About SiftrCode
We build context optimization technology between software repositories and autonomous AI coding agents.
The Evolution from Token Reduction to Outcome-Awareness
SiftrCode began with a clear observation: coding agents consume far more source code than they need to solve tasks. Ingesting tens of thousands of lines of raw function bodies leads to context saturation, response delays, and attention dilution.
However, experience with production agents proved an equally critical truth: pure token reduction is not enough. If a context optimizer strips implementation logic from the file the agent actually needs to edit, the agent fails.
SiftrCode is the context optimizer for coding agents: give your coding agent focused repository context instead of dumping the whole codebase into its prompt. Through AST discovery, deterministic ranking, variable resolution, and token budget knapsacking, SiftrCode ensures causal edit targets are delivered at high fidelity while supporting modules degrade safely to exact AST interface contracts.
Outcome-Aware Variable Resolution
Distinguishes causal edit targets from background dependencies. Edit files receive full body preservation, supporting files receive clean AST interface skeletons, and distractors are pruned completely.
100% Local-First & Zero Egress
Your intellectual property never leaves your machine. Multi-channel recall, AST parsing, and budget solving execute entirely on local CPU with 0 bytes of network transmission.
Sub-100ms 6-Stage Optimization
Engineered with a deterministic knapsack solver and submodular bundle composer, condensing 100,000 lines of code into high-precision context in under 100 milliseconds.
Open Source & Permissive MIT
Freely licensed under the permissive MIT License for individual engineers worldwide, backed by transparent public audits, explainable ranking, and zero telemetry.
Closed-Loop Outcome Verification
Bridges pre-flight context budgeting with post-execution test passes, regressions, and actual provider token tracking (siftr_outcome) to verify task completion objectively.
Air-Gapped Learning Plane
Multi-dimensional TrainingEvidenceRecord lineage captures feedback without raw source code leakage (ContextPlanMetadataRecord) in local SQLite.
Empirical SiftrBench Rigor
Audited 25-task real-world pilot study (500 bounded evaluations across Express, FastAPI, and SiftrCode) validates semantic LLM signals with genuine continuous probability distributions (relevance 0.3445, edit target 0.2159), sub-10ms latency (P50 4ms, P95 6ms), and 100% normalized decision plan invariance before production influence.
Structured Egress Gateway
Automated regex sanitization scrubs API tokens (AWS, Stripe, Bearer credentials) before external provider dispatch, enforcing local-first trust domain boundaries.
Canonical AgentEnvironment Lineage
Cryptographically binds tasks, sessions, workspace snapshots, plans, decision observations, JEV signals, and SQLite durability to systemConfigurationHash, ensuring complete lineage integrity.
Sanctioned Training Exporter Boundary
Raw observations cannot bypass the exporter boundary. Persistent training rows strictly require sanctioned texport_ export IDs with verified multi-dataclass data rights.
Open Source + Team Infrastructure
We believe standard developer tooling must remain open, auditable, and freely accessible. The SiftrCode CLI (npx siftrcode) is free forever for developers working on open source or individual commercial projects.
For fast-scaling engineering teams and enterprises, we provide SiftrCode Pro: a synchronized context cache that deduplicates codebase context across engineering seats, enforces architectural boundaries, and runs pull-request context audits in CI/CD.
Get in Touch
We love pair-programming with engineering teams who are pushing coding agents to their limits. Whether you're benchmarking a 2M-token monorepo or building custom MCP agent loops, we're here to help:
Ready to test it on your codebase?
Run a 5-second audit or register SiftrCode directly with your agent.
npx siftrcode audit