Open-source evidence infrastructure for AI-native software.
Raintree Technology builds open-source tools for testing and reviewing AI systems. Our projects make three commonly hidden layers inspectable: the context agents consume, the policies governing their actions, and the interfaces they help generate. Each tool produces durable local evidence and can enforce release criteria in CI.
Built for engineers working on agents, RAG pipelines, governed data products, and AI-assisted application development.
Explore our repositories · Visit raintree.technology · Start a technical conversation · Follow @raintree_tech
| Project | Problem | What it does | Best for |
|---|---|---|---|
| DocPull | External context changes without a reviewable dependency trail | Locks sources, builds cited context packs, detects changes, and gates context quality | Agents, RAG, evals, and context engineering |
| PolicyStrata | Policy drifts between agent intent, application logic, SQL, databases, and output | Finds cross-layer disagreement and produces reproducible witnesses and release decisions | Text-to-SQL, BI copilots, analytics agents, and governed AI |
| HIG Doctor | Coding agents and teams lack consistent, automatable interface guidance | Combines structured Apple HIG knowledge with deterministic source-level UI audits | Apple platforms and cross-framework interface review |
- DocPull uses dependency manifests, content hashes, lockfiles, cited pack contracts, and Context CI to make agent context reviewable.
- PolicyStrata ships a reproducible paper artifact reporting 1,720/1,720 killed non-clean injected cases and zero false positives on 80 clean controls.
- HIG Doctor evaluates 348 deterministic rules across 12 native and web framework families, backed by 14 agent skills and 156 reference topics.
- All three prioritize deterministic, local-first workflows before optional hosted or model-dependent execution.
Lock, inspect, and test the external context your agents depend on.
DocPull treats context like a software dependency. A project declares public web, package, repository, standard, API, dataset, or local-file sources; DocPull syncs them into versioned packs with stable citations, provenance, coverage, rights metadata, and content hashes. Context CI checks whether those artifacts are fresh, complete, attributable, and ready for an agent or evaluation loop.
Use DocPull when a model must answer from specific, reviewable sources rather than an opaque scrape or an unversioned collection of documents. Its base workflow is local and browser-free; rendering and paid-capable routes require explicit opt-in.
Repository · PyPI · Pack contract · Context CI
Detect policy drift across agent, application, SQL, and database layers.
PolicyStrata compares a canonical policy with the behavior exposed by manifests, semantic plans, validators, compilers, database controls, and release filters. Its scanner imports representative traces, exercises deterministic mutations, localizes the first disagreement, and reduces failures into compact witnesses that can be reviewed or enforced in CI.
The project also includes a TypeScript recorder and default-deny authorizer, plus a customer-hosted Agent Trust Gateway for local runtime evaluation and sanitized decision envelopes. These controls add evidence and defense in depth; they do not replace application authorization or database policy.
Repository · Paper · Reproduction artifact · PyPI · Node runtime · Agent Trust Gateway
Give coding agents deterministic Apple HIG guidance and source-level audits.
HIG Doctor combines a progressively disclosed Human Interface Guidelines corpus, a cross-framework static auditor, and a small MCP surface for coding agents. It checks accessibility, semantic color, typography, layout, motion, focus, input, and interaction patterns, then returns file-level findings, positive signals, Markdown reports, JSON, and configurable severity gates.
Use HIG Doctor as a fast, repeatable first pass across SwiftUI, UIKit, React, Vue, Svelte, Angular, Compose, Android XML, React Native, Flutter, CSS, and HTML. It provides source-level evidence—not rendered UI testing, assistive technology testing, or a complete compliance certification.
- Deterministic before probabilistic: test observable state and explicit contracts before asking a model to judge a system.
- Evidence over dashboards: keep hashes, traces, reports, citations, and witnesses portable and reviewable.
- Explicit trust boundaries: state what each tool observes, what it can gate, and which responsibilities remain with the host system.
- Local-first execution: make the useful base workflow run without a hosted account or automatic paid route.
agent-starter generates portable skills and MCP configuration for Claude Code, Codex, and Cursor from one versioned manifest.
Issues and pull requests are welcome. Each repository documents its own build, test, and contribution workflow.
To report a vulnerability, do not open a public issue. Email security@raintree.technology with the affected repository, version or commit, reproduction steps, and potential impact.
Building an agent, governed AI product, or developer tool? Start a technical conversation →