From intent to
a working method.
SolKraft is a local skill discovery, routing, and retrieval layer for agents. It converts requested capabilities into inspectable selections and ordered stages, then serves the procedure text. The host agent supplies tools, execution authority, and verification.
1. Why a routing layer?
A library of procedures is useful only when the right procedure reaches the task. Compound requests add another problem: an agent must distinguish active tasks from context, preserve order, avoid unrelated methods, and identify missing intent. SolKraft makes those decisions inspectable rather than hiding them in a catalog search result.
2. Architecture
The bundle contains 173 skills. Discovery begins with compact metadata; full entrypoints and supporting resources are retrieved selectively. CLI, REST, and MCP interfaces expose the same catalog and routing layer. Seven read-only MCP tools cover search, routing, skill retrieval, resource retrieval, selection graph inspection, skill contracts, and the contract index.
Public routing uses hardened contract policy. Contracts describe requirements, inputs, outputs, effects, and trust. Admission checks run before full instruction loading. A route retains execution_authorized: false: selecting a procedure does not grant permission to deploy, publish, delete, or perform another effect. See the client interface documentation and contract documentation.
3. Matching the work, without extra work
The repaired router handles capability conjunctions, quoted text, exclusions, deferred work, and delivery framing. It exposes unresolved requested stages. Ambiguous wording requests clarification rather than selecting every plausible skill. Stage repair clears stale unresolved state when a valid admissible selection is recovered. Regression tests exercise these boundaries in the test suite.
4. Executed validation
The full local CI completed 22 steps, including 369 Python tests, JavaScript checks, wheel installation outside the checkout, packaged-plugin MCP integration, and semantic proof publication. The semantic run used the production router and frozen catalog for 473,000 calls across 32 slices; all 15 acceptance gates passed. The published receipt supplies the measured results and source identity.
| Population | Executed result |
|---|---|
| 111,000 eligible single-skill cases | 100% exact selections |
| 62,000 policy-boundary cases | 100% expected hardened policy decisions |
| 100,000 clarified compounds | 100% exact sets and order; zero extras; zero unresolved stages |
| 100,000 unchanged original compounds | 44.704% exact; 55,296 clarification requests; zero extras |
| 100,000 stability calls | Zero mismatches across ten interleaved rounds |
The earlier 27.697% compound exact-match result was rejected as completion evidence. Clarified requests add capability-specific wording while preserving original topics and targets; the unchanged corpus remains diagnostic. Policy decisions are distinct from eligible exact skill selections. Methodology, gates, and reproduction commands are in the semantic proof specification.
5. What this evidence establishes
The result establishes exact matching for the clarified generated acceptance corpus and repeatable routing under the tested conditions. It does not establish perfect understanding of arbitrary language, successful execution of every skill, measured token savings, superiority over another system, or physical engineering qualification. Ambiguous original requests remain a meaningful limitation, visible alongside acceptance results.
6. Using the system
Install the local Python runtime and native Codex plugin, or connect a client through MCP or REST. Search for a clear specialist method; route compound work; inspect the selection and unresolved stages; retrieve the selected instructions; then apply them using the host’s authorized tools. The local service requires no hosted account or model API key. The agent still needs its own runtime and access. See the installation guide.
Explore the features ↗ Inspect the evidence ↗
The Pages console can also compose custom routes in a dedicated browser worker. It loads pinned Pyodide and PyYAML, then executes the bundled production Python router with hardened contract policy. Startup needs the runtime CDN; subsequent requests reuse the engine in the tab. The request is processed locally unless an API is explicitly connected. Skill details foreground their specific method and contract inputs and outputs, while retaining full original instructions.