Local Development¶
Everything for running ARC-1 locally against your own SAP — one developer, one laptop.
Already did the Quickstart? This page is the full toolbox: all install methods, .env patterns, MCP client configs, safety profiles, and the cookie extractor for SSO-only on-prem systems.
For multi-user / production deployments see deployment.md.
Install methods¶
Pick how you want to install and update ARC-1. The executable has a few limitations listed below.
npx (zero install)¶
npx arc-1@latest --url https://your-sap-host:44300 \
--user YOUR_USER --password YOUR_PASS \
--client 100
npx downloads on first run and caches. @latest follows the npm latest tag, which can advance across minor or major releases. Best for trying things out.
Slow startup on Windows? Try the single-file executable. Claude Code users can also allow more startup time.
npm install -g (faster startup)¶
npm install -g arc-1
arc1 --url https://your-sap-host:44300 --user YOUR_USER --password YOUR_PASS --client 100
Startup is ~1s faster than npx. Update with npm install -g arc-1@latest.
Docker (local)¶
export ARC1_LOCAL_KEY=$(openssl rand -hex 32)
docker run -d --name arc1 -p 127.0.0.1:8080:8080 \
-e ARC1_API_KEYS="$ARC1_LOCAL_KEY:admin" \
-e SAP_URL=https://your-sap-host:44300 \
-e SAP_USER=YOUR_USER \
-e SAP_PASSWORD=YOUR_PASS \
-e SAP_CLIENT=100 \
ghcr.io/arc-mcp/arc-1:latest
Defaults to HTTP Streamable on :8080. Connect MCP clients to http://localhost:8080/mcp with Authorization: Bearer <ARC1_LOCAL_KEY value>. Full Docker reference → docker.md.
For stdio mode inside Docker (Claude Desktop wraps the docker run in the MCP config), add -e SAP_TRANSPORT=stdio and use docker run -i --rm instead of -d.
Single-file executable (no Node.js, no node_modules)¶
Starting with 1.7.0, each GitHub Release carries one executable per platform plus SHA256SUMS.txt:
| Platform | File |
|---|---|
| Windows x64 | arc-1-<version>-windows-x64.exe |
| Linux x64 / arm64 | arc-1-<version>-linux-x64, arc-1-<version>-linux-arm64 |
| macOS Apple Silicon / Intel | arc-1-<version>-macos-arm64, arc-1-<version>-macos-x64 |
On Linux or macOS, replace the filename below with the one you downloaded, then make it executable and run it from that directory:
chmod +x arc-1-1.7.0-linux-x64
./arc-1-1.7.0-linux-x64 --url https://your-sap-host:44300 --user YOUR_USER --password YOUR_PASS --client 100
The npm package installs several thousand files that antivirus scanners inspect on first start; the executable is one file (#957). It runs the same code on the Node.js runtime it was built with (a Node SEA), so behavior matches the npm package apart from these differences:
- SQLite cache — the native
better-sqlite3addon cannot be embedded, soARC1_CACHE=sqliteuses Node's built-innode:sqlite(prints anExperimentalWarningon stderr; stdout is unaffected). Cache files use the same schema. - Not available — the read-only local UI (
public/ui) is not embedded, andARC1_PLUGINSextensions are untested in the executable. Use npm or Docker for those. - First start can still be slow — Windows Defender and similar tools can scan each new executable (~100 MB); startup time depends on your machine and security policy. The binaries are not publisher-signed; verify downloads with
SHA256SUMS.txt. Corporate policy may block the Windows executable. macOS executables are ad-hoc signed; if Gatekeeper blocks a download, runxattr -d com.apple.quarantine <file>. - Updating — download and verify the new file; there is no built-in auto-update. See the Windows steps below.
Build one yourself with npm run build && npm run build:sea (current platform only; output in dist-sea/), then check it with npm run test:sea-smoke -- dist-sea/<file>.
Windows setup and updates¶
- Download the Windows x64
.exeandSHA256SUMS.txtfrom the same official release. -
Verify in PowerShell from your download folder. Replace
1.7.0with your downloaded version:(Get-FileHash .\arc-1-1.7.0-windows-x64.exe -Algorithm SHA256).Hash Select-String -Path .\SHA256SUMS.txt -SimpleMatch 'arc-1-1.7.0-windows-x64.exe'The hashes must match, ignoring uppercase/lowercase. If they differ, do not run the file. A checksum detects download corruption; it is not a publisher signature.
