depwire/depwire

**Your AI doesn't know your architecture. Depwire does.**
What makes Depwire different
**16 programming languages, plus Angular templates. Depwire parses symbols and builds dependency relationships across the constructs it supports. It records relationships it cannot resolve instead of inventing an edge.**
Depwire builds a deterministic dependency graph from source parsing, without embeddings or similarity search. For a proposed change, it traverses the relationships present in that graph to show known dependents, import chains, and health movement. Results depend on the language constructs and files the parser can analyze.
**Not a build graph either.** Depwire tracks symbol-level relationships for What If simulation, graph-aware security scanning, and blast radius analysis.
Contents
- What makes Depwire different
- Performance evidence
- The problem
- Start here
- The infrastructure layer
- Tested on real-world projects
- What If simulation
- Security scanner
- Pre-action verification
- Structural diff between commits
- Visualization
- Temporal graph
- All commands
- MCP server — AI integration
- Cross-language edge detection
- Architecture health score
- Language support
- SDK
- Cloud dashboard
- VSCode Extension
- GitHub Action — PR Impact Analysis
- pre-commit / prek hook
- Depwire Action Token (DAT)
- Roadmap
- Security posture
- Paths and file counts
---
Depwire provides graph context to AI coding assistants. Parse a project to inspect recorded relationships, assess health, and simulate a proposed change.
⭐ If Depwire saves you from a broken build, star the repo — it helps this project grow.
---
Performance evidence
The previously published agent benchmark has been withdrawn after an audit found
that the task prompt exposed its answer key, the scored file set was narrower
than the change required by the monorepo, and one arm started in a different
working directory. Three corrected exploratory sessions have since been run;
they do not establish a performance or correctness advantage. No conclusion
from the withdrawn runs should be cited.
Audit and corrected harness →
---
The problem
AI coding tools are getting smarter. But they still have a fundamental blind spot: they don't know your architecture before they touch it.
For example, you ask an AI assistant to delete a utility file. It deletes it without warnings.
Then the build reveals downstream consumers that still import it.
The assistant saw one file without inspecting those consumers.
This isn't a model problem. It's a context problem. The AI is flying blind.
---
The infrastructure layer
Depwire is the context and safety layer for AI-generated code.
Depwire builds a dependency graph from supported source constructs using tree-sitter and serves it to AI assistants through 24 MCP tools.
How it works:
- **Local** — the CLI analyzes source code on your machine and collects no usage telemetry. It does not upload source code, file names, graph data or analysis. Applicable security scans run dependency vulnerability and supply-chain checks by default; these may contact package registries and public advisory databases. connect_repo accesses GitHub when given a GitHub URL. The What If browser page loads D3 from a CDN. Cloud is a separate service.
- **Security scanner** — runs local code checks and applicable dependency audits by default; no Depwire API key is required.
- **Token-efficient** — Depwire serves pre-computed graph data so agents can request focused dependency context instead of broad file dumps.
- **Deterministic** — tree-sitter provides consistent structural parsing without relying on model inference.
---
Start here
Three commands for a supported codebase:
---
Tested on real-world projects
| Project | Language | Files | Symbols | Edges | Health |
|---------|----------|-------|---------|-------|--------|
| google/guice | Java (multi-module, 13 modules) | 647 | 30,592 | 10,081 | 31/100 |
| honojs/hono | TypeScript | 352 | 6,462 | 2,194 | 41/100 |
| apache/commons-lang | Java (single-module) | 624 | 29,723 | 9,037 | — |
| pallets/flask | Python | 79 | 2,005 | 851 | — |
| dart-lang/shelf | Dart | 108 | 1,639 | 219 | — |
| rstudio/plumber | R | 197 | 1,194 | 219 | — |
| payloadcms/payload | TypeScript | 645 | 9,292 | 3,511 | — |
> Historical depwire parse runs on public repositories, last validated with v1.8.2 (June 2026).
>
> **Pre-1.9.0 measurement.** v1.9.0 fixed parser bugs (double-emitted symbols in the TypeScript/Python/C#/C++/Java parsers, dropped type-only-import edges, false orphans) that directly affect symbol counts, edge counts, and health scores. Later releases changed graph contents and health methodology again. These numbers have not been re-measured and are not current results.