LeanCTX is positioned as a Context Gateway for AI systems, with its core function being to control what AI can see. It sits between AI and the systems it reads from, responsible for selecting the relevant context needed for a task, applying supported control rules, and recording what the integration can observe. The official emphasis is that "source checking and model delivery are separate": a permitted file can be read locally, and then its returned content is filtered before delivery.
Main capabilities
- Context selection: selects the permitted information per task, prepares the needed details, and retains source references. Examples in the materials include selecting the relevant implementation from the billing code total.ts, providing filtered diagnostic context from logs, rejecting customer export files, and redacting email fields and customer_id.
- Source and version control: uses sources authorized by the user and task; only content explicitly marked as the current version in the source metadata is selected, while superseded older versions are retained only as references; sources beyond the user's permission scope are denied access.
- Pre-delivery rules: uses source permissions, configured content checks, and budget constraints to support rules such as allowing/denying sources and masking fields in supported read results.
- Read-time checks: enables checks for known injection patterns and qualification-style checks on supported text results; secret redaction and custom patterns can be added to the rules. Filtering metadata reports classifications and counts, without including sensitive matched values.
- Filtering rules and auditing: for detected content, you can choose to warn, redact, or block; audit details depend on the specific call, and the integration can observe source references, policy outcomes, and recorded usage.
- PathJail: restricts reads to allowed directories and checks resolved paths, including symbolic links and path traversal attempts. The official documentation states that this is an application-layer check, not an OS-level sandbox, and cannot eliminate OS or race condition limitations.
Usage
- Context service mode: lets LeanCTX retrieve, shape, and filter context; the application receives the prepared results and decides for itself how to use them, with model execution still under user control. No managed service is required when using only the context functionality.
- Model request routing: qualifying requests can be routed through the configured gateway; the enterprise edition adds organizational identity, policy, and reporting on top of this. Coverage depends on the chosen edition and integration.
- Retain your own tools and repositories: you can use the open-source Engine locally, or route supported requests through configured providers and proxy handling. Controls such as routing and proxy budgets depend on the configured integration; local processing does not equal local inference.
- For AI product builders: Supported Engine capabilities can be embedded through five stable SDK primitives and six Agent Tools language contracts. Preview operations are located in the Python namespace leanctx_sdk.preview. Production and OEM rights require a separately signed written agreement.
- For enterprise buyers: The Enterprise Edition is a commercial offering that provides organizational identity, policy, and reporting in supported AI workflows. Deployment, retention, and support depend on the licensed edition and agreement.
Applicable scenarios The materials give application directions including: embedding context control into your own application, copilot, or agent; connecting to a compatible LeanCTX Engine within a product through the SDK, running locally or using it with a compatible remote deployment. The materials also mention that Wavect uses LeanCTX in its development workflow, reducing the context delivered to its agents (this snapshot is a reporting observation as of July 26, 2026, and is not an audited vendor bill or a promised saving).
Limitations and notes The materials do not provide specific pricing, charging methods, a list of supported models, deployment environment details, performance data, or a complete integration list. PathJail is explicitly an application-layer check rather than an OS sandbox. Enterprise Edition features, deployment, retention, and support all depend on the licensed edition and agreement. The specific contents of the five SDK primitives and six Agent Tools language contracts are not listed one by one in the materials. Production and OEM rights require a separately signed written agreement.