-
Configure your MCP client. Move and rename the verified file to a stable path, for example
C:\Tools\arc-1\arc-1.exe(create the folder first). In your existing server entry, replace thenpxcommand and remove its-yandarc-1@…arguments:This fragment assumes SAP settings are in
env: keep that existing block. If you use ARC-1 flags such as--url, keep them inargs. Restart the client. For a first setup, use the MCP client configuration below with this command. -
Update by downloading and verifying the next version first. Stop all clients and ARC-1 processes using this file, keep the old executable as a backup, replace it at the same path, then restart. If replacement fails, keep using the old file and retry once it is no longer running.
Automating updates? Check outside MCP startup so downloads do not delay connections. Verify each download, stage it separately, and replace the executable only when it is stopped. Serialize updates when several clients share the file; leave the working copy intact on failure.
git clone (contributing or running from source)¶
For every contribution, check the idea roadmap before starting and before opening or updating the PR. Record useful deferred work, narrow partially completed ideas, and remove completed items from both the overview and details in the same PR. State the affected IDs or "No roadmap impact" in the PR description. This also applies to fixes, documentation, refactors, and research.
For SAP-facing or runtime changes, test through ARC-1 on a real, authorized SAP test system when possible. Use the PR template to record the tested build, SAP release, deployment/authentication route, observed results, and untested scenarios. If you lack access, explain why and what still needs verification; this does not prevent you from contributing. Documentation-only changes can mark live testing as N/A.
For contributor tooling (lint-staged), use Node.js 22 (22.22.1 or later) or 24; the published package's
runtime minimum remains Node.js 22.19.
git clone https://github.com/arc-mcp/arc-1.git
cd arc-1
npm ci
cp .env.example .env # then edit for your SAP system
# Pick one:
npm run dev # stdio from source; restart after edits
npm run dev:http # builds + runs HTTP streamable on 0.0.0.0:8080
npm run build && npm start # production-style: compile to dist/, run from there
npm ci fails looking for Python or Visual Studio. An npm bug
can run node-gyp rebuild for better-sqlite3 13 when installing from a lockfile, despite its
bundled prebuilt binaries. This commonly affects Windows machines without the build toolchain.
For this specific failure, skip install scripts for this one install:
npm ci --ignore-scripts
npm run build
node -e "const Database = require('better-sqlite3'); const db = new Database(':memory:'); console.log(db.prepare('SELECT 1 AS ok').get()); db.close()"
The SQLite check should print { ok: 1 }. If it fails, the workaround is insufficient;
resolve the native-module error before continuing. --ignore-scripts skips all install
scripts, including Git-hook setup: run npm run prepare before committing. Keep this flag
on the affected install command rather than setting it globally. See #1038
for the reported Windows environment and results.
| Script | What it does | When to use it |
|---|---|---|
npm run dev |
tsx src/index.ts — runs from source over stdio. No build step; does not watch files or restart automatically. |
Iterating on stdio-mode code; pairing with node debugger. |
npm run dev:http |
npm run build && tsx src/index.ts --transport http-streamable |
Iterating on HTTP-mode code; testing OAuth/XSUAA flows; pointing remote MCP clients at your laptop. |
npm run build |
tsc + copies AFF schemas to dist/. |
Producing the dist/index.js you ship in Docker or invoke as node dist/index.js. |
npm start |
node dist/index.js (assumes you ran npm run build first). |
Production-equivalent local run; useful when you want to test the same artifact CI publishes. |
All four read .env from the repo root (their CWD). CLI flags and shell env vars override .env — see Where do my config values come from? below.
Read-only local UI¶
For Claude Desktop, Claude Code, Cursor, or any stdio client, add ARC1_UI=local (or pass --ui) to start a loopback browser console next to the MCP subprocess:
npx arc-1@latest --ui --ui-open true \
--url https://your-sap-host:44300 \
--user YOUR_USER --password YOUR_PASS \
--client 100
The sidecar binds to 127.0.0.1:8711 by default and serves http://127.0.0.1:8711/ui/. It is read-only: sanitized config, safety/auth state, feature status, cache metadata, and recent sanitized audit events. It never exposes cached ABAP source bodies or mutation controls.
