Connect your AI (MCP)
Keep working with the AI agent of your choice. Through the Saka MCP, the agent only uses the tools we provide, not a root shell that can do anything.
1. Create a token
- Open Pengaturan (Settings) in the panel, on the Sambungkan AI Anda (MCP) & token API (Connect your AI (MCP) & API tokens) card.
- Give the token a name, for example "Claude on my laptop", then press Buat token (Create token).
- Choose the token access: Can make changes, or View only when the AI only needs to watch (tools that change things are hidden and refused, and so are changes through the API). Or Selected permissions for a token that may only change certain areas (sites, servers, databases, kubernetes, networking, monitoring, delete). Through MCP the AI only sees and can call the tools in those areas; other tools are hidden and refused, and tools that delete need the delete permission. See Team and roles.
- Copy the token (it starts with
saka_) now. It is shown only once.
The same token works for MCP and the REST API. The panel shows when it was last used, and a token can be revoked at any time; any AI using it loses access immediately.
2. Connect
Load only the tools you use
Without a choice, the AI agent loads all 148 Saka tools (about 25k tokens in every conversation; some clients such as Cursor also limit the number of tools). Add ?alat= to the MCP address to load only some groups, e.g. https://app.saka.work/mcp?alat=situs,pantau. Core tools (server list & status, tasks, alerts, action history