@alex-apps/ng-token-saver
# ng-token-saver
[](https://www.npmjs.com/package/@alex-apps/ng-token-saver)
[](https://www.npmjs.com/package/@alex-apps/ng-token-saver)
[](#requirements-and-setup)
[](LICENSE)
An MCP server that lets an AI agent understand Angular **templates** by asking the same
compiler that builds the project, instead of guessing from file text. The token saving is
measured, not promised.
Angular ships a first-class language server, but no official AI integration exposes it — the
Angular CLI MCP (`ng mcp`) works at the docs-and-build level and does not touch templates.
This server is that missing layer, and it answers **for the Angular version the project
actually runs**.
Everything below marked as measured was produced by running code against six real Angular
workspaces (17.3.12, 18.2.14, 19.2.25, 20.3.26, 21.2.18, 22.0.8) and two production projects.
Every number can be reproduced with the commands in
[Reproducing the measurements](#reproducing-the-measurements).
| Measured on a production Nx monorepo | Result |
|---|---|
| A component's contract instead of its whole file | **67% fewer tokens** over 1298 components |
| Where one input is bound on a component used 577 times | **50 527 → 1 035 tokens** against the grep path |
| Errors of one template | **81 tokens, 1 ms warm** against a 395-token compiler listing |
| Whole Claude Code sessions, five read-only questions × 3 | **no saving**: +15% input tokens, −4% cost, 60 of 60 answers right in every arm — [why](#measured-whole-sessions) |
The last row stays in the table: a per-call saving is not a per-session one, and the
difference is explained below.
## Quick start
The language server ships as a regular dependency — nothing to install besides the package:
```
npm install -g @alex-apps/ng-token-saver
```
Claude Code:
```
claude mcp add ng-token-saver -- ng-token-saver
```
Or, in Claude Code, as a plugin — the same server through npx, plus the one-line
[hints](#hints-in-claude-code) toward it; no global install needed:
```
/plugin marketplace add Alex88Ryabov/alex-apps-ng-token-saver
/plugin install ng-token-saver@ng-token-saver
```
A server already registered with `claude mcp add` would then run twice; remove it with
`claude mcp remove ng-token-saver`. The plugin's tools carry a plugin prefix
(`mcp__plugin_ng-token-saver_ng-token-saver__…`), so permission rules and agent tool lists
written for the bare names need the new ones.
Codex CLI:
```
codex mcp add ng-token-saver -- ng-token-saver
```
Any other MCP client:
```json
{ "mcpServers": { "ng-token-saver": { "command": "ng-token-saver" } } }
```
Or skip the install and let the client fetch it through npx:
```json
{ "mcpServers": { "ng-token-saver": { "command": "npx", "args": ["-y", "@alex-apps/ng-token-saver"] } } }
```
In Cursor that npx form is one click:
[](https://cursor.com/install-mcp?name=ng-token-saver&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsIkBhbGV4LWFwcHMvbmctdG9rZW4tc2F2ZXIiXX0%3D)
The npm and npx paths are verified by running: the packed tarball (79 kB, dist only) was
installed into a clean prefix and all four tool kinds answered through a real MCP client,
and the npx form connects in 1.2–2.0 s from a warm npm cache (the very first run on a
machine also downloads the dependency tree — the language server alone unpacks to 13.6 MB).
A first question to ask it — the contract of a component whose members are scattered across
an extends chain. Asked for `fixtures/v17/src/app/derived-card.component.ts` (a fixture in
this repository), `ng_component_info` answers, verbatim:
```json
{"found":true,"angularVersion":"17.3.12","className":"DerivedCardComponent","kind":"component","selector":"app-derived-card","standalone":true,"inlineTemplate":true,"styleUrls":[],"imports":[],"hostDirectives":[],"extends":"BasePanel","ancestors":[{"name":"BasePanel","file":"<repo>\\fixtures\\v17\\src\\app\\base-panel.ts"},{"name":"BaseWidget","file":"<repo>\\fixtures\\v17\\src\\app\\base-widget.ts"}],"inputs":[{"name":"accent","type":"boolean"},{"name":"heading","type":"string"},{"name":"disabled","type":"boolean"}],"outputs":[{"name":"blurred","type":"void"}],"publicMembers":[{"name":"focus","kind":"method","signature":"focus(): void","noop":true},{"name":"collapse","kind":"method","signature":"collapse(animated: boolean): void","noop":true}]}
```
811 characters with this checkout's absolute paths in place of `<repo>`, 305 ms on the
session's first call (it loads the project's own TypeScript), single-digit milliseconds after.