Using a .env file¶
Copy .env.example to .env and uncomment what you need. Priority is CLI > env > .env > defaults.
Minimal .env for basic auth:
SAP_URL=https://your-sap-host:44300
SAP_USER=YOUR_USER
SAP_PASSWORD=YOUR_PASS
SAP_CLIENT=100
SAP_LANGUAGE=EN
The .env file loads automatically for npm run dev, npm start, and the arc1 CLI. npx also reads .env if one exists in its launch directory. Docker requires explicit --env-file/-e settings or a mounted file.
Full grouped template with every option: see .env.example. The file is grouped into Layer B (ARC-1 → SAP) and Layer A (MCP Client → ARC-1) blocks with fail-fast rules documented inline.
Where do my config values come from?¶
ARC-1 resolves every config field from four sources, in order: CLI flag > process.env > .env file (in CWD) > built-in default. The .env file is loaded by dotenv from the running process's CWD and never overrides values already in process.env — it only fills in missing keys.
What process.env contains depends on how you launched ARC-1:
npm run dev/npm startfrom a shell — your shell exports + the repo's.env. Shell wins over.env.npx arc-1from an MCP client (stdio) — theenvblock in yourmcp.json/claude_desktop_config.jsonbecomes the subprocess's environment. Dotenv also checks the subprocess working directory; do not assume which directory your client chooses.- Remote HTTP — MCP client connects via
"url"— only the server's startup environment matters. Puttingenv:in mcp.json next to aurl:does nothing. - Docker / BTP CF —
-eflags,--env-file,cf set-env, manifest properties, andVCAP_SERVICES. No.envinside containers.
Full per-mode table + debugging tips: Configuration Precedence.
MCP client configuration¶
All MCP clients that speak stdio work the same way — they spawn npx arc-1 as a subprocess and talk JSON-RPC over stdin/stdout. The env block is where credentials and safety flags go, for example SAP_ALLOW_WRITES, SAP_ALLOW_FREE_SQL, and SAP_ALLOWED_PACKAGES.
Claude Desktop¶
~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows).
{
"mcpServers": {
"sap": {
"command": "npx",
"args": ["-y", "arc-1@latest"],
"env": {
"SAP_URL": "https://your-sap-host:44300",
"SAP_USER": "YOUR_USER",
"SAP_PASSWORD": "YOUR_PASS",
"SAP_CLIENT": "100"
}
}
}
}
To keep the same local setup but allow SQL + named table preview while staying read-only, add "SAP_ALLOW_DATA_PREVIEW": "true", "SAP_ALLOW_FREE_SQL": "true" inside that same env block.
Claude Code¶
Project-scoped: create .mcp.json in the repo root with the same shape as above. User-scoped: ~/.claude.json with a mcpServers block.
Claude Code startup timeout¶
For slow Windows starts, prefer the executable or install with npm ahead of time. The npm/npx package is unchanged; a new @latest release can require another download/install on first start.
If Claude Code still times out during startup, allow up to three minutes by running this in PowerShell before launching Claude Code:
This applies to Claude Code in that terminal session. Put it in Claude Code's environment, not ARC-1's .env or MCP server env block. It changes the startup deadline, not tool-call timeouts or startup speed. Claude Code currently documents a 30-second default; see its environment-variable reference.
Cursor¶
Cursor Settings → MCP — same JSON shape as Claude Desktop.
