sift-gateway 0.2.2

Keep big MCP responses out of your context window. Query them.

Sift

Sift is a local, single-tenant gateway for AI agent tool work. It fixes context bloat by storing large tool outputs as artifacts and returning only the slices an agent needs.

Use it with agents that talk through MCP, or run it directly in CLI workflows.

One Command, Two Modes

Sift uses one command handle: sift-gateway.

Mode	How you run it	What it does
MCP gateway mode	sift-gateway (default server behavior)	Proxies MCP traffic and persists tool outputs as artifacts
CLI mode	sift-gateway <artifact-command>	Captures, queries, and computes over artifacts directly

These modes are independent. You can run only MCP mode, only CLI mode, or both.

MCP Gateway Mode

Use this when an agent/client talks to tools over MCP.

Quick start

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

Then restart your MCP client.

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

Common gateway commands

sift-gateway --help
sift-gateway upstream add '<json>' --from claude
sift-gateway install pandas
sift-gateway uninstall pandas
sift-gateway --transport sse --host 127.0.0.1 --port 8080

Runtime behavior

• Mirrors upstream MCP tools with original schemas.
• Persists mirrored outputs as artifacts.
• Returns small responses inline and larger responses as handles.

CLI Mode

Use this when you want artifact workflows directly in terminal automation.

Quick start

pipx install sift-gateway
sift-gateway run -- echo '{"items":[{"id":1,"name":"a"}]}'
sift-gateway list --limit 10

Capture sources

# Capture command output
sift-gateway run -- git status --porcelain

# Capture stdin
cat payload.json | sift-gateway run --stdin

Query and inspect artifacts

sift-gateway schema art_123
sift-gateway get art_123
sift-gateway query art_123 '$.items' --select id,name --limit 50
sift-gateway code art_123 '$.items' --expr 'len(df)'
sift-gateway diff art_left art_right

CLI mode uses local state in .sift-gateway by default. Use --data-dir to target a different instance.

MCP Artifact Query Model

In MCP workflows, retrieval happens through the artifact tool with action="query".

Supported query_kind values:

query_kind	Purpose
describe	Schema and metadata
get	Full payload retrieval
select	Projection/filtering from a root path
search	Search and list session artifacts in the current workspace
code	Execute constrained Python over artifact data

Example query_kind="search":

artifact(
    action="query",
    query_kind="search",
    query="github issues",
    limit=25,
)

Example query_kind="select":

artifact(
    action="query",
    query_kind="select",
    artifact_id="art_123",
    root_path="$.items",
    select_paths=["id", "name", "status"],
    limit=50,
)

Context-Bloat Controls

• SIFT_GATEWAY_PASSTHROUGH_MAX_BYTES controls inline-vs-handle threshold.
• Default: 8192 bytes.
• Set to 0 to force handle-first behavior.

Configuration Highlights

Env var	Default	Description
SIFT_GATEWAY_DATA_DIR	.sift-gateway	Instance root directory
SIFT_GATEWAY_PASSTHROUGH_MAX_BYTES	8192	Inline response threshold
SIFT_GATEWAY_CODE_QUERY_ENABLED	true	Enable code queries
SIFT_GATEWAY_SECRET_REDACTION_ENABLED	true	Redact likely outbound secrets
SIFT_GATEWAY_AUTH_TOKEN	unset	Required for non-local HTTP binds

Full reference: docs/config.md

Security Notes

• Code queries use AST/import/time/memory guardrails, not full OS sandboxing.
• Outbound secret redaction is enabled by default.

Disable code queries if needed:

export SIFT_GATEWAY_CODE_QUERY_ENABLED=false

More: SECURITY.md

Documentation

• docs/quickstart.md
• docs/config.md
• docs/api_contracts.md
• docs/recipes.md
• docs/deployment.md
• docs/errors.md
• docs/observability.md
• docs/architecture.md
• docs/openclaw/README.md

Development

git clone https://github.com/lourencomaciel/sift-gateway.git
cd sift-gateway
uv sync --extra dev

UV_CACHE_DIR=/tmp/uv-cache uv run python -m ruff check src tests
UV_CACHE_DIR=/tmp/uv-cache uv run python -m mypy src
UV_CACHE_DIR=/tmp/uv-cache uv run python -m pytest tests/unit/ -q

License

MIT - see LICENSE.