AI工具输出的可靠性网关：schema-stable、secret-safe、分页完整JSON。

Sift 是 MCP 和 CLI 工具输出的可直接嵌入可靠性层。它以工件形式保留完整有效载荷，返回内联有效载荷（full）或紧凑引用（schema_ref），并允许代理用 Python 代码对存储数据进行查询所需内容。

基准总结：在12个真实JSON数据集中，103个事实性问题中，Sift将准确率从33.0%提升到99.0%，同时剔除输入令牌减少了95.4%（10,757,230 - > 489,655）。完整详情请见benchmarks/README.md。

它是如何运作的

流程：

1. 执行上游工具/命令并捕获 JSON。
2. 在 SQLite 中作为工件持久化完整输出，并确定性地映射模式/根提示。
3. 返回full（小）或schema_ref（大/分码）。
4. 明确继续页面直到pagination.retrieval_status == COMPLETE。
5. 对单个工件或完整的分页链运行聚焦的Python查询。

主要MCP痛点

这些内容在MCP客户端问题跟踪器和协议生产环境中的使用中反复出现：

• 大型工具定义和大型工具结果会迅速消耗上下文。

Sift

Reliability gateway for AI tool output: schema-stable, secret-safe, pagination-complete JSON.

Sift is a drop-in reliability layer for MCP and CLI tool output. It persists full payloads as artifacts, returns either inline payload (full) or compact references (schema_ref), and lets agents query what they need with Python code over stored data.

Benchmark summary: on 103 factual questions across 12 real JSON datasets, Sift improved accuracy from 33.0% to 99.0% while cutting input tokens by 95.4% (10,757,230 -> 489,655). Full details: benchmarks/README.md.

How it works

                           ┌─────────────────────┐
  MCP tool call ──────────▶│                     │──────────▶ Upstream MCP server
  CLI command   ──────────▶│        Sift         │──────────▶ Shell/API command
                           │                     │
                           │   ┌─────────────┐   │
                           │   │  Artifacts  │   │
                           │   │  (SQLite)   │   │
                           │   └─────────────┘   │
                           └─────────────────────┘
                                     │
                                     ▼
                         Small output -> `full` inline
                         Large output -> `schema_ref`
                         Agent queries artifacts with code

Flow:

1. Execute upstream tool/command and capture JSON.
2. Persist full output as an artifact in SQLite and deterministically map schema/root hints.
3. Return full (small) or schema_ref (large/paginated).
4. Continue pages explicitly until pagination.retrieval_status == COMPLETE.
5. Run focused Python queries on one artifact or the full pagination chain.

Main MCP pain points

These are recurring across MCP client issue trackers and protocol usage in production:

• Large tool definitions and large tool results consume context quickly.
• Upstream API pagination often sits outside MCP list-cursor flows, so agents can stop early and answer on partial data.
• Tool output shape differs across servers, which makes follow-up parsing brittle.
• Tool output is untrusted input and can contain sensitive values that should not re-enter model context.
• Raw outputs scroll away in chat history, so provenance and reproducibility degrade across multi-step runs.

Background and references: docs/why.md.

What Sift adds (without changing upstream servers)