SiftrCode V2 • TypeSafe SystemOne Live Pilot Certified →

Context optimization, outcome-aware

SiftrCode discovers repository context, protects edit targets, and synthesizes interface skeletons — eliminating 90% of token bloat while boosting coding agent accuracy.

$ npx siftrcode context "task prompt"

See Interactive Comparison ↓

−90%

Token Reduction

AST skeletons eliminate internal function bodies

100%

Plan Invariance

Bit-for-bit identical baseline vs shadow decision plans

175ms

Median Latency

Verified on 25 real tasks via TypeSafe SystemOne

0 Egress

Local & Secure

Zero raw source or secrets leave your machine

See the Difference: Raw Dump vs SiftrCode

Select a task to compare how an AI agent experiences a raw repository dump versus a SiftrCode outcome-aware bundle.

Without SiftrCode

94,200 Tokens

Context Bloat & Agent Confusion

Agent is handed entire repositories or full uncompressed files. Critical edit target is buried inside 18 files of irrelevant boilerplate.

Injected Files:

Cost per turn: $0.283 100% Uncompressed

OPTIMIZED

SiftrCode V2 Bundle

3,030 Tokens −96.8%

★ EDIT TARGET (FULL BODY PROTECTED) 100% Fidelity

src/webhooks/signature_verifier.ts

Preserved with complete implementation so agent edits succeed without hallucinations.

DEPENDENCIES (AST SKELETON SYNTHESIS) Excised Bodies

14 Unrelated Files Suppressed Zero context pollution

Cost per turn: $0.009 Save $0.274 / turn

AST Interface Skeletonizer Demo

See how function bodies are excised across 4 languages while preserving type interfaces.

RAW IMPLEMENTATION 240 tokens

...

SIFTRCODE AST SKELETON 38 tokens (−84%)

...

How SiftrCode Optimizes Agent Context

From full repository scan to compressed context plan in under 100 milliseconds.

Multi-Channel Recall

Extracts AST symbols, traverses the repository call graph, and analyzes git co-change history to find candidate files related to the task prompt or stack trace.

Exact + Lexical + Graph + Git

Outcome-Aware Scoring

Evaluates candidates against point-in-time features to distinguish exact edit targets from supporting dependencies. Verified against production TypeSafe SystemOne JEV signals.

Edit Target vs Dependency

Submodular Budgeting

Linear programming solver allocates each file into its optimal resolution (FULL, BODY, SKELETON, SIGNATURE, OMIT) under strict token caps with zero hallucination.

Linear Optimization Solver

Empirical Validation

Live TypeSafe Pilot Results

Target Repositories: Express • FastAPI • SiftrCode

Tasks Evaluated

25 / 25

100% Success Rate (0 Failures)

Decision Plan Invariance

100.0%

Normalized cryptographic equality

Remote Latency (P50)

175ms

P95: 294ms • 0 retries required

Ground-Truth Correlation

r = 0.513

Strong root cause identification

Lineage Coverage: 9,018 / 9,018 decision records referentially joined with 0 orphans.

Read Full Architecture →

Get Started in 30 Seconds

Works seamlessly with Claude Code CLI, Cursor, Windsurf, or as a standalone CLI tool.

claude mcp add siftrcode -- npx siftrcode mcp

Claude Code will automatically invoke siftr_context before executing edits.

Add to your IDE's MCP configuration (~/.cursor/mcp.json):

{
  "mcpServers": {
    "siftrcode": {
      "command": "npx",
      "args": ["siftrcode", "mcp"]
    }
  }
}