The best static intelligence for agents*.
Quick start · Why SolidKG · Benchmarks · Capabilities
Coding agents are good at reasoning once they have the right context. Finding that requires fine-tuned automated assistance.
Tested error-free on 112 of the largest and most impactful OSS codebases, SolidKG indexes a project once, keeps it in sync, and gives agents an optimized structural view.
- Follow real flows. Trace callers, routes, callbacks, UI render chains, framework conventions, and cross-language bridges.
- Return evidence, not summaries. Results include verbatim source, graph relationships, provenance, and explicit uncertainty when static analysis cannot prove a path.
- Understand change risk. Query callers, callees, impact radius, affected tests, snapshots, package boundaries, and layer coupling.
- Stay local. SQLite, tree-sitter, and optional SCIP run on your machine. No source uploads, embeddings, vector database, API keys, or telemetry are required.
Published packages bundle the Node.js 24 runtime SolidKG executes on, so you do not need Node.js 24 installed separately.
Agent-led setup: For a first installation, give
docs/setup.mdto your coding agent. It guides the agent through read-only preflight, approved configuration changes, project indexing, and a realsolidkg_exploreacceptance check.
npm install -g solidkg
solidkg install
solidkg init -i .
solidkg status .Standalone GitHub Release bundles also include the runtime and require no separately installed Node.js, npm, or build tools.
Then ask your coding agent a normal question:
How does authentication work in this project?
The default MCP profile exposes one answer-producing tool, solidkg_explore, so the agent can send the question it already has and receive ranked source plus structural evidence in one response.
Prefer the CLI?
solidkg context "how does authentication work?" --path .
solidkg callers UserService --path .
solidkg impact UserService --path .
solidkg affected src/auth.ts --path .SolidKG stores deterministic facts derived from source, not LLM-generated descriptions:
source files
↓ tree-sitter extraction
symbols + calls + imports + references
↓ framework and dynamic-dispatch resolution
routes + callbacks + render chains + semantic hyperedges
↓ optional SCIP linking
compiler-grade definitions + references + implementations
↓ query surfaces
MCP + CLI + TypeScript API
Today that covers 43 language/source modes and 25 framework/runtime resolvers:
| Ecosystem | Language/source modes | Framework/runtime resolvers |
|---|---|---|
| Web and JavaScript | TypeScript, JavaScript, TSX, JSX, Svelte, Vue, Astro, Liquid | Hono, Express/Koa-style routing, NestJS, React/React Router, Svelte/SvelteKit, Vue/Nuxt |
| Python and templates | Python, Jinja | Django/DRF, Flask/Flask-RESTful, FastAPI |
| JVM | Java, Kotlin, Scala, Groovy | Spring, Play Framework |
| Native and Apple | C, C++, Objective-C, Swift, SystemVerilog | SwiftUI, UIKit, Vapor, Swift–Objective-C bridge |
| .NET | C#, VB.NET, Razor | ASP.NET Core |
| PHP and Ruby | PHP, Ruby, Twig | Laravel, Drupal, Rails |
| Go and Rust | Go, Rust | Go routers (Gin, chi, gorilla/mux), Rust web (Axum, actix-web, Rocket) |
| Mobile bridges | Dart, ArkTS | React Native (legacy/TurboModules), Expo Modules, React Native Fabric/Codegen |
| Scripting and scientific | Lua, Luau, Julia, Fortran, R, Pascal | — |
| Data, config, and shell | SQL, HCL, Nix, SCL, Solidity, Bash, PowerShell, YAML | dbt |
See the exact language and framework coverage and the tested codebases.
This July 2026 detailed report describes a maintainer-run local, post-indexing OpenCode diagnostic: ten architecture-explanation tasks, seven arms, four repeats per launchable repository-arm cell, and 276 completed sessions. It is a bounded diagnostic, not a product ranking or independently reproducible benchmark.
The strongest matched SolidKG core versus no-graph baseline result was lower agent interaction count:
| Paired median result | SolidKG core versus baseline |
|---|---|
| Mechanism F1 difference | 0 |
| Tool calls | 8 fewer, 70.0% lower |
| Tokens | +0.2% |
| Session time | +9.6% |
Answer quality was ceiling-limited, with arm mean F1 values from 97.1% to 98.1%. Token and session-time bootstrap intervals crossed zero, so this run does not show a reliable reduction in either. The tool-call raw sign-flip p-value was 0.00195, but its Holm-adjusted p-value was 0.140625. Treat the reduction as descriptive, not multiplicity-adjusted significant.
