Build agents that understand before they act.

One family. Four public products. Explicit trust boundaries.

Core is network-free and source-read-only. SEO is a Search Evidence Graph over the live site and the code that produced it. Refactor adds reviewed local writes. Online adds an authorized network connection. All four public packages are MIT.

$ npx -y weavatrix mcp <repoRoot>
$ cargo install weavatrix

$ npx -y weavatrix-seo audit --site https://example.com --json
$ npx -y weavatrix-seo mcp --allow-root .

Clear boundary: Core, SEO, Refactor, and the public Online connector are MIT. SEO reads the origins you name; it does not write pages. Write authority still requires Refactor's gates.

The protocol-independent Rust engine beneath Core.

weavatrix-rust@2.9.0 is the reusable repository-intelligence engine and library. It does not implement an MCP server; weavatrix@1.10.0 owns the MCP adapter and ships the same native product through Cargo and npm.

[dependencies]
weavatrix-rust = { version = "2.9.0", default-features = false }

The basic 2.0.2 verification uses the native engine against its committed 192-file repository: 1,531 nodes and 7,287 edges. Coverage also records 80.57% of functions and 85.30% of regions. The short load run is intentionally bounded rather than presented as an exhaustive performance study. Reproducible methodology and component evidence remain published with the Rust repository.

The graph your agent is missing

This is not one report or three fixed workflows: 44 methods project the same reusable graph into the smallest view needed for each task. Grep sees text; Weavatrix sees structure: files, symbols, runtime edges and compile-time coupling — including Rust module paths, Java OOP relationships, project-internal Java receiver call flow, and Swift same-module types with HTTP/WebSocket clients. Typed adapters build semantic evidence for code, contracts and infrastructure, while the lossless parser preserves every supported source surface without pretending that CSS and Rust have the same semantic depth. A bounded, package-pinned TypeScript/JavaScript LSP overlay verifies ambiguous references locally without a native compiler or network request.

What the agent actually gets

Six real task shapes from local dogfooding. The output stays bounded, exposes uncertainty, and carries the same graph evidence from orientation through verification.

1. Graphmodules, symbols, routes
2. Contextexact source workset
3. Impactruntime and type radius
4. RatchetsHealth, clones, architecture
5. VerdictPASS, BLOCKED or incomplete evidence

› Where does attack mitigation start?

module_map       → product territories
list_endpoints   → 462 observed routes
trace_endpoint  → mount → controller
                  → service → task/messaging

Result: one bounded call path with
call-site excerpts, not a repository dump.

› Can this handler be removed?

trace_api_contract
  3 registered repository graphs
  163 backend endpoints observed
  267 client callsites joined

NOT_DEAD_EXTERNAL_USE
dynamic URLs remain POSSIBLE / unproven

› What breaks in this BranchPilot diff?

change_impact
  removed methods / signatures separated
  additive exports kept low-risk
  98 impacted nodes observed
  runtime vs type-only radius retained

Then: context_bundle → edit →
verified_change phase=verify

› Show actionable debt, not boilerplate.

run_audit
  missing direct dependency: mongodb
  runtime cycle evidence + lockfile drift
find_duplicates
  repeated nudge() family observed
  router boilerplate suppressed by default

Every item stays review-required;
no auto-delete or cosmetic green score.

› Who produces and consumes this event?

query_graph
  seed_symbols=["handleAttackEvents"]
  relation_filter=["calls","references"]
  flow_direction="both"
context_bundle → production callers first

Exact seeds keep tests and fuzzy config
matches out of the bounded flow.

› Preserve this architecture from now on.

get_architecture_contract action=preview
  adaptive source roots + candidate rules
  observed directions: NOT ENFORCED
  exact content + hash + one-time token
action=approve confirm_token=…

Creates only a missing local contract;
never overwrites active policy.

Release proof at the boundary we can reproduce.

The 2.0.2 engine was exercised against its own repository: 192 files became 1,531 nodes and 7,287 typed edges. Three clean graph builds establish the basic latency range; an unchanged refresh, literal search, and 1,000 hot graph_stats calls cover the short-load path. The installed Core 1.1.2 package then passed the same bounded MCP check.

Measured July 30, 2026 with the native 2.0.2 engine executable. This is deliberately a basic release verification plus short load, not a claim about every repository, machine, or workload. Methodology and reproducible evidence remain versioned with the engine; installed-package evidence remains with Core.

Intelligence that grows with the software.

The long-term direction is an evidence substrate for increasingly capable agents: a repository becomes a living model of code, intent, decisions, runtime signals, and verified outcomes. Models may change; durable evidence, explicit uncertainty, and human-controlled trust boundaries remain.