The asked file declares one input and an `extends` clause; `heading`, `disabled`, the output
and both methods live in `BasePanel` and `BaseWidget` and are resolved statically, with their
files named in `ancestors` for the next edit. `"noop": true` on `focus()` is the subclass
shadowing it with an empty body — the kind of fact that otherwise costs a whole file read per
ancestor.
Configs for Cursor, VS Code, Windsurf, Codex CLI and JetBrains, the Node floor, and running
from source are in [Requirements and setup](#requirements-and-setup).
## The two problems it solves
**1. No template awareness.** `grep` over an `.html` file cannot tell you where
`{{ user().fullName }}` is declared, and no amount of reading gives you `TS2339 Property
'emailAddress' does not exist on type 'UserVm'`. That is compiler output, not text.
**2. Version drift.** The AI context files on angular.dev (`llms.txt`) describe only the
newest major and carry no version markers; the versioned archive sites serve none at all.
On v17–v21 they hand the agent instructions that produce APIs which do not exist. Measured
examples are in [Version facts](#version-facts).
## Tools
Six tools, 1040 characters of descriptions in total. Answers are dense JSON with no markdown.
| Tool | What it answers | Needs the language server |
|---|---|---|
| `ng_template_definition` | where a symbol on a template line is declared — the line as Read shows it, the symbol by name | yes |
| `ng_template_diagnostics` | Angular compiler errors for a template, or for a `files` batch; an entry anchored in the companion `.ts` carries `file`, a message repeated on several lines comes once with its `lines` | yes |
| `ng_component_info` | the public contract of a component or directive | no |
| `ng_workspace_map` | projects, versions, `strictTemplates` and zone.js per project | no |
| `ng_version_rules` | what exists and what does not in this project's Angular version | no |
| `ng_find_usages` | where a component, directive, pipe or service is used; with `input` — where that input is bound | no |
Four of the six never start the language server, so they answer in milliseconds and keep
working on workspaces where the server refuses to load.
Since 0.2.0 every position is 1-based, as Read shows it: `ng_template_definition` takes the
line that way and the symbol by name (`character` is an alternative for a name that repeats on
the line). 0.1.x took a 0-based `line` and `character`, the LSP convention — the same numbers
now point one line lower.
## Measured: contract instead of the whole file
`ng_component_info` returns the public contract of a component rather than its source.
Measured across two production codebases through a real MCP client:
| | Nx monorepo | CLI workspace\* |
|---|---|---|
| Angular / TypeScript | 19.2.18 / 5.8.3 | 17.3.8 / 5.3.3 |
| Components in the tally | 1298 | 407 |
| Parse errors | 0 | 0 |
| Sources | 5 404 708 chars | 1 735 223 chars |
| Contracts (base-class members included) | 1 759 331 chars | 463 563 chars |
| **Saved** | **67%** | **73%** |
| **Saved in tokens** (o200k_base proxy) | **67%** | **71%** |
| Contract shorter than source | 1201 of 1298 (93%) | 386 of 407 (95%) |
| Flagged as partial | 1 of 1298 | 3 of 407 |
| First call (loads the project's TypeScript) | 306 ms | 564 ms |
\* measured with the pre-0.1.2 wire format; the current format is leaner, so this saving
is a floor.
The largest component in the monorepo shrinks from **177 863 to 10 770 characters** while
listing 117 contract members. Contracts include members inherited from base classes — the
extends chain is resolved through relative imports, tsconfig path aliases and barrels.
Before that resolver, 91 of 1298 monorepo contracts were flagged as partial; now 1 is.
Three caveats that travel with these numbers:
- **Tokens are counted through a proxy** — OpenAI's `o200k_base`, since Claude's tokenizer
is not public; `cl100k_base` agrees within one point on this data. Characters are exact.
