A live map of how code moves through a repository.
Eyrie parses a repository locally, builds a code knowledge graph, derives a small architecture graph from it, and presents it as an interactive map that stays synchronised with the source. When a file is saved, only the affected region is recomputed and the change animates into the map.
System lens over the bookstore fixture. Captured from the working Eyrie application.
View full-size captureStill true after an hour of editing.
That is the promise Eyrie is built to keep: the map follows the code, save by save. The captures below show the working tool against a bookstore fixture: the four lenses, the Health findings, then the change tape after a save.
These captures show the working tool against a bookstore fixture. Open either image at full size to inspect the interface.
The video walkthrough is not available yet. The captures above show the current UI.
The list describes documented behaviour. Still captures do not demonstrate animation or agent execution.
Five stages. One traceable map.
Follow the source from syntax facts to a live architecture map. The first three panels explain the mechanism; the final two show the working tool.
tree-sitter, against the checkout on your disk
Eyrie reads the repository already open on your machine and extracts syntax facts with tree-sitter. Sixteen language groups are supported, including Kotlin, Swift, C/C++, Dart, Vue, Objective-C and Lua. TypeScript and JavaScript still have the deepest support.
export function listOrders() {
return db.query(orders)
}listOrdersdb.queryevery edge traces to a parsed fact
Each relationship between symbols comes from that analysis and carries a confidence score. There is no inference by resemblance and no model in this path — the same commit yields the same graph, every time.
OrderRoutecalls OrderServiceapp/api/orders/route.ts:14listOrders()call_expressionstable identity over fresh layout
A node's id is a hash of repository, file path, symbol kind and qualified name. Unchanged code keeps its identity across edits, so the map does not reshuffle under you. The cache is a single file you can delete at any time.
hash(repo, path, kind, name)~/.eyrie/repos/<hash>/graph.dbsix lanes, read in request order
Presentational components, hooks, helpers and SDK wrappers are lifted out, and edges pass through them — leaving the components that actually carry the architecture. Classification follows framework conventions and names, so it is heuristic; a doc comment on a component's main export overrides what was derived.
on save, only what moved
Eyrie invalidates the files that changed, plus the files that consulted them — and those only when their exported surface changed. It emits an exact diff, animates it into the map, and records it on the change tape. This is the part that stops a diagram from going stale.
Four fundamentals,
made concrete.
On save, only what moved
Eyrie invalidates the files that changed, plus the files that consulted them — and those only when their exported surface changed. It emits an exact diff, animates it into the map, and records it on the change tape. This is the part that stops a diagram from going stale.
Every edge traces to a parsed fact
Each relationship comes from tree-sitter analysis and carries a confidence score. There is no inference by resemblance and no model in this path — the same commit yields the same graph, every time.
OrderRoutecalls OrderServiceapp/api/orders/route.ts:14listOrders()call_expressionAn agent can ask what it is about to break
Codebases are modified by agents now, and the same graph is reachable over MCP: twelve tools, from impact before an edit to recent changes after it. One command wires it into the agent with eyrie init-agents.
$ eyrie mcpimpactcomponents_for_filesrecent_changesOn your machine — with the exception named
AI notes are off by default, and in that state Eyrie makes no network requests at all: a loopback server with a per-run token, no accounts, no telemetry. Turn notes on and your own CLI sends commit metadata — never source code — under your account.
127.0.0.1Not the only way to read a codebase. Just the one that stays true.
Eyrie refuses one thing the category accepts: a picture that goes stale the day it is produced. How that refusal compares, row by row.
Eyrie’s column is highlighted. Scroll horizontally to compare all categories.
| Capability | EyrieIn development | Static diagram generators | IDE dependency views | AI coding assistants |
|---|---|---|---|---|
| Stays true while you edit | Yes — on save, only the affected region is re-resolved and the diff animates in. | No — a new export each time; stale between runs. | Partial — reflects the open workspace, not a map you keep. | No — answers come from a snapshot of context, not a live model. |
| Every relationship traces to a parsed fact | Yes — tree-sitter analysis with a confidence score; same commit, same graph. | Varies — many infer from imports alone. | Yes — from the compiler's own model of the code. | No — asserted from patterns; edges can be hallucinated. |
| Stable identity across edits | Yes — a node id hashes repository, path, kind and name. | Rarely — layouts typically rebuild from scratch. | Not applicable — views rebuild; nothing persists. | No — there is no persistent model to keep stable. |
| Coding agents can query it | Yes — twelve MCP tools over stdio, onboarded with eyrie init-agents. | No — pictures have no query surface. | Rarely — usually no agent-facing API. | Partial — the agent is the interface; there is no inspectable graph. |
| Runs locally without an account | Today, yes — loopback server, no telemetry; AI notes are opt-in. | Varies — some local, some hosted. | Yes — local by nature. | Usually no — inference typically runs in the cloud. |
| Edits your code | No — Eyrie reads; your editor or agent writes. | No. | Yes — it is an editor. | Yes — generation is the point. |
What is typical of each category; verify per tool. The Eyrie column follows its own record: in development, launch status Hold.
Three questions your codebase should answer in a minute.
The map answers with impact: every caller, up through the route that reaches it — before you touch a line.
answers with · impact
The change tape lists exactly what moved since your last look: added, modified, removed — each wired to its place on the map.
answers with · recent_changes
Trace the path from the route to data. An auth guard either sits on that path or it does not; the graph can prove which.
answers with · trace
$ pnpm start -- /path/to/repo→ http://127.0.0.1:4319 · requires Node.js 22.13+ and pnpm 9From source. There is no published binary and no npm package yet, and the desktop build — macOS only, unsigned — is not distributed.
Sample answers describe documented behaviour; component names are illustrative.
One graph, asked four different questions.
Every pointer action has a key
An agent can ask what it is about to break.
Codebases are modified by agents now, and the same graph is reachable over MCP. An agent can ask what a file belongs to before an edit, and what its edits changed afterwards.
On your machine — and here is the exception.
Turned on for a project, AI notes run your own Claude Code or Codex CLI, which sends commit messages, author names, changed paths with line counts and component names — never source code — to Anthropic or OpenAI under your account. Neither model can act on the machine. It is opt-in per project, and the content is named before you turn it on.
One command, against a checkout you already have.
$ pnpm start -- /path/to/repo→ http://127.0.0.1:4319Requires Node.js 22.13+ and pnpm 9. From source. There is no published binary and no npm package yet, and the desktop build — macOS only, unsigned — is not distributed.
An eyrie is a bird of prey's nest built high on a crag — the vantage point you watch the whole territory from.