governance-platform

governance-platform: проверка решений AI-агентов до исполнения, с вердиктами PASS/FAIL/REVIEW

От сообществаРаботаетБез входаГлобальныйБесплатноТолько чтение

Что умеет

  • Проверять и чинить блюпринты действий агента
  • Выносить вердикт PASS/FAIL/REVIEW с трассировкой решения
  • Сравнивать семантическую эквивалентность вариантов
  • Строить контрфактуальные сценарии и искать пути восстановления
  • Вести цепочки этапов и аудит передачи задач между агентами

Какие данные видит

Доступ без авторизации: сервер работает с переданными в запросе блюпринтами и данными, не хранит внешние учётные данные.

Нужен ли аккаунт

Не нужен: сервер работает без входа

MCP-сервер проверяет действия AI-агентов до их выполнения и выдаёт детерминированный вердикт с воспроизводимым доказательством. Подходит для построения цепочек проверки, анализа аномалий, восстановления после сбоев и аудита передачи задач между агентами.

Как подключить

Как подключить governance-platform к Claude

Адрес сервера, который нужно вставить в настройках Claude:

https://app.geodesiclabs.ai/mcp

  1. Откройте Claude (сайт claude.ai или приложение для компьютера).
  2. Перейдите в Настройки → Коннекторы (в английском интерфейсе: Settings → Connectors).
  3. Нажмите Добавить пользовательский коннектор (Add custom connector).
  4. Вставьте адрес сервера, скопированный ниже, в поле URL удалённого MCP-сервера и нажмите Добавить.
  5. В чате нажмите + → Коннекторы и включите новый коннектор.

Пользовательские коннекторы доступны на тарифах Free, Pro, Max, Team и Enterprise. На Free можно добавить только один. В Team и Enterprise коннектор сначала добавляет владелец организации в Настройки организации → Коннекторы.

Авторизация

После нажатия Добавить откроется окно входа в сервис. Войдите под своим аккаунтом и подтвердите доступ. Пароль Claude не видит.

Как подключить governance-platform к ChatGPT

Адрес сервера, который нужно вставить в настройках ChatGPT:

https://app.geodesiclabs.ai/mcp

  1. Откройте ChatGPT в браузере (chatgpt.com). Нужен тариф Plus, Pro, Business, Enterprise или Edu.
  2. Один раз включите режим разработчика: Настройки → Приложения → Дополнительные настройки → Режим разработчика (Settings → Apps → Advanced settings → Developer mode).
  3. Откройте Настройки → Коннекторы и нажмите Создать (Create).
  4. Заполните форму: Название (любое), Описание (одна фраза, что делает сервис), URL MCP-сервера (скопируйте ниже).
  5. В поле Аутентификация выберите OAuth, если сервис требует вход, иначе Нет. Нажмите Создать.
  6. В новом чате откройте + → Приложения/Коннекторы и включите коннектор.

Название раздела у OpenAI менялось (Connectors → Apps/Plugins). Если пункт называется иначе, ищите «режим разработчика» в настройках. В рабочих пространствах Business/Enterprise администратор должен разрешить пользовательские коннекторы.

Авторизация

При первом использовании ChatGPT откроет окно входа в сервис. Войдите и подтвердите доступ.

Как подключить governance-platform к Cursor

Адрес сервера, который нужно вставить в настройках Cursor:

https://app.geodesiclabs.ai/mcp

Быстрый способ: нажмите кнопку Открыть в Cursor ниже, Cursor сам предложит добавить сервер.

Вручную:

  1. В Cursor откройте Settings → Cursor Settings → MCP и нажмите Add new global MCP server.
  2. Вставьте JSON, скопированный ниже, в открывшийся файл ~/.cursor/mcp.json (для одного проекта: .cursor/mcp.json в корне проекта).
  3. Сохраните файл. Сервер появится в списке MCP; при необходимости пройдите авторизацию по кнопке рядом с ним.

Авторизация

Если сервис требует вход, рядом с сервером в списке MCP появится кнопка авторизации.

Список инструментов сервера (37)

Технические названия из tools/list. Нужны только разработчикам.