- **The baseline is reading the whole file**, which is what an agent does by default.
- **JSON tokenizes slightly worse than TypeScript**, so token savings sit a point or two
below character savings. The table carries both.
On small components there is no saving at all: a 17-line component produces a 578-character
contract against a 315-character source. The contract grows with the number of members, the
source with method bodies — and on production code the second wins almost always.
## Measured: rename and diagnose
**Renaming an input across usages.** The grep path an agent actually takes — read the
component file to learn the selector, then grep the selector and the binding spellings
repo-wide — against `ng_component_info` plus `ng_find_usages` with its `input` filter,
which returns only the tags that bind the name, each entry pointing at the binding itself.
On the production monorepo, 6243 files scanned:
| Component | grep path | bridge | saved (o200k) |
|---|---|---|---|
| 577 usages, `mask` bound on 5 tags | 50 527 tokens | 1 035 tokens | **98%** |
| 466 usages, `icon` bound on 463 tags | 41 698 tokens | 22 566 tokens | **46%** |
| 1211 usages, `name` bound on 1134 tags | 129 227 tokens | 30 650 tokens | 76%\* |
\* the tool returns at most 500 entries per answer, and the answer says so.
The saving is decided by how many of the usages actually bind the input: `mask` is bound on
5 of 568 tags, and grep still prints every selector line plus 67 binding-shaped lines from
across the repo — 62 of them somebody else's `mask` — while the bridge answers with exactly
those five sites.
**Diagnosing a template that will not compile.** A whole-project compiler listing costs
**395 tokens and 1321 ms**; one `ng_template_diagnostics` call answers with the asked
file's diagnostics in **81 tokens, 1 ms warm**. The gap only widens with project size: the
listing grows with the project, the answer does not. After an edit, fresh diagnostics
arrive as a 340–400 ms push — against a rebuild.
## Measured: whole sessions
The numbers above are per call. Whether they reach the bill was measured on whole sessions:
five read-only questions on the production monorepo — an inherited contract, a plain contract,
where one input is bound, where a component is used, which APIs the installed version has as
stable — each answered three times by `claude -p` (Claude Code 2.1.322, `claude-opus-5-5`) in
four arms, isolated from the machine's own CLAUDE.md, memory, hooks and MCP servers, with
read-only tools and no subagents. Every one of the 60 answers was right.
| Arm | Input tokens | Output tokens | Cost\* | Turns |
|---|---|---|---|---|
| A — no server | 1 322 306 | 20 560 | $1.27 | 65 |
| B — no server, told to read only what the question needs | 1 347 280 | 21 062 | $1.39 | 70 |
| C — B plus this server | 1 552 429 | 21 634 | $1.34 | 83 |
| D — C plus the plugin hints | 1 537 528 | 20 828 | $1.23 | 79 |
\* Claude Code's own estimate at API list prices; cached input is cheap, so cost and token
totals move apart.
Against the control (C versus B) the server **added 15% input tokens and saved 4% of cost** —
inside the run-to-run spread, which reaches 3× within a single cell. Two reasons, both visible
in the logs. Claude Code defers MCP tools, so the first call costs an extra ToolSearch turn,
and every turn re-sends some 20 000 tokens of context. And on questions like these the agent
spends little on files: for the inherited contract it read the 348-line component whole but
grepped its 671-line base class for decorators — a few thousand tokens, which one extra turn
outweighs.
The per-call savings hold where the files are large and read whole. In 81 field transcripts
on the same monorepo agents read a component or directive file whole 319 times against 99
contract calls, and 254 of those reads came from subagents — see
[Hints in Claude Code](#hints-in-claude-code). Whether the hint
turns those reads into contract calls is not measured yet. Edits, refactors and builds were
not part of this run.
## Version facts
`ng_version_rules` contains no rule taken from documentation: the data comes from importing
the packages actually installed in each fixture and from running the compiler.
**The zoneless provider is renamed between v19 and v20.**
| API | v17 | v18 | v19 | v20 | v21 | v22 |
|---|---|---|---|---|---|---|
| `provideExperimentalZonelessChangeDetection` | – | yes | yes | – | – | – |
| `provideZonelessChangeDetection` | – | – | – | yes | yes | yes |
Advice to "enable zoneless" without a version breaks on three majors out of six.
