP-INDEX.md	Which AgenticBlox DPs we adopt, defer, or skip.
docs/specs/	One folder per spec type (user / functional / test).

Process

Vault is the source of truth (DP-024); docs/ is the mirror, updated at sprint review by scripts/sync-docs-from-vault.sh.

Status flow Draft → Review → Accepted → In Progress → Done → User Accepted. Implementation only starts on a Functional Spec at status Accepted. Review→Accepted and Done→User Accepted always require explicit owner approval.

Branches: feature/US-NN-short-title → develop → main. Commit prefix: [US-NN] short description.

License

MIT. See LICENSE.

Project home

This repo is the implementation. The design source-of-truth is the project hub in the Obsidian vault at 020 Projects/LLM-Cockpit/.