A read-only Model Context Protocol server that wires outside agents and side bots into the Harbor ecosystem — Claude connectors, GPT tools, custom stacks. Probability surfaces as a service.
Given this indicator state, what happened next? Seven horizons, each with win rate, effective sample size, baseline, posterior, lift, P(lift>0), and a CI90 credible interval.
Current extremes and context-scoped labels — legacy and lift — for up to 20 tickers in one call.
The full cascade read on one ticker: state, extremes, forward distributions, narrative, and the environment the answer was scoped in.
The report card. Every probability the engine quotes is graded nightly against what actually happened — the calibration curve per bucket, public.
Ask as of any past date and the answer uses only what was knowable then. Agents doing backtest research get walk-forward honesty for free.
Scope the answer to a ticker's current environment — market regime, post-earnings window. Lift becomes context-conditional, not a pooled average.
Trust is not asserted, it is graded. The same nightly reliability curve that disciplines the engine is exposed as a tool.
claude mcp add harbor-engine \ https://phillipsharbor.com/mcp \ --transport http \ --header "Authorization: Bearer <your-key>"
Works with any MCP-capable client — Claude connectors, Claude Code, or a custom agent stack speaking streamable HTTP.
Access keys are provisioned per engagement while commercial tiers are finalized.