validateValidate structured data against a Blueprint's rules BEFORE the result is used. Returns PASS, FAIL, or REVIEW with plain-language findings, repair suggestions, a determinism hash, and a re-verifiable certificate. Same input + same rules = same verdict, every time.
validate_repairValidate structured data against a Blueprint and, when it fails, include repair suggestions (corrected values with the rule each fix is based on) in the same call. Same verdicts as validate: PASS, FAIL, or REVIEW, with reasons and proof.
create_blueprintCreate a Blueprint - the governance contract validation runs against. A Blueprint defines what correct means for your data: fields, the math that must hold between them, and acceptable ranges. Start from load_rule_pack or discover_patterns if you have no rules yet; invoke the blueprint_guide prompt for the full rule/constraint reference. Returns the new Blueprint's API key.
list_blueprintsList the Blueprints on this account with field/rule/constraint counts and mode. Use the returned workflow_name as 'blueprint' in validate.
repairOne-shot repair: return corrected values that would make failing data valid under the Blueprint. Use repair_path to see the steps instead.
check_blueprint_healthStatic pre-deploy analysis of a Blueprint's rule set. Returns a health verdict - healthy, acceptable, fragile, rigid, split, brittle_islands, or unsatisfiable - with advice, including joint conflicts pairwise checks miss.
compare_semantic_equivalenceCompare two payloads under the dual-hash design: content_hash is content_hash normalizes field order and numeric formatting. Semantic comparison preserves field roles; renaming requires an explicit bijection. With supplied rules, scalar types and whitespace remain significant. Structural similarity alone does not establish decision equivalence.
govern_inferenceQuality-govern an in-progress AI generation step BEFORE its output is used (complements validate, which checks finished documents). Returns an action - STOP, CONTINUE, REPAIR_REGION, REUSE_MOTIF, REVIEW, ESCALATE - with a plain-language explanation. Structural scores do not establish task correctness. Check safe_to_finalize and acceptance coverage. Persistence success is reported; read traces in the same Blueprint namespace.
get_inference_traceRetrieve the durable audit trail for a governed generation: every recorded decision and its reasons.
recent_inference_decisionsRecent generation-governance decisions in this owner's Blueprint version.
verify_certificateIndependently re-verify a validation certificate. Integrity mode checks the hash chain; full mode (certificate + original data) recomputes every attested rule from scratch - trust nothing, recheck everything.
profile_blueprint_robustnessSweep the Blueprint's numeric constraint bounds and report verdict stability: the stable band, the scales where the verdict first flips, and advice. Use before deploying bound changes.
forecastDeterministic forward reasoning: from the current data state, generate and rank the valid next states reachable under the Blueprint's rules.
discover_patternsLearn candidate validation rules and structural document types from a batch of your records, deterministically - no Blueprint required. Promote results with approve_rule. Source data is not stored.
repair_pathFind the shortest sequence of field changes taking invalid data to a valid state, as an ordered path of intermediate states. Different from repair (one-shot nearest fix): use repair_path to explain or audit the fix, or compare alternative repairs.
counterfactualRun the same data under two rule sets and compare which future states remain valid - what-if analysis for rule changes.
analyze_anomalyExplain whether a record fits the usual pattern for records like it, and which fields stand out. No Blueprint required.
create_chainCreate a multi-agent sequential chain: stages validate in order against one Blueprint, repairs propagate forward, TTL bounds the run. Siblings: submit_chain_stage advances the chain; handoff_audit verifies a transition between stages. Returns chain_id.
submit_chain_stageSubmit data for the chain's current stage; the platform validates it and advances the chain if it passes. Response includes next-stage info and accumulated repairs.
handoff_auditAudit a handoff between two chain stages: a context capsule of verified facts from the prior stage, and (if proposed_data is given) a compatibility verdict that catches fields mutated in transit. Siblings: create_chain, submit_chain_stage.
approve_rulePromote a rule discovered by discover_patterns into Blueprint-ready form.
reject_ruleReject a discovered candidate rule so it will not be promoted into a Blueprint. Pair with approve_rule after discover_patterns.
structural_typesRetrieve the document categories a discover_patterns session identified (counts, distinguishing fields, domain hints). Read-only; returns status=no_session if discovery has not run for this namespace.
decompose_failureSplit the error between original and corrected values into direct rule violations, boundary violations, and systemic structural error, with per-field contributions. Use with a known-correct version to diff against; use analyze_anomaly when you only have the suspicious payload. Diagnostics-tier tool.
geometric_confidenceSummarize an already-computed state_vector into a confidence level (high/medium/low) with a recommendation. Post-hoc digest - use analyze_anomaly or check_drift for fresh analysis of raw data.
check_realizationStructural realization analysis of a payload against the Blueprint's reference configuration (requires a 'realization' block; otherwise status=skipped). Diagnostics-tier tool; prefer validate or analyze_anomaly for standard checks.
check_driftCheck whether recent submissions still match the established pattern for this Blueprint. Returns a stability verdict and observation count.
authorize_executionGo/no-go for a real-world action (payment, filing, API write): runs full validation, then the Blueprint's execution gate. authorized=true only on PASS; REVIEW means do not proceed automatically. Different from validate: validate asks is this data correct, authorize_execution asks should this action happen.
load_rule_packLoad a prebuilt Blueprint template (invoices, timecards, legal, POs, claims). Call without pack_id to list packs; then create_blueprint to save a customized copy.
get_execution_traceRun validation and return the per-node execution trace (node names, deterministic flags, timing) plus the verdict and determinism hash. Use validate for normal operation; this is for debugging and audit preparation.
verify_replayCheck replay commitment integrity and compare recorded execution components. Version 4 includes reference context and the final result/status. A match compares commitments; this tool does not rerun the workflow or reconstruct historical reference populations, and does not prove factual correctness.
account_statusThis account's plan, key usage, Blueprint counts, and the deployed platform build fingerprint (version, build, deployed).
delete_blueprintPermanently delete a Blueprint and revoke its API keys. Irreversible; requires confirm=true. Account-level keys are unaffected.
update_blueprintUpdate an existing Blueprint in place. Only passed fields change; pass [] to clear a list. workflow_name cannot be renamed and existing API keys keep working. Different from create_blueprint: modifies an existing Blueprint, mints no new key.
list_api_keysList this account's API keys (masked) with their Blueprint bindings.
rotate_api_keyReplace an API key with a fresh one. The old key stops working immediately; the new key inherits its bindings.
delete_api_keyPermanently delete one of the caller's API keys. DESTRUCTIVE — agents using the deleted key will receive auth errors immediately. The Blueprint a key was tied to (if any) is NOT affected; only the credential is revoked. To delete a Blueprint and all its keys, use delete_blueprint. The target key can be specified two ways: - As the full key string (gai_...). - As a key_id (SHA-256 hash from list_api_keys).