VS Code / GitHub Copilot¶
For local stdio mode in .vscode/mcp.json, use servers (Claude Desktop uses mcpServers):
{
"servers": {
"sap": {
"type": "stdio",
"command": "npx",
"args": ["-y", "arc-1@latest"],
"env": {
"SAP_URL": "https://your-sap-host:44300",
"SAP_USER": "YOUR_USER",
"SAP_PASSWORD": "YOUR_PASS",
"SAP_CLIENT": "100"
}
}
}
}
For a long-running local server, first generate ARC1_LOCAL_KEY as in the Docker example, then run ARC-1 as an HTTP Streamable server:
npx arc-1@latest --url https://host:44300 --user dev --password secret \
--client 100 --api-keys "$ARC1_LOCAL_KEY:admin" \
--transport http-streamable --http-addr 127.0.0.1:3000
Then in VS Code MCP settings:
{
"servers": {
"sap": { "type": "http", "url": "http://localhost:3000/mcp", "headers": { "Authorization": "Bearer <ARC1_LOCAL_KEY value>" } }
}
}
If you want read-only SQL access in VS Code / Copilot, change the command that starts ARC-1, not the MCP JSON. The JSON only tells VS Code where the already-running server lives:
SAP_ALLOW_DATA_PREVIEW=true SAP_ALLOW_FREE_SQL=true \
npx arc-1@latest --url https://host:44300 --user dev --password secret \
--client 100 --api-keys "$ARC1_LOCAL_KEY:admin" \
--transport http-streamable --http-addr 127.0.0.1:3000
For a local loop, bind to
127.0.0.1not0.0.0.0— stops other machines on the network from hitting your instance. HTTP authentication is required even on loopback; for setup see api-key-setup.md.
Gemini CLI / Goose / OpenCode / other stdio clients¶
Same pattern: spawn npx -y arc-1@latest with the same env block. Use each client’s configuration syntax; the subprocess command is the same.
Pointing an MCP client at a locally-built instance¶
When you're iterating on ARC-1 itself (or just want to skip the npx download), point the client at your local clone instead.
Stdio against compiled dist/ — closest to what the published package does. Run npm run build first.
{
"mcpServers": {
"sap-local": {
"command": "node",
"args": ["/absolute/path/to/arc-1/dist/index.js"],
"env": {
"SAP_URL": "https://your-sap-host:44300",
"SAP_USER": "dev",
"SAP_PASSWORD": "secret",
"SAP_CLIENT": "100"
}
}
}
}
Dotenv reads .env from the subprocess CWD, which may differ from the directory containing dist/index.js. Pass settings via env, or set the working directory if your client supports it.
Stdio against tsx (no build step) — handy while iterating on source.
{
"mcpServers": {
"sap-local": {
"command": "npx",
"args": ["tsx", "/absolute/path/to/arc-1/src/index.ts"],
"env": { "SAP_URL": "…", "SAP_USER": "…", "SAP_PASSWORD": "…", "SAP_CLIENT": "100" }
}
}
}
HTTP against a long-running npm run dev:http — useful for testing OAuth/XSUAA or sharing one instance across multiple clients. Start the server in a terminal:
cd /path/to/arc-1
ARC1_API_KEYS="$ARC1_LOCAL_KEY:admin" npm run dev:http -- --http-addr 127.0.0.1:3000 # reads .env from repo root
Then point any MCP client at the URL:
{
"servers": {
"sap-local": { "type": "http", "url": "http://localhost:3000/mcp", "headers": { "Authorization": "Bearer <ARC1_LOCAL_KEY value>" } }
}
}
For local HTTP loops, bind to
127.0.0.1(not0.0.0.0) so other machines on your network can't hit the instance. HTTP authentication is required even on loopback — see api-key-setup.md.
In HTTP mode, the env block in mcp.json does nothing — the server already has its own environment from when you launched npm run dev:http. Change config by editing .env (or shell-exporting) and restarting the server. Full mode-by-mode breakdown: Configuration Precedence.
Safety flags¶
ARC-1 ships read-only. For local development, enable only what you need via positive-opt-in flags.
What's blocked by default¶
| Capability | Default | Opt-in flag |
|---|---|---|
| Object writes | off | SAP_ALLOW_WRITES=true |
| Named table preview | off | SAP_ALLOW_DATA_PREVIEW=true |
| Freestyle SQL | off | SAP_ALLOW_FREE_SQL=true |
| Transport mutations | off | SAP_ALLOW_TRANSPORT_WRITES=true (also needs SAP_ALLOW_WRITES=true) |
| Git mutations | off | SAP_ALLOW_GIT_WRITES=true (also needs SAP_ALLOW_WRITES=true) |
| Writes to any package | $TMP only |
SAP_ALLOWED_PACKAGES='$TMP,Z*' (or * for any) |
Transport reads (list / get / check / history) and Git reads (list_repos / whoami / branches / ...) work without any opt-in flag — only mutations are gated. The Git opt-in currently enables gated abapGit workflows and SAP-side Git egress; every gCTS mutation is quarantined before HTTP. Accepted abapGit mutations without an authoritative postcondition return error/incomplete, so inspect state before retrying.
