Перейти к содержанию

Multi-surface AI agent toolkit with credentialed gateway

Дата: 2026-09-11 Репозиторий: https://github.com/sageox/agent-toolkit Уверенность: в карантине Модель: claude-haiku-4-5-20251001 Промпт версия: v2.0 Источник: файл: README.md локация: не указана цитата: "Hosted agents, on your own infra. One agent, one memory, across Buzz and Slack. Your team's context, inherited and live, when SageOx is connected."

What Changes in the Ecosystem

The ecosystem gains a reference architecture for multi-surface agents that decouples the decision-making model from credential management and tool execution, shifting from "model holds keys" to "gateway holds keys, brain only asks." Teams deploying this project can achieve cross-channel agent presence (Slack + Nostr/Buzz) with unified memory while maintaining fine-grained tool access control and audit trails. The pattern enables smaller organizations and self-hosted deployments to run sophisticated agents without building their own gateway and credential infrastructure.

Reasoning

The project implements a working multi-surface agent architecture with architectural innovation (gateway-model separation, credential isolation, MCP-in-gateway pattern) that goes beyond conventional agent frameworks. Manifest shows real infrastructure (Helm, Terraform, audit logging) and test coverage. However, adoption signals remain limited to early adopters; the 3 maturity reflects functional code with some production deployment capability but not yet field-hardened at scale.

Maturity x Novelty

Maturity: 3/5 Novelty: 4/5

Self-Check (CoVe)

Cross-validation (README vs manifest/files): Yes. The README claims: (1) "Hosted agents, on your own infra" with gateway owning credentials and connections—manifest shows deploy/ with Helm and Terraform, deployment-contract.md referenced. (2) "One memory, one persona, one profile across surfaces"—file tree shows .sageox/ and docs/guide/memory-and-tools.md structure. (3) MCP servers run "inside the gateway, not beside the brain"—docs/guide references MCP credential handling. (4) Team memory via SageOx integration—.sageox/ directory and references to ox CLI. (5) Audit log with tool_call tracking—docs/guide/reference.md referenced. Package.json shows TypeScript tooling and vitest test coverage. The manifest structurally supports the README claims without contradictions. Подтверждено: Да

Novelty checklist: New protocol? Yes—the gateway-model credential separation and MCP-in-gateway pattern represent a new architectural protocol for safe agent tool access that isolates the language model from credential handling. New standard? Partially—team memory integration with SageOx is a procedural standard for team knowledge inheritance but not a widely adopted formal standard yet. New architectural layer? Yes—the gateway layer as a credentialed multiplex with per-surface egress routing and refuse-able refusals is a new architectural abstraction not found in earlier agent toolkits. New market interaction model? Yes—the model of self-hosted sovereignty over agent infrastructure with team-scoped memory over SageOx represents a new way teams can deploy and run agents without vendor lock-in while gaining organizational context. Проходит: Да

Falsifiable Hypothesis

Если права: Within 12 months, at least 3 independent organizations (verified by public GitHub discussions, case studies, or conference talks) deploy agent-toolkit in production across Slack and Buzz concurrently, confirming the multi-surface architecture is viable and adoption is growing beyond proof-of-concept. Если ошиблась: Within 12 months, the project receives no new external contributors, no documented production deployments emerge, and issue velocity drops below 2 issues/month, indicating the architecture remains a reference design without meaningful ecosystem adoption.

Оценка Claude

Правка человека

Мнение Ольги

История оценок

  • 2026-09-11 - CANDIDATE: первая оценка

Связи

  • [[MCP как слой интеграции сервисов 2026-06-14]]
  • [[Self-hosted суверенитет над агентной инфраструктурой 2026-06-14]]
  • [[Верификация и доверие к действиям агентов 2026-06-14]]
  • [[Локализация памяти и идентичности агента 2026-07-23]]
  • [[Collaborative_AI_агенты_в_shared_workspaces 2026-06-24]]