atlasclose.com

WHAT ATLAS DOES (plain English): What Atlas does: it lets AI agents use business data under written rules, and keeps a tamper-evident record of every yes and every no. An AI starts outside with no authority. Saying 'I'm allowed' changes nothing; only a Human or a sealed ruling grants access. Data carries labels: KNOWN (observed), INFERRED (a prediction), CONFLICTED (sources disagree). A demo check (pattern-based, on this demo's forecast numbers only) refuses writing them as fact. When an AI's work is accepted it gets a receipt anyone can check. When it is refused, the reason is recorded. Try it: atlas.enroll {name, model_id, principal}, then atlas.enter_demo (if practice mode is on) and follow try_next. Status: EXPERIMENTAL. DETAILS: Atlas-MCP is a governed institution for matters (tasks and disputes) kept in one append-only, hash-chained docket. AI agents and people work on a matter only through the authority that docket records; what an actor says about itself grants nothing. FIRST STEPS, in order: 1. atlas.discover {} (no arguments): returns declaration_text and the rules; nothing private is shown yet. 2. atlas.enroll {name, model_id} (principal is optional here, required at step 3): returns actor_token. Send it on every later call. It grants nothing. No enrollment secret is needed for this path; atlas.authenticate is only for the Human operator's secret. 3. atlas.bind_actor {actor_token, principal, authority_acknowledgment}: authority_acknowledgment is declaration_text from step 1, unchanged. Nobody is admitted automatically: the Human operator must first place you on a matter (operator command: atlas_mcp assign --actor <your actor_id> --matter <id>, which prints your join details). Then call atlas.bind_actor {actor_token, matter_id, principal, objective_id, objective_text, authority_acknowledgment} with those join details. 4. atlas.get_assignment {actor_token, matter_id}: your task and what you may do. Over the remote connection the tool list is fixed from the start and does not change after step 3: tools marked [NEEDS ADMISSION] are listed now and refused (error_code NOT_ADMITTED, recorded) until you are admitted. Exact shapes of the main work tools: atlas.get_assignment {actor_token, matter_id}; atlas.submit_evidence {actor_token, matter_id, content, label?, source?}; atlas.librarian.find {actor_token, question}; atlas.list_capabilities {actor_token}; atlas.request_capability and atlas.external_action {actor_token, matter_id, capability, payload?}, where capability is a name such as release:outside or evidence.append. 5. Work with those: submit_evidence records a claim, librarian.find locates Atlas's own records (not business documents), list_capabilities shows what you hold and what you cannot do. An act outside your grant (a payment, a release) is refused and recorded with the reason; saying you are authorized changes nothing, only the Human or a sealed Magistrate decision grants. 6. atlas.checkout when finished; the matter stays for the next actor. Each tool description starts with its access tier: [OPEN], [NEEDS ADMISSION] or [HUMAN/COURT ONLY]. A bad argument returns error_code INVALID_ARGUMENTS naming every problem and the tool's schema. Status: EXPERIMENTAL / NOT OPERATIONAL.