Common local starting points¶
- Just explore (default): no config needed.
- Table preview + SQL:
SAP_ALLOW_DATA_PREVIEW=true SAP_ALLOW_FREE_SQL=true. - Developer:
SAP_ALLOW_WRITES=true SAP_ALLOWED_PACKAGES='$TMP,Z*'— optionally addSAP_ALLOW_TRANSPORT_WRITES=trueorSAP_ALLOW_GIT_WRITES=true. - Full local dev: all 5
SAP_ALLOW_*=true+SAP_ALLOWED_PACKAGES='*'(quote the*in shell so it isn't expanded to filenames).
For fine-grained blocking even after the above flags are set, use SAP_DENY_ACTIONS — e.g. SAP_DENY_ACTIONS="SAPWrite.delete,SAPManage.flp_*". See authorization.md.
Full model and per-user scope handling: authorization.md. Production hardening: security-guide.md.
SSO-only on-prem: cookie extractor¶
⚠️ Developer-only escape hatch. Single user, short-lived session, never for deployed / shared instances. The script refuses to run if
SAP_PP_ENABLED=true.
Some corporate on-prem SAP systems return an HTML login page on /sap/bc/adt/ instead of accepting Basic Auth — typically when SAML2 / SPNEGO / X.509 / Kerberos SSO is enforced. You're expected to authenticate through a browser.
For a single-developer local loop, the included extractor scrapes your existing SAP session cookies from Chrome and writes them to a file ARC-1 can reuse:
What it does:
- Launches Chrome with remote-debugging enabled (CDP).
- You complete your normal SSO login in the browser window (IdP redirect, MFA, whatever your corp flow is).
- Return to the terminal and press Enter. The script then reads the SAP session cookies (
SAP_SESSIONID_*,MYSAPSSO2,sap-usercontext) out of Chrome. - Writes them to
cookies.txtwith mode0600.
Then point ARC-1 at the cookie file:
Startup log:
When to use it¶
- ✅ SSO-only on-prem SAP, solo developer loop.
- ✅ Your IdP enforces MFA / X.509 client certs that can't be scripted.
- ❌ Multi-user / deployed ARC-1 — cookies are one user's session.
- ❌ BTP ABAP Environment — use service-key OAuth.
- ❌ Combined with
SAP_PP_ENABLED=true— the extractor refuses, and ARC-1 fails at startup unless the explicit escape-hatchSAP_PP_ALLOW_SHARED_COOKIES=trueis set.
Limitations¶
- Cookies expire (usually minutes to hours). Re-run the script to refresh.
- SAP sees whichever user you logged in as, not "the MCP caller" — fine for solo dev, wrong for a shared service.
- No refresh token — you get whatever session the browser has.
For per-user SAP identity with a deployed ARC-1, use a per-user BTP Destination: Cloud Connector Principal Propagation for on-premise SAP, or OAuth2UserTokenExchange for BTP ABAP Environment. See deployment.md.
What you get at startup¶
Every ARC-1 startup prints a one-line auth summary on stderr:
INFO: auth: MCP=[none] SAP=basic (shared)
INFO: auth: MCP=[api-key] SAP=cookie (shared)
INFO: auth: MCP=[oidc] SAP=pp (per-user) [disable-saml=on]
This line tells you which Layer A / Layer B methods are active. If it disagrees with what you thought you configured, that's the first place to look.
Full auth reference (all methods, combinations, coexistence rules): enterprise-auth.md.
CLI usage (outside MCP)¶
Sometimes you just want to shell-test an ADT endpoint without running the full MCP server:
# Works off .env
npm run cli -- search ZCL_CUSTOMER
npm run cli -- source clas ZCL_CUSTOMER
# Verbose (shows every HTTP request)
SAP_VERBOSE=true npm run cli -- search ZCL_CUSTOMER
Full CLI reference → cli-guide.md.
Next¶
- Deploy for a team → deployment.md
- All flags → configuration-reference.md
- Where config values come from → configuration-precedence.md
- Auth internals and combinations → enterprise-auth.md
- Update ARC-1 → updating.md