One Engine. From source to outcome.

Follow the request

LeanCTX is a Context Intelligence Engine. An AI tool, agent or application supplies a task and its identity. The Engine resolves permitted sources, compiles the relevant evidence and applies the controls that govern the next operation.

Task + identity + limits
          ↓
Permitted sources → retrieval → context plan
                                    ↓
                          input checks + budget
                                    ↓
              prepared context OR governed execution
                                    ↓
                     receipt + usage + outcome
                                    ↓
                       memory + checkpoint

The same flow applies to code, service records, documents and structured data. A source remains evidence: text retrieved from a file, ticket or website cannot grant itself a new permission.

Choose who makes the model call

Mode	LeanCTX	Your application
Context preparation	Returns selected, checked context with source references	Chooses the model and sends the request
Governed execution	Enforces the permitted route, provider, budget, retry and fallback policy; records the result	Supplies the goal, authorization and acceptance criteria

Controls cover operations routed through the Engine. Direct calls made outside that path remain the host’s responsibility. A context tool result alone does not prove that an unrelated provider request was filtered.

The host remains the authorization authority. The Engine enforces the granted limits; an optimizer or worker cannot enlarge them by proposing another action.

Keep four boundaries explicit

Boundary	What to establish
Information	Which sources, records, fields and representations the request may use
Execution	Which models, tools and actions are permitted, at what cost and with which approval
Evidence	What was actually delivered or run, what usage was observed, and what remains estimated or unknown
Continuity	Which findings and artifacts should survive, for how long, and with which provenance

A receipt records an operation. A successful task requires a separate acceptance check. Reduced input tokens, a completed call and a useful outcome are different measurements.

Use one core through different interfaces

CLI, MCP, proxy and SDK are entry points into Engine capabilities. Native SDK packages expose the versioned contract; they do not maintain six competing implementations of context policy.

The standalone Enterprise Engine adds organizational identity, governed sources and providers, budgets, audit controls and managed operation. It can serve external applications without requiring a particular application above it. See organizational deployment.

Retain selected state

Caches, source-linked knowledge, session progress and portable checkpoints serve different purposes. A checkpoint transfers supported state; it does not copy an entire running process or authorize a new recipient.

Verify source freshness, destination identity, policy and format compatibility when recovering state. Cancellation and interrupted work remain explicit outcomes rather than being recorded as completed work.

See memory and knowledge, context packages and multi-agent coordination.

Engine contract & references2 references · Reviewed 19 Sept 2026

• registry.rsrust/src/core/providers/registry.rsGitHub
• state.rsrust/src/core/session/state.rsGitHub

Engine concepts describe responsibilities and behavior. Exact commands, package versions and platform contracts belong to their versioned references.

Versions & compatibility