**Existing is not the same as ready.** The `@experimental` and `@developerPreview` tags live
only in declaration JSDoc and are invisible at runtime:
| API | v17 | v18 | v19 | v20 | v21 | v22 |
|---|---|---|---|---|---|---|
| `input`, `output`, `model`, `viewChild`, `contentChild` | preview | preview | stable | stable | stable | stable |
| `effect`, `toObservable` | preview | preview | preview | stable | stable | stable |
| `linkedSignal`, `afterRenderEffect` | – | – | preview | stable | stable | stable |
| `resource`, `rxResource`, `httpResource` | – | – | **experimental** | **experimental** | **experimental** | stable |
So "rewrite `@Input()` as `input()`" on a v17 or v18 project means moving to a non-public
API, and `resource()` was experimental all the way through v21. A batch of signal APIs
appears exactly at v19: `linkedSignal`, `resource`, `rxResource`, `httpResource`,
`afterRenderEffect`, `provideAppInitializer`.
**Two documentation claims that measurement contradicted:** `standalone` becomes the default
at **v19**, not v20; and `*ngIf` is not removed in 22.0.8 — it reports hint `TS6385` (TypeScript's
deprecation hint) and keeps working, with `NgIf` still exported from `@angular/common`.
**Compiler gates**, read in the 22.0.8 bundle and confirmed by running it — keyed on the
project's Angular version, which the server detects from `@angular/core` next to the tsconfig
(`--angularCoreVersion` is passed too, as a fallback):
| Feature | Gate |
|---|---|
| `@if` / `@for` / `@switch` blocks | ≥ 17.0.0 |
| signals in two-way bindings | ≥ 17.2.0-0 |
| `@let` | ≥ 18.1.0 |
| implicit `standalone` | ≥ 19.0.0 |
| DOM event type assertion | ≥ 20.2.0 |
Also measured: Signal Forms (`@angular/forms/signals`) exist only from v21 and are stable on
22.0.8; `AbstractControl.events` from v18; `TestBed.tick` from v20.
Outside the measured v17–v22 range `ng_version_rules` returns nothing and says so —
extrapolating "it was in v22, so it is in v23" is exactly the failure it exists to prevent.
## Honesty as a feature
An incomplete answer says so, in words, inside the answer:
- A contract merges base-class members and host-directive exposures, resolved statically
through relative imports, tsconfig aliases and barrels. Where the walk cannot continue —
a base class from a package, a mixin call — the answer carries `incomplete`, naming the
class and the file to ask about next.
- `ng_version_rules` reports `notMeasured` topics, and a `caveat` when your minor differs
from the measured one.
- `ng_find_usages` labels declarations as `declaration`, admits when class-name matches were
found without resolving imports, and names selector twins — a second declaration of the
same selector elsewhere in the workspace — instead of silently mixing their usages.
- `ng_template_diagnostics` separates three states that all look like an empty list: the
template is clean; the server is silently down (caught by a canary probe); or template
checking is off for this project — then the answer carries `checksDisabled` naming the
tsconfig responsible. In the measured monorepo `strictTemplates` is off in two of seven
applications, so the distinction is not theoretical.
## What this is not
- **Not better than a careful `grep` at finding usages.** `grep -rn "<app-widget"` finds the
same 62 usages. What `ng_find_usages` adds: the selector and class name are derived from
the file for you, usage kinds are labelled, all four attribute-binding spellings are
covered, and `path` scopes the scan — 33 selectors in the measured monorepo are declared
in two applications at once, and an unscoped search would mix them.
- **Not a replacement for the Angular CLI MCP** — a different layer. `ng mcp` covers docs,
best practices and build orchestration; none of its nine tools touches templates or the
language server. The two complement each other.
- **Not a type checker of its own.** The LSP-backed tools deliver the project's own compiler
output, undistorted.
Known gaps are recorded, not hidden: a base class from a package or behind a mixin call
stops the ancestor walk (1 of 1298 components in the measured monorepo), Nx projects with
inferred targets come without `tsConfig`, and pipe twins are not detected.
