Seven agents, each doing one real job

Ask anything about your status — the answer is built live, not recalled.

Every question runs a real multi-step search: the agent decides which official sources actually cover it, fetches them live (checking a 12-hour cache first), and only then writes an answer — grounded strictly in what those pages actually say this run. It never answers from its own training memory, and it explicitly lists what it couldn't verify rather than filling a gap with a guess. One category — the J-1 two-year home-residency rule — is hard-coded to always redirect to the Department of State directly, because the only page that can give a real answer can't be fetched automatically, and a partial answer there is genuinely dangerous.

Every source is checked for real rule changes, not just a date stamp.

Compass keeps a snapshot of each official source it relies on. On every check, it compares the current page against that snapshot and — if anything differs — asks a model to judge whether the change is substantive (an actual rule, timeframe, or requirement) or trivial (a wording tweak, a date-stamp bump). Only substantive changes get surfaced. It also tracks which other pages depend on a given source, so a change to the OPT page flags the STEM OPT and travel pages for re-review too.

Your grace period, OPT window, and STEM OPT dates — computed, and double-checked.

Given your program end date (and OPT start date, if relevant), this agent computes every date that follows from official timeframes — deterministically, the same way every time, never an LLM's arithmetic. But it doesn't just trust its own constants: before showing you a date, it fetches the live source behind that timeframe and asks a model whether the source still agrees with the number baked into the app. If it disagrees, you see that flagged, not a silently wrong date.

Paste your I-20, DS-2019, or EAD — get back only what's actually printed on it.

Paste the text of your document and Compass extracts only what's explicitly present — document type, category, program end date, work authorization start date — never an inferred or estimated value. Every field can come back null, and null is treated as the correct answer when something isn't clearly stated, not a failure. If a usable date comes through, it feeds straight into the same deadline engine as Agent 3, so you get your actual timeline from your actual document.

Figures out whether this needs your DSO, an attorney, or just an answer.

This agent doesn't answer your question — that's Agent 1's job. It triages: is this a routine procedural question, a rights question, or something closer to an emergency? Is your DSO the right first call, or does this genuinely need a licensed immigration attorney? It's deliberately conservative about escalating to "you need a lawyer" — reserved for things with real legal consequence already in motion, not every unfamiliar process question.

See how F-1 stacks up against M-1 or J-1 on any topic, neutrally.

Curious how your category differs from another? This runs Grounded Retrieval independently for both categories on the same topic, then synthesizes only the differences — strictly from what those two answers actually said, never introducing a new claim. It never says one category is "better"; that's not a judgment this tool makes for anyone.

Composes retrieval and your timeline into one checklist before you fly.

The highest-stakes moment in the app, so it's the one built from the other agents rather than a single prompt: it pulls the real document/travel-signature requirements from Agent 1 and your actual computed deadlines from Agent 3, then cross-references your travel date against those deadlines — flagging anything landing within about a month of your trip as worth confirming with your DSO before you go.