Lighthouse Agentic

Tugboat: Daily Model Intelligence for the Council

Know which model to use, for which task, at which cost, every day.

The AI model landscape changes fast. What was the best choice last week may not be the right call today. Tugboat gives the Lighthouse Council a daily, sovereign map of the best model per task per budget, built from public benchmark and pricing data, processed locally, with no internal data leaving the box.

Learn more
What It Is

The Problem with Picking a Model by Feel

Most teams either pick one model and stick with it, or they spend real time manually comparing prices and benchmarks every time something changes. Neither is good enough when you are running a fleet of agents across multiple task types and budget tiers. Tugboat automates the map: it reads public data daily, scores each available model for each task class, and tells the Council what to route where.

The name comes from the job. A tugboat doesn't build the ship or plan the route. It guides heavy things into the right position efficiently and gets out of the way. That is exactly what this layer does: it positions the right model for each job and lets Hermes and Pilot take it from there.

How It Works

Public Data In, Routing Map Out

Tugboat reads from public sources: benchmark leaderboards, public pricing APIs, and our own internal usage ledger. It runs the scoring locally, produces a ranked map for each task class and budget tier, and writes it to a file the Council reads each day. No private context, no proprietary code, and no internal data travels outside the box to produce that map.

Input

Public Data Only

Tugboat reads benchmark results from public leaderboards, current model pricing from OpenRouter and provider APIs, and our internal usage ledger. None of these reads require sending any internal context outside the machine.

Scoring

Task-Class Scoring

Each model is scored per task class: executive reasoning, coding, summarization, fast drafting, and so on. Cost-per-token is folded in at each budget tier. The output is a ranked list, not a single recommendation, so the Council can see the tradeoffs.

Cadence

Daily, Not Monthly

A benchmark comparison you ran last month is probably stale. Models get updated, prices shift, new entrants arrive. Tugboat re-scores daily so the routing map reflects what is actually available today, not what was available when we last had time to check.

Governance

The Council Still Governs

Tugboat informs; it does not decide alone. Capability rules set by the Council govern when a model can be used regardless of score. The weekly OBS-grounded reassessment can override any Tugboat recommendation. The map is an input, not an automatic switch.

The Thinking Behind It

Sovereign by Construction

Most model-selection services work by sending your usage data, your prompts, or your outputs to an external service that scores them for you. For a system built on the principle that your data stays yours, that approach is a non-starter. Tugboat's sovereignty is architectural.

Where It Stands

Baseline-Ready, Not Yet Routing Live

Tugboat is built and running. Here is the honest current state.

Shipped

Architecture and Scoring

The sovereign architecture is built: local scoring from public benchmark and pricing data, written to a map the Council reads. The runner and scoring are built and runnable; the daily schedule activates once the cron entry is set. The fail-closed posture is in place.

Next gate

Second Corroborating Source

Active production re-routing waits on a confirmed second data source: either a benchmark API key that can produce real-time scores independently, or the internal usage ledger growing to a size that corroborates the public benchmark picture on its own. One source alone is not enough to make automatic routing decisions in production.

Architecture: built and running Daily scoring: scheduled (not yet armed) Live production routing: next gate

Tugboat is an internal infrastructure layer, not a product we offer externally. It is described here because we think transparency about what the Council actually uses to make model decisions is worth the space. If you are building something similar, reach out.

The Right Tool for the Job

Know which model. Know the cost. Keep the data home.

A tugboat does not sail to the destination. It guides the right vessel into the right channel, efficiently, and steps aside. Tugboat does the same for model selection: daily, sovereign, and out of the way once the routing is set.

Lighthouse Agentic a faith-led AI studio — AI is a tool; God is the source