mcp-server

The MCP server for abap2UI5 — gives any AI coding agent (Claude Code, Cursor, VS Code Copilot, or any MCP client) the full abap2UI5 development loop, without an SAP system:

examples -> app_guide -> validate_view + screenshot_view -> deploy_app -> build_backend -> run_app -> pitfalls
(has somebody  (how an app  (SECONDS, no system:        (write ABAP,  (transpile      (boot headless,  (what a green
 built it       is built)    is the view legal,          lint)         to Node)        errors +         run still
 already?)                   and what does it LOOK like)                               SCREENSHOT)      does not prove)

The agent writes an ABAP class, validates the view and looks at a picture of it in seconds, deploys it, boots it in a real browser and looks at the running app — then iterates. Everything runs locally on infrastructure that already guards the abap2UI5 ecosystem in CI: the abaplint transpiler + open-abap runtime, the framework's express shim, the samples-controls build and boot gates, and the linter validation core.

Documentation

→ The MCP server, in full — what MCP means here, the three setup levels and what each one buys, how to register the server with your client, every tool with what the agent gets from it, and the loop they are meant to be used in.

→ Building with AI — the whole AI setup in rising order of effort. This server is the top rung; the cheaper ones matter first.

Related MCP server: UI5 MCP Server

Quick start

One-click install

Every button registers the same stdio server, named abap2ui5:

{"command":"npx","args":["--yes","-p","@abap2ui5/mcp-server","abap2ui5-mcp"]}

The badges go through a web redirect because GitHub does not render links to an editor's own URL scheme; the direct links, to paste into a browser or xdg-open/open, are:

vscode:mcp/install?%7B%22name%22%3A%22abap2ui5%22%2C%22command%22%3A%22npx%22%2C%22args%22%3A%5B%22--yes%22%2C%22-p%22%2C%22%40abap2ui5%2Fmcp-server%22%2C%22abap2ui5-mcp%22%5D%7D
cursor://anysphere.cursor-deeplink/mcp/install?name=abap2ui5&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyItLXllcyIsIi1wIiwiQGFiYXAydWk1L21jcC1zZXJ2ZXIiLCJhYmFwMnVpNS1tY3AiXX0%3D

(vscode-insiders: in place of vscode: for Insiders.) Claude Code:

claude mcp add abap2ui5 -- npx --yes -p @abap2ui5/mcp-server abap2ui5-mcp

Claude Code plugin. The abap2UI5 plugin bundles the framework's agent skills (building an app, the ABAP and UI5 pitfall catalogues, the view-chain layout, ...) together with this server, registered the same way as above — one install for both:

/plugin marketplace add abap2UI5/abap2UI5
/plugin install abap2ui5@abap2ui5

Releases from 1.0.0 on are also listed in the official MCP Registry as io.github.abap2UI5/mcp-server (server.json), for clients that install from the registry. The same command is what the Smithery (smithery.yaml) and Glama (glama.json) directories list: one npm package, one version, whichever catalogue a client reads.

That is level 1 below: validate_view and fix_view work out of the box, and setup_status tells the agent what else works on this machine and what is missing.

What each install brings

Level 1 — validate_view, fix_view and screenshot_view, the tools most work happens at. They need the linter, which the server declares as a peer dependency (@abap2ui5/linter, >=0.8.0 <0.10.0 — the range this server is built against). npm 7+ installs a non-optional peer by itself, so the one-liner below brings the linter along and validate_view/fix_view work out of the box. The render gate behind screenshot_view needs the UI5 libraries and Playwright on top (@abap2ui5/linter-render, declared as an optional peer: the compatible range is stated, the package is NOT installed for you — plan on ~150–200 MB and a few minutes the first time you add it):

claude mcp add abap2ui5 -- npx --yes -p @abap2ui5/mcp-server abap2ui5-mcp                            # validate_view, fix_view
claude mcp add abap2ui5 -- npx --yes -p @abap2ui5/mcp-server -p @abap2ui5/linter-render abap2ui5-mcp # + screenshot_view

The server looks for the linter, in this order: AI_VIEW_CHECK_HOME; a linter checkout next to the server (which is also where npm puts the peer for an npx run: node_modules/@abap2ui5/{mcp-server,linter}); the project it is started in (node_modules/@abap2ui5/linter — app-template has it as a devDependency, so npm install there is all it takes); the server's own node_modules; whatever Node's module resolution finds from the server's location. A set env var decides alone. Or work from a checkout:

git clone https://github.com/abap2UI5/linter        # AI_VIEW_CHECK_HOME
cd linter && npm ci

The server itself is on npm, so it needs no checkout; the registration above is the whole install. (With npm 6 or --legacy-peer-deps, add -p @abap2ui5/linter yourself.)

