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)

• Artifact-backed outputs: keep full data out of prompt context while preserving it losslessly.
• Tool inspection helper: keep mirrored tools/list descriptions compact and pull full docs with gateway.inspect_tool.
• Schema-aware references: schema_ref returns query guidance for stable follow-up analysis.
• Exact structured retrieval: run Python against stored artifacts instead of relying on prompt-sized payloads.
• Exact structured retrieval via artifact(action="query", query_kind="code", ...) (MCP) or sift-gateway code (CLI).
• Explicit pagination contract: continue with artifact(action="next_page") or run --continue-from.
• Completion signaling: do not stop until pagination.retrieval_status == COMPLETE.
• Pagination-chain analysis: query one artifact or all related pages (scope="all_related"; CLI default).
• Outbound secret redaction enabled by default before output returns to the model.

MCP vs CLI positioning

• MCP: Sift is a reliability gateway for mirrored tool calls and artifact-based follow-up queries.
• CLI/OpenClaw: same artifact contract for command output (sift-gateway run + sift-gateway code).
• CLI pitfall: ad-hoc extraction can silently scope analysis to partial data (for example, inspecting only one row).
• CLI note: for one-off local extraction, plain jq can be enough. Sift is for repeatable, pagination-complete, policy-controlled workflows.

60-second quickstart

MCP clients

pipx install sift-gateway
sift-gateway init --from claude

Restart your MCP client, then use mirrored tools normally.

Supported --from shortcuts: claude, claude-code, cursor, vscode, windsurf, zed, auto, or an explicit config path.

CLI flow

# 1) Capture JSON output as an artifact
sift-gateway run --json -- kubectl get pods -A -o json

# 2) Query artifact data with Python
sift-gateway code --json <artifact_id> '$' --code "def run(data, schema, params): return {'rows': len(data)}"

Use $ when rows are at root. If nested, use metadata.usage.root_path from run --json (or metadata.queryable_roots in MCP schema_ref).

Pagination continuation

sift-gateway run --json --continue-from <artifact_id> -- <next-command-with-next-params-applied>

Do not claim completion until pagination.retrieval_status == COMPLETE.

Python codegen over all pages

For complex questions, generate Python once and run it over the entire pagination chain:

sift-gateway code --json --scope all_related <artifact_id> '$' --file ./analysis.py

CLI default is --scope all_related. Use --scope single for anchor-only analysis.

Benchmarks

Tier 1 result (claude-sonnet-4-6):

That is +66.0 points accuracy with 95.4% fewer input tokens on the same question set.

Methodology, scripts, and Tier 2 autonomous-agent results: benchmarks/README.md.

Documentation library

Start here: docs/README.md

Getting started

• Quick Start
• Installation
• Your first artifact (CLI)
• Your first artifact (MCP)
• Adding MCP servers after initial setup
• Troubleshooting

Core contracts

• API Contracts
• Mirrored Response Contract (full vs schema_ref)
• Response Mode Selection
• Pagination Metadata
• Code Query Contract
• CLI output contract
• CLI default scope (all_related)

Operations and security

• Deployment Guide
• Authentication tokens
• Outbound secret redaction
• Configuration Reference
• Code query runtime
• Error Contract
• Security policy

Patterns and deep dives

• Recipes
• Pagination chain (CLI)
• Pagination chain (MCP)
• Architecture
• Pagination model
• Observability
• Why Sift exists
• OpenClaw integration pack
• Upstream registration design

Security

See SECURITY.md for threat model and hardening guidance.

Development

git clone https://github.com/lourencomaciel/sift-gateway.git
cd sift-gateway
uv sync --extra dev
uv run python -m pytest tests/unit/ -q

Full contributor workflow: CONTRIBUTING.md

License

MIT - see LICENSE.