Tool calls count agent actions across each packaged tool surface, including graph calls and built-in reads and searches. SolidKG's default one-tool MCP profile contributes to this interaction-economy result. It is not independent proof of superior retrieval quality, less computation, lower cost, or higher productivity. Aggregate mean token use was lower largely because of VS Code, not a general token result.
CodeGraph completed 36 of 40 cells because four .NET Runtime cells were unavailable. Its aggregate therefore covers a different task mixture and is not a matched comparison. Direct MCP latency had only four process-level observations on one host, making tail percentiles unsuitable for headline inference. Indexing and preparation time were excluded, so this is not install-or-index-to-answer timing.
Two blinded, same-family AI reviewers disagreed in at least one scored field on 269 of 276 answers, and separate AI adjudication resolved those cases. Reported F1 differences should not imply human-validated precision. Raw trajectories, exact source snapshots, resolved configurations, scoring inputs, and the analysis package are not published in this repository, so the diagnostic cannot be independently audited or reproduced from the repository alone.
One architecture question per famous public repository may be easy or familiar. The study does not test patches, builds, regressions, security, maintenance, or productivity. The defensible conclusion is comparable ceiling-limited answer quality and approximately 70% fewer agent tool calls, but no reliable token or end-to-end time reduction. Larger reproducible coding benchmarks with executable outcomes are required.
SolidKG goes beyond a flat symbol index:
- Semantic hyperedges preserve multi-symbol relationships such as route bindings, event channels, UI render chains, dynamic dispatch, native bridges, and type contracts.
- Framework resolvers connect routes, handlers, composition, and native bridges that syntax alone misses.
- Optional SCIP adds compiler-grade navigation without replacing the always-available AST graph.
- Source-body retrieval improves natural-language discovery while exact symbol, path, trace, semantic, and SCIP evidence stays higher priority.
- Architecture and snapshots make package coupling and graph changes explicit instead of leaving them as review intuition.
Retrieval stays deterministic and transparent:
- Persisted local source-body/code-token chunks can seed or boost a result and appear as the
source_body_chunkranking reason. Exact symbol, path, trace, semantic, and SCIP evidence still outranks body-only matches. - External embeddings, vector databases, rerankers, and hybrid retrieval providers are not enabled by default; the optional hybrid retrieval seams remain disabled.
- Query exclusions are retrieval constraints, not search terms. A
missing_query_conceptlabel means the returned source did not cover every explicit concept, so the answer is intentionally partial.
import SolidKG from "solidkg";
const graph = await SolidKG.init("/path/to/project", { index: true });
const matches = graph.searchNodes("UserService");
const match = matches[0];
if (!match) throw new Error("UserService not found");
const impact = graph.getImpactRadius(match.node.id);
const overview = graph.getOverviewText({ maxDepth: 2 });
console.log(matches.length, impact.nodes.size, overview);
graph.close();solidkg overview . --max-depth 2
solidkg snapshot create before-refactor --path .
solidkg architecture analyze --path .
solidkg scip generate --path .Set SOLIDKG_MCP_TOOLS=all on the MCP server to expose the advanced trace, precise-reference, snapshot, architecture, and status tools. Run solidkg <command> --help for the complete CLI.
The public source archive includes the maintained SCIP integrations and pinned native grammar manifest needed to reproduce supported behavior.
Source checkout and development require Node.js 24 and the pnpm version pinned in package.json (10.34.5).
git clone https://github.com/Valid-Systems/SolidKG.git
cd SolidKG
corepack enable pnpm
pnpm install --frozen-lockfile
pnpm run verify:source:fullverify:source:full builds the project, validates required assets and vendored sources, indexes a real fixture, exercises the stdio MCP server and solidkg_explore, runs the public test suite, and checks release-package contents.
SolidKG is open source under the MIT License.
Third-party notices are in docs/THIRD_PARTY_LICENSES.md. SolidKG builds on ideas explored by Codegraph, Sverklo, Tilth, and Cartog.
* “The best static intelligence for agents” is based on SolidKG's analyses of non-commercial alternatives; commercial offerings were excluded.