(-p … abap2ui5-mcp names the bin explicitly, which every published version answers. The shorter npx --yes @abap2ui5/mcp-server needs a bin named after the package, which 0.2.0 and earlier do not have — npx stops there with "could not determine executable to run".)

(The install is ~47 MB: 19 MB of it a Playwright driver only run_app uses, 2 MB the linter — paid on the first start, cached after. From a checkout instead: git clone https://github.com/abap2UI5/mcp-server && cd mcp-server && npm ci, then claude mcp add abap2ui5 -- node /path/to/mcp-server/server.mjs.)

Cursor, VS Code and Claude Desktop take the standard stdio shape — the documentation has the JSON, and the two further levels (the sample catalogues and deploying, then the headless build-and-boot loop). A tool whose prerequisites are missing answers with a message naming what it needs; the server starts either way.

The abap2UI5 VS Code extension registers this server for you, and adds a second one of its own for the tools that need a real SAP system.

Everything in one command — for a machine (or Codespace) dedicated to the full build-and-boot loop, setup.sh clones the framework, corpus and linter checkouts next to this repo, installs their dependencies and the headless browser (existing checkouts are reused, safe to re-run):

git clone https://github.com/abap2UI5/mcp-server && ./mcp-server/setup.sh

setup.sh --no-corpus clones the framework and the linter only, which is the whole loop for an app of your own since the framework sandbox exists; the corpus is for scope_of, the full build and the locally served UI5.

A Claude Code started inside the checkout picks the server up automatically via the committed .mcp.json; the devcontainer runs the same setup on create.

Level 0 — no checkout at all. The knowledge tools (app_guide, api_reference, pitfalls, capabilities, examples, read_example, docs_search, scaffold_app, generation_rules) read committed files, so when no checkout resolves and no env var is set they read them from GitHub instead: the files land in a per-user cache (<tmp>/abap2ui5-mcp-remote, a day at a time - used only while it is a directory of the user's own, created 0700; A2UI5_MCP_REMOTE_DIR moves it) that the server treats as a read-only checkout; add_agent_setup reads the template from there too, and writes only into the project it is given. And the expensive half runs on the npm package @abap2ui5/node-runtime (below), so npx --yes -p @abap2ui5/mcp-server abap2ui5-mcp in a fresh project covers deploy_app, build_backend, run_app, interact_app, run_unit_tests and verify_app too, with no clone of anything. What still needs a checkout is the corpus' part (scope_of, build_backend mode full, the locally served UI5). A2UI5_MCP_REMOTE=0 or A2UI5_MCP_OFFLINE=1 switches the mirror off. The docs mirror lists its pages through GitHub's API, which allows 60 unauthenticated requests an hour per address; a GITHUB_TOKEN (or GH_TOKEN) in the environment raises that and is sent to api.github.com only.

The backend without a framework checkout: @abap2ui5/node-runtime. The framework is published on npm already transpiled, with the ABAP sources of the same release commit to transpile apps against. With no abap2UI5 checkout (and none configured), build_backend installs that package once per release into ~/.abap2ui5-mcp/runtime/<version> — with the transpiler it names, express, and app-template's abaplint for the lint, at exact versions with a lockfile, --ignore-scripts — fetches open-abap-core at the commit the release was built with, and then transpiles only the deployed apps against it: seconds, never the framework. deploy_app writes into ~/.abap2ui5-mcp/sandbox (and lints with app-template's own config against the package's sources), run_app boots the package's server with the apps registered, run_unit_tests runs the apps' own tests. The release is the registry's latest (asked through npm, cached a day) unless A2UI5_MCP_RUNTIME_VERSION pins one. Measured here with a cold npm cache: the first install about 10 s, open-abap-core about 1 s, then 7–8 s per build and about a second for the unit tests. setup_status shows the release, the workspace and what the next build would do.