## Requirements and setup
- **Node ≥ 18.20.8** — the measured floor: the published package answered on clean Node
18.20.8, 20.20.2 and 22.x. This is about the MCP server process only — **your project
keeps building on its own Node**. To give just the server a newer runtime, point the
client config at that binary: `"command": "C:\\node22\\node.exe"`.
**From source** (instead of npm):
- The project's own dependencies: `npm install`, then `npm run build`.
- The shipped language-server branch lives in `tools/servers/ls22` and needs `npm ci` there once.
- In every client config, replace the `ng-token-saver` command with `node <path>/dist/index.js`;
for Claude Code: `claude mcp add ng-token-saver -- node <path>/dist/index.js`.
`node_modules` folders are not committed, including the twelve inside the stand. To restore
the full measurement environment:
```
npm install && npm run build
cd tools/servers/ls22 && npm ci # the branch actually shipped
cd fixtures/v22 && npm ci # repeat per fixture you want to run
```
### Clients
The server is a plain stdio MCP server with no client-specific features, so any MCP client
can launch it; installation and Claude Code registration are in [Quick start](#quick-start).
Usage guidance for the agent travels in the MCP `instructions` field — see
[Using with agents](#using-with-agents).
**Cursor** — the one-click button in Quick start, or the same `mcpServers` JSON as in
Quick start, in `~/.cursor/mcp.json` (all projects) or `.cursor/mcp.json` (one project).
**Windsurf** — the same JSON, in `~/.codeium/windsurf/mcp_config.json`.
**VS Code (Copilot agent mode)** — `.vscode/mcp.json`; the key is `servers` and the entry
takes a `type`:
```json
{ "servers": { "ng-token-saver": { "type": "stdio", "command": "ng-token-saver" } } }
```
**Codex CLI** — `codex mcp add ng-token-saver -- ng-token-saver`, or `~/.codex/config.toml`:
```toml
[mcp_servers.ng-token-saver]
command = "ng-token-saver"
```
**JetBrains AI Assistant / Junie** — Settings → Tools → AI Assistant → Model Context
Protocol accepts the same JSON as Quick start's; for Junie, additionally enable "Pass
custom MCP servers".
Configuration, all variables optional:
- `NG_TOKEN_SAVER_IDLE_MS` — a language-server session unused this long shuts its ngserver
down; the next call pays the cold start again. Default 900000 (15 minutes); `0` keeps
sessions alive until the server exits. A session with a call in flight is never shut down.
- `NG_TOKEN_SAVER_PREWARM=1` — `ng_workspace_map` with a `path` inside an app starts loading
that app into the language server in the background, so the first diagnostics or definition
call finds it ready: 21 s → 1.4 s on the production monorepo. The cost is ~1 GB of memory
until the idle shutdown, whether or not an LSP tool follows; a folder holding several projects
(a workspace root, an Nx `apps/`) is not warmed, since which app to load is unknown.
- `NG_TOKEN_SAVER_SERVERS_DIR` — where the language-server branch lives, if not in
`tools/servers` next to the build.
## Using with agents
The server sends its usage guidance in the MCP `instructions` field (924 characters), which
Claude Code keeps in the agent's context even when tool descriptions are deferred until a tool
search (checked on Claude Code 2.1.282: the whole text arrives while the tools show only their
names). For a client that ignores `instructions`, the same text goes into `CLAUDE.md` or
`AGENTS.md`:
> Angular tools that answer from the compiler and the source instead of whole files. Prefer
> them to reading and grepping: ng_component_info for what a component or directive accepts
> (inputs, outputs, members, inherited ones included) instead of reading its .ts;
> ng_find_usages for where a component, directive, pipe or service is used, and with input for
> where one input is bound, instead of grep; ng_template_diagnostics after editing a template
> instead of ng build; ng_template_definition for what a template symbol is; ng_version_rules
> before suggesting an Angular API; ng_workspace_map once for projects, versions and
> strictTemplates. The first diagnostics or definition call in a workspace loads the project
> into the Angular language server and takes up to a minute on a large workspace; later calls
> take milliseconds, so a slow first answer is not a hang. Lines and characters are 1-based
> everywhere, as Read shows them.
