Skip to content

Self-Hosted Honcho on ai-nuc

Status: running on ai-nuc as of 2026-09-13
API: http://127.0.0.1:8000 (healthy). Tailscale: http://100.73.71.29:8000
Data: /opt/stacks/honcho → /mnt/ai-data/stacks/honcho on the ai-data disk
Inference: host Ollama only. Deriver/summary/dream/dialectic = qwen3.5:9b. Embeddings = nomic-embed-text (768-d). No cloud key.

The earlier elkimek setup.sh path (clone into ~/honcho, cloud embeddings) is superseded. We used the current plastic-labs/honcho config format (model_config + embedding.vector_dimensions) and the /opt/stacks convention.

Overlay in this repo: stacks/honcho/. Deploy script: scripts/ai-nuc/deploy-honcho.sh.

What is running

Hermes (Mac) ──Tailscale──► ai-nuc:8000  Honcho API
                                 │
                                 ├── PostgreSQL + pgvector   (127.0.0.1:5432, on ai-data)
                                 ├── Redis                   (127.0.0.1:6379, on ai-data)
                                 └── Deriver worker
                                         │
                                         └── host Ollama :11434
                                              ├── qwen3.5:9b          (every-message work)
                                              └── nomic-embed-text    (vectors)
Service Container Notes
API honcho-api-1 0.0.0.0:8000, /health → {"status":"ok"}
Deriver honcho-deriver-1 Small model so it does not evict qwen3.8:27b-hermes
Database honcho-database-1 pgvector/pg15, bind mount ./data/pgdata
Redis honcho-redis-1 bind mount ./data/redis
MCP not started docker compose --profile mcp up -d publishes 8787, never 3000 (OpenWebUI)

Docker network: ai-nuc. Compose plugin lives at ~/.docker/cli-plugins/docker-compose (the distro Docker package on this box has no compose).

Why 9B for Honcho, 27B for Hermes

Honcho's Deriver runs on every message. If it loaded the 27B, it would kick Hermes off the 24 GB card. qwen3.5:9b (~6.6 GB) is the always-on memory worker. See tokens, context, and picking a model.

Point Hermes on the MacBook at this instance

Done 2026-09-13. Mac dotfiles live under ~/.config. The tracked file is:

~/.config/.honcho/config.json

Hermes itself does not read that XDG path. It looks at ~/.honcho/config.json (after $HERMES_HOME/honcho.json and ~/.hermes/honcho.json). The install script wrote the tracked file and pointed ~/.honcho at it:

# already run on the Mac; re-run only if the pointer is missing
bash scripts/macbook/install-honcho-config.sh

That copies stacks/honcho/hermes-config.json (baseUrl http://100.73.71.29:8000) and runs ln -sfn ~/.config/.honcho ~/.honcho if ~/.honcho is not already a real directory.

On ai-nuc itself, ~/.honcho/config.json still points at http://127.0.0.1:8000 (NUC-local, not part of the Mac dotfiles repo).

Model in Hermes stays qwen3.8:27b-hermes, context 64000. Honcho does not replace that. It is the filing cabinet, not the desk.

Deploy / update

~/office-plan/scripts/ai-nuc/deploy-honcho.sh

First boot creates pgvector columns at 1536 (Honcho default). The script then runs scripts/configure_embeddings.py --yes so they match nomic-embed-text (768). Do not skip that if you rebuild the database from scratch.

cd /opt/stacks/honcho
docker compose ps
docker compose logs -f api deriver
curl -s http://127.0.0.1:8000/health

Backup

mkdir -p /opt/backups/honcho
cd /opt/stacks/honcho
docker compose exec database pg_dump -U postgres postgres \
  > /opt/backups/honcho/honcho-$(date +%Y%m%d).sql

What this is not

  • Not Buzz. buzz.bryanwills.dev is a separate onboarding on the MacBook.
  • Not OpenWebUI chat history. That still lives in the open-webui Docker volume until that stack is migrated.
  • Not a 262k context window. Honcho stores observations on disk. The live window is still the GPU desk.

Known limits

  • Deriver quality is "good enough local 9B," not Plastic Labs' Neuromancer 8B. Observations will be less sharp. Data stays on the NUC.
  • USE_AUTH = false. Port 8000 is on all interfaces. Keep this Tailscale/LAN only until auth is on.
  • Ollama sees request content in the clear at inference time (function calling). Stored rows stay on ai-data.