With a framework checkout (A2UI5_HOME or a sibling) everything works as before: build_backend downloads the release's prebuilt backend (backend-<version>.tar.gz) into the checkout the first time and re-transpiles the deployed apps from its node/zz_dev afterwards. A2UI5_MCP_BACKEND=npm runs on the package beside a checkout; build_backend mode prebuilt (or A2UI5_MCP_BACKEND=clone) clones the release into ~/.abap2ui5-mcp/abap2UI5 instead — the path this server took without a checkout until the package existed. The full corpus build stays available as mode: "full" for a checkout on an unreleased commit.

Tools

Every tool reads live from a sibling checkout, and each one needs a specific sibling — there is no "optional" repository, only tools you do or do not use. The Needs column says which checkout a tool is dead without: the linter alone carries validate_view, fix_view and screenshot_view (the fast loop, where most iterations happen — and the one dependency npm installs with the server, as its declared peer), the framework checkout carries the guide, the pitfalls and the interface (all three mirrored from GitHub when it is absent), the npm package @abap2ui5/node-runtime carries the backend when no framework checkout is there, and the corpus carries almost everything else. A tool whose checkout is missing answers with the clone command and env var that fix it.

verify_app is the loop in one call: validate, deploy, build, unit tests and boot, stopping at the first stage that fails and reporting every stage before it. interact_app is run_app with hands: after the boot it clicks, fills and presses through a short script and photographs the result, which is how the event branch of an app becomes visible without a system; run_unit_tests runs the test classes deploy_app wrote beside the app (a local z2ui5_if_client double, see the app guide's chapter 9) in the open-abap runtime and answers with assertions.

scaffold_app and add_agent_setup both execute app-template's own template.json, for the two kinds of project. scaffold_app hands back a whole NEW project, app class included, and writes nothing. add_agent_setup is for the abapGit repository that already exists - most of them never started from the template - and does what npm create abap2ui5-app -- --agent-setup does there, by the template's agentSetup key: it writes the agent setup and the two gates into project_dir (default: the server's working directory, when that looks like a project), never overwrites a file the project has (each one is listed as skipped), merges package.json and .gitignore by only adding what is missing (each value kept that differs from the template's is a warning), points the configs at the source folder .abapgit.xml names and writes nothing into it. A second run changes nothing; dry_run: true answers the same report without writing. Every path it writes comes from the template's file list and is checked first - a plain relative path, a shared file of the template, outside the source folder; one that fails refuses the whole call before anything is written. A symbolic link, or a folder that leads out of the project through one, is never written through: that file is skipped, and the answer says why.

Every answer fits the client. A client shows the agent nothing of a tool result over its cap (Claude Code: 25,000 tokens), so no answer is larger than about 60,000 characters. The tools whose answers grow with upstream content page instead of cutting: matches counts every hit, returned the page, and more names the exact call for the rest (offset - examples, capabilities, api_reference, app_guide, pitfalls, docs_search; files - scaffold_app; from_line - read_app, read_example; build_log names its next offset under cut). A list that is shortened rather than paged says so in a ...Cut field (findingsCut, rulesCut, errorsCut, ...), and a last-line backstop trims anything that still passes the cap and marks it (__answerGuardCut). The resources follow the same budget (below).

examples degrades per catalogue instead of failing: it searches the checkouts it finds and names the ones it could not, so a thinner answer never reads as "nobody has built this". It reads each repository's committed catalogue.json where the checkout has one — which is what carries a control port's verification status (checked over reviewed over generated, used to break ranking ties), the learning-path stage, and what a stack sample needs from the system — and falls back to parsing SAMPLES.md on a checkout from before that file existed. Where samples and samples-controls commit catalogue-derived.json beside it, a query also matches the control types each sample's view builds, so sap.m.Dialog finds every sample that builds one, not only those filed under it. screenshot_view and run_app answer the same question at three orders of magnitude apart: the first photographs the reconstructed view with no backend, the second the running app after a build. Most iterations should end at the first.