### Hints in Claude Code
The instructions alone did not change the habit. Counted with `npm run bench:adoption` over 81
Claude Code transcripts on the production monorepo, with the server installed and these
instructions in context:
| | Main sessions (53) | Subagents (28) |
|---|---|---|
| `ng_component_info` calls | 70 | 29 |
| Component files read whole (`Read`, `cat`, `Get-Content`), pre-edit Reads left out | 53 | 254 |
| `ng_find_usages` calls | 67 | 59 |
| Greps for a component element | 66 | 11 |
| `ng_template_diagnostics` calls / `ng`/`nx` builds | 112 / 15 | 13 / 0 |
Subagents are where the gap is, and a capture of Claude Code's own requests shows why: a
subagent receives the MCP instructions only after its first tool call, so its first move is
made without them. `ng_template_definition` was not called once in these transcripts, and
`ng_version_rules` twice.
The plugin's hook (`plugin/hooks/hints.mjs`) targets exactly these calls, and the counts
above use its own definition of them. Before a component or
directive file is read in full, an element selector is grepped, or an `ng`/`nx` build runs
inside an Angular workspace, it adds one line naming the tool that answers it — once per agent
and kind; the call itself runs unchanged. A subagent gets a one-sentence pointer before its
first prompt. The line arrives next to the tool result, so it steers the next call rather
than this one. `if` filters keep the hook off unrelated commands; where it does run it costs
about 95 ms, most of it Node's start.
## Reproducing the measurements
```
npm test build plus 192 unit tests (node:test, no dependencies)
npm run smoke end-to-end check with a real MCP client over stdio
npm run bench:settle whether a pause after didOpen is needed (it is not)
npm run bench:standalone where standalone becomes the default (v17..v22)
npm run bench:api which Angular APIs exist in which majors, and their stability
npm run bench:contract <project root> contract size against reading whole files (--tokens adds token counts)
npm run bench:rename <component> <input> the grep path against the bridge for an input rename
npm run bench:diagnose a compiler listing against one diagnostics call
npm run bench:matrix resolution probes across a fixture
npm run bench:negative what the server returns when things break
npm run bench:didchange diagnostics timing after an edit
npm run bench:adoption <transcripts dir> tool calls against the habits they replace, in Claude Code transcripts
npm run bench:session <tasks.json> whole-session A/B through claude -p, four arms (see the file's header)
```
The whole-session tasks name files of a production project, so the tasks file is not in the
repository; `tools/bench-session.mjs` documents its format, and any workspace will do.
The stand is `fixtures/v17..v22` — six real Angular workspaces, each with its own
`node_modules` and its own pinned TypeScript, plus `fixtures/negative/*` for failure cases.
## Status
All six tools verified on the six fixtures and on two production codebases — 1298 and 407
components, zero parse errors — plus one Angular 16 project to check that out-of-range
refusals are structured rather than silent. 192 unit tests, all green.
Measured latency: the two LSP-backed tools pay 8–28 s of cold start on the first call and
answer in 2–9 ms after it; the four static tools answer in 250–600 ms on the first call and
in milliseconds once the project's TypeScript is cached. A session idle for 15 minutes shuts
its language server down, and the next call pays the cold start again — see
`NG_TOKEN_SAVER_IDLE_MS` above.
## License
GPL-3.0-or-later — see [LICENSE](LICENSE). Copyright (C) 2026 Alex Ryabov.
Use and modify freely; derivative works must stay open under the same license.
TDQS
Scored across 6 tools
Each tool targets a distinct aspect of Angular development: component metadata, template positions, template diagnostics, workspace configuration, version-specific rules, and cross-reference usages. There is no overlap or ambiguity between them.
All tools follow a consistent 'ng_' prefix with snake_case names that clearly indicate their purpose. The naming pattern is uniform across all six tools, making it easy to predict tool functionality from the name alone.
With exactly six tools, the server is appropriately scoped for Angular code analysis. Each tool serves a distinct purpose without redundancy, and the count is within the ideal range.
The tool set covers the core workflows for Angular analysis: understanding components, navigating from templates to definitions, checking diagnostics, mapping workspace structure, understanding version features, and finding usages. Minor gaps like batch processes are addressed by the batch capability in diagnostics, and the set feels complete for a read-only analysis server.