Operating an app without a browser

app_start, app_act and app_describe make every abap2UI5 app agent-operable: they speak the JSON protocol the UI5 frontend speaks — the app start, the event with its arguments, the model delta of what was typed — against the local backend, and answer with an agent snapshot derived from the response's view XML and model: the fields an agent may fill (model path, label, kind, current value, choice values, editable), the actions it may fire (event name, static and row-dependent arguments), the tables (columns, the first rows, selection - a SelectDialog/TableSelectDialog is a table too, and app_act with row picks from it), the messages (toast, message box, MessageStrip, field value states, MessagePopover and MessageView items) and some static text. An act is validated against the snapshot before anything is sent: an unknown event, a field that is not on the screen or not editable, a choice outside its values is refused with the list of what is allowed. A short session against z2ui5_cl_smp_app_009 of abap2UI5/samples:

app_start { app: "z2ui5_cl_smp_app_009" }
  -> fields f1..f5 (f3 "Input with value", /S_SCREEN/COLOR_02, text, ""),
     actions a1 POPUP_TABLE_VALUE (valueHelpRequest of f3), ..., a5 BUTTON_SEND
app_act { session, values: { f4: "Smith" }, event: "POPUP_TABLE_VALUE" }
  -> layer "popup", table t1 /T_SUGGESTION_SEL (6 rows, Single, editableCells [SELKZ]),
     action a1 POPUP_TABLE_VALUE_CONTINUE
app_act { session, values: { "/T_SUGGESTION_SEL/2/SELKZ": true }, event: "POPUP_TABLE_VALUE_CONTINUE" }
  -> layer "main", f3 = "BLACK", f4 = "Smith", message { toast, "value selected" }

The snapshot shape is shared with the VS Code extension and the ABAP agent addon; docs/agent-snapshot.md is its reference — the derivation rules, the operations, the deviations from the original contract and what the snapshot cannot see yet. interact_app stays the tool for what only a browser shows (the rendered page, client-side behaviour).

The screen in the chat: MCP Apps and Adaptive Cards

The snapshot is text for the agent. Two optional surfaces show the same screen to the user in the chat, and send what the user does back through app_act - the same validation as the agent's own acts, so nothing the snapshot does not allow can be fired, and the agent sees the act:

• MCP Apps (SEP-1865, stable 2026-01-26, extension io.modelcontextprotocol/ui). app_start, app_describe and app_act name the UI resource ui://abap2ui5/app-screen (_meta.ui.resourceUri, MIME type text/html;profile=mcp-app) when the client advertises the extension in its initialize. A host that renders MCP Apps (the extension's own README lists Claude, ChatGPT, VS Code and Goose) shows it as a sandboxed iframe beside the result: fields to fill, buttons, row actions, table cells, messages. A click is a tools/call of app_act through the host, built from the snapshot on screen; afterwards the page tells the model what happened (ui/update-model-context) and renders the new snapshot. The page is one self-contained HTML document - no external URL, no network, no eval - so it runs under the spec's restrictive default CSP; it needs nothing else. A2UI5_MCP_UI=on declares the screen for a host that renders MCP Apps without advertising the extension, off never; the tools' text answers do not change either way.

• Adaptive Cards for Copilot/Teams-style hosts: format: "adaptive-card" on app_start, app_describe and app_act (or A2UI5_MCP_APP_FORMAT=adaptive-card for every call) adds the screen as an Adaptive Card 1.5 - an embedded resource of type application/vnd.microsoft.card.adaptive after the unchanged snapshot - rendered by abap2UI5/protocol's renderer (vendored under lib/vendor/adaptive-cards). Off by default. Each Action.Submit carries session (and row for a row's action); a submitted payload maps to one app_act call: event and row as they are, every input whose value changed as values["<binding path>"], a message box button as the box's close action with the button as its $action argument, the popup/popover close wire as @CLOSE_POPUP / @CLOSE_POPOVER (cardSubmitToAct in lib/adaptive-card.mjs; the full table is in docs/agent-snapshot.md).

Migrating a classic report

migrate_report takes the source of a classic report (REPORT, PARAMETERS, SELECT-OPTIONS, the event blocks, a WRITE list, an ALV) and its .prog.xml text pool, and answers the class of the abap-cloud-gui addon that keeps its programming model — selection_screen( ), start_of_selection( ), write( ), alv( ), message( ) — together with the migration report (the TODOs, what the list does not carry over) and the statements of the user interface it refused with file:row:col (a dynpro it cannot build, reading the list back by index). It converts the user interface only: the logic of the report - its tables, function modules, native SQL, batch input - is copied as it is, and whether it runs on the target system is yours to check. A report that carries open-abap's @KERNEL escape (WRITE '@KERNEL ...'., JavaScript once transpiled) is refused before anything is converted, and a dynamic WHERE (cond) is converted but flagged under warnings - the runtime evaluates it with eval() (SECURITY.md). The converter is the addon's own report2UI5 (tools/report2ui5; called report2cloud in the addon's 1.0.0, whose path still loads), run from a checkout of the addon — it is not on npm:

git clone https://github.com/abap2UI5-addons/abap-cloud-gui   # ABAP_CLOUD_GUI_HOME
cd abap-cloud-gui && npm ci

With deploy: true the class is also written into the dev sandbox together with the addon's runtime (src/01, its database tables included - the backend creates a deployed table at boot) and the popups it calls (a checkout of popups at POPUPS_HOME, else the addon's .deps/popups, ../popups beside it or the addon's build/popups, in that order), the backend is built, and the answer carries app_start's snapshot of the selection screen — app_act with CGUI_EXECUTE runs the report. The database tables a report reads are not in the local backend: the screen runs, a run that reads them does not.

System mode: operating apps on a real SAP system

Everything above runs without a system. With A2UI5_MCP_SYSTEM_URL set, the same server is something else: the app tools against a real SAP system, logged on as the configured user — for MCP clients that are not the VS Code extension (whose own system server does this for the clients of its window): Claude Desktop, Claude Code, Cursor, any stdio client. The sandbox tools are not offered in this mode; register the server twice, once with and once without the variable, to have both.

Claude Desktop (claude_desktop_config.json: Settings → Developer → Edit Config), the sandbox server beside the system one:

{
  "mcpServers": {
    "abap2ui5": {
      "command": "npx",
      "args": ["--yes", "-p", "@abap2ui5/mcp-server", "abap2ui5-mcp"]
    },
    "abap2ui5-dev-system": {
      "command": "npx",
      "args": ["--yes", "-p", "@abap2ui5/mcp-server", "abap2ui5-mcp"],
      "env": {
        "A2UI5_MCP_SYSTEM_URL": "https://host:44300/sap/bc/z2ui5?app_start={class}&sap-client=100",
        "A2UI5_MCP_SYSTEM_USER": "DEVELOPER",
        "A2UI5_MCP_SYSTEM_PASSWORD_CMD": "security find-generic-password -s abap2ui5-dev -w"
      }
    }
  }
}

Claude Code, the same as one command:

claude mcp add abap2ui5-dev-system \
  -e A2UI5_MCP_SYSTEM_URL='https://host:44300/sap/bc/z2ui5?app_start={class}&sap-client=100' \
  -e A2UI5_MCP_SYSTEM_USER=DEVELOPER \
  -e A2UI5_MCP_SYSTEM_PASSWORD_CMD='security find-generic-password -s abap2ui5-dev -w' \
  -- npx --yes -p @abap2ui5/mcp-server abap2ui5-mcp

What to know before switching it on:

• It acts for real. An app_act event runs on the system as the configured user and may save, post or delete data. Use a development system and a user whose authorizations fit what the agent may do; the agent sees only what that user may see.

• A rejected logon is sent once. SAP locks a user after a few failed logons, and an agent retries. After a 401 nothing more is sent with that password — every tool says so — until the server is restarted with the corrected one (Claude Desktop: quit and reopen), or the password command answers a different one.

• Basic authentication only. SAML, X.509 or SSO logons are not supported in this mode; the VS Code extension's system server (its auth proxy) is the way there.

• One system per registration. Two systems are two entries with two names.

• No proxy support. The requests go straight to the host; HTTPS_PROXY is not read.

• The protocol client is the sandbox's (lib/appclient.mjs): it does the CSRF token handshake and keeps the stateful session's sap-contextid itself; this mode adds the logon, the system's cookies and the start location (lib/system.mjs). app_list needs the ADT services and the user's authorization for them — app_start works without both when the class name is known.

Unit tests in CI, without a system

The same runtime runs an app repository's ABAP Unit tests in GitHub Actions (or at a terminal): @abap2ui5/node-runtime at the release the project's abaplint.jsonc pins, installed once and cached, the classes transpiled against it, the tests run through the generated runner — no framework clone, none of the framework's devDependencies, and no npm ci of this package either (the runner imports none of its dependencies). Against app-template's starter app, measured on one machine: 16 s cold and 8 s with the cache, where the clone of 0.2.0 took 30 s and 23 s and left a 231 MB workspace instead of 67 MB.

- uses: abap2UI5/mcp-server@v1
  with:
    paths: src

framework: X.Y.Z pins another release than the project's, backend: clone takes the old path (the release cloned, its backend downloaded or built) — which is also what a pin older than the package (1.145.0) gets by itself.

@v1 is a floating major tag: the release workflow moves it to every 1.x release it publishes, so a workflow naming it picks up fixes without picking up a breaking change. It is created by the 1.0.0 release; until that release is out, pin @v0 (0.3.0, where it stays) or a release tag such as @v0.3.0.

npx -p @abap2ui5/mcp-server abap2ui5-unit src     # the same, locally

Every class and interface under paths is deployed with all of its files (an app or not; test and local-class includes too), and every test include runs. On the package the run uses a sandbox and a build of its own, so an MCP session's apps on the same machine are neither part of it nor touched by it. The result is the job's verdict plus a step summary naming every test method and the first failure; an object that cannot be deployed (a namespaced name) fails the run rather than leaving its tests out. app-template ships the job in its check.yml and the command as npm run test:unit. What the runner cannot see is what the open-abap runtime cannot model (see pitfalls, area abap); a test that passes here passes on the system short of that, and a PARTIALLY IMPLEMENTED test double has to implement every method the code under test calls, because the runtime generates no empty stubs.

Resources

The knowledge documents behind those tools are also MCP resources, for clients that surface them (context pickers, attach-a-document UIs) and for agents that want a document whole instead of sliced. Same live reads from the same sibling checkouts: listing is free (no checkout needed), reading a resource whose checkout is missing answers with the same actionable error the tool gives. A document longer than a client shows of one read (about 60,000 characters - today the abap-check catalogue and the capability map) is cut at a section heading, and its last line names the tool call that reads on from the first section left out (pitfalls { area, offset }, app_guide { offset }, capabilities { status }). The one resource that is not a document is ui://abap2ui5/app-screen, the MCP Apps screen of the app tools (see "The screen in the chat" above); it is part of this server and needs no checkout.

Prompts

Two prompts — one per job this server serves — put an agent straight into the loop instead of leaving it to reconstruct the order from the tool descriptions alone. Each renders an orchestration script over the tools above and duplicates none of their content:

• build-an-abap2ui5-app (argument: task, what the app should do) — orient with examples/capabilities, learn the shape from app_guide, write the class, iterate through validate_view/screenshot_view in seconds, prove it with deploy_app → build_backend → run_app, close with pitfalls.

• port-a-ui5-sample (argument: sample, the demo-kit sample) — the corpus job: generation_rules as the brief, scope_of and capabilities before writing, neighbouring ports from examples, then the same validate/screenshot/deploy/run loop.

Notes

• Dev sandbox: deployed apps land in the samples-controls checkout's gitignored src/zz_dev/, else the abap2UI5 checkout's gitignored node/zz_dev/, else ~/.abap2ui5-mcp/sandbox (A2UI5_MCP_WORKSPACE moves it) — nothing an agent deploys can leak into a commit.

• Backend: without a framework checkout the npm package @abap2ui5/node-runtime — A2UI5_MCP_RUNTIME_VERSION pins its release (default: the registry's latest, asked once a day), A2UI5_MCP_BACKEND=npm uses it beside a checkout too, A2UI5_MCP_BACKEND=clone restores the framework clone as the default. Everything it installs lives in ~/.abap2ui5-mcp and is safe to delete.

• Port: the backend listens on 3000 (A2UI5_MCP_PORT overrides).

• The lint: deploy_app lints with app-template's abaplint config. On the npm backend a long-lived worker keeps @abaplint/core (installed beside @abaplint/cli at the same version) with the framework parsed, so a deploy's lint takes milliseconds instead of a fresh abaplint process (2-3 s, mostly the framework's parse); whenever the worker cannot stand in for the CLI - a checkout without that core, a config it does not mirror - the CLI runs as before. A2UI5_MCP_LINT_WORKER=0 switches the worker off; A2UI5_MCP_LINT_IDLE_MS (default 10 minutes, 0 = never) is how long it is kept without a request - with the framework parsed it holds about 300 MB - before it exits and the next lint starts a fresh one.

• Timeouts: every spawned child is killed (whole process tree) when it exceeds its limit — lint/scope 5 min, unit tests 10 min, build (with the prebuilt download and the npm install) 30 min by default; A2UI5_MCP_LINT_TIMEOUT_MS, A2UI5_MCP_SCOPE_TIMEOUT_MS, A2UI5_MCP_UNIT_TIMEOUT_MS and A2UI5_MCP_BUILD_TIMEOUT_MS override (values in ms).

• Download limits: what is read from GitHub is capped by size as well as by time - 8 MB for a file of the read-only mirror, the docs tree listing and the release list (the largest today is 0.34 MB), 200 MB for the prebuilt backend archive (2.7 MB for 1.146.0). An answer over the cap is refused (when it declares its size) or cut at the cap, and the tool says so by URL; a cached mirror or the previous build stays in place.

• The prebuilt backend is verified before it is unpacked: the framework attaches a backend-<version>.tar.gz.sha256 sidecar beside the archive (releases after 1.148.0), and build_backend compares the download's sha256 against it - a mismatch discards the download and keeps the previous build; a release without the sidecar unpacks as before and the build log says it was unverified. The archive is code run_app boots, so A2UI5_MCP_PREBUILT_URL must be https: - a plain http: mirror off the loopback is refused unless A2UI5_MCP_PREBUILT_INSECURE=1 says the network is trusted (the sidecar is looked for at <url>.sha256 there too).

• UI5 sources are served from the samples-controls checkout's @openui5 packages when that checkout is there, so booting needs no network; without it run_app and interact_app load UI5 from the CDN (and say so at once when it cannot be reached, instead of waiting out the boot timeout). The built theme CSS is not in those packages — with network access it loads from the CDN (styled screenshots); without, apps render unstyled but structurally complete. What the CDN answers is kept in memory across calls (A2UI5_MCP_UI5_CACHE_MB, default 64; 0 switches it off), so a second boot does not download UI5 again. A2UI5_MCP_OFFLINE=1 forces the hermetic behaviour.

• Chromium: A2UI5_MCP_CHROMIUM (or CHROMIUM_BIN, which the linter reads too) names the executable; otherwise the Playwright-managed browser (npx playwright install chromium); otherwise a system chromium. setup_status says which one it found and where it came from. validate_view's warm render gate keeps 2 pages of that browser, which concurrent calls queue on; A2UI5_MCP_RENDER_PAGES (1 to 8) sets it.

• Screenshots: run_app and interact_app write their PNG to <tmp>/abap2ui5-mcp-screenshots/<class>-<stamp>.png (the newest 50 are kept, older ones of that pattern removed) and return the path beside the image — deliberately not into the install directory, which is inside node_modules when you install from npm. The last build's log, which build_log reads after a restart, lives there too. That directory is created 0700 and used only while it is your own (/tmp is shared on Linux); otherwise the image is returned but not saved, and screenshotNotSaved says why. A2UI5_MCP_SCREENSHOT_DIR puts them somewhere you keep.

• scope_of needs an OpenUI5 checkout as well as the corpus: it reads the JSDoc from OPENUI5_SRC, or from ../fork-openui5 beside the samples-controls checkout when that variable is unset.

• If you set this up earlier: the corpus repository was ai-demokit, then abap2UI5-api, and is samples-controls today. Nothing needs changing — an existing checkout is still found under any of the three directory names, and AI_DEMOKIT_HOME is still read alongside SAMPLES_CONTROLS_HOME.

• Real-system deployment stays what it is today: abapGit. This server is the inner dev loop; the real-system half lives in the VS Code extension, whose own MCP server exposes it as run_app_on_system. Both servers are registered in the same editor window, which is why that tool is not called run_app.

Working on this repository

npm ci
npm test

AGENTS.md carries the conventions, CONTRIBUTING.md and RELEASING.md the rest of the workflow.

bench/ is abap2UI5-bench: how often an AI agent's abap2UI5 app passes the template's gates, with and without this server. It is a package of its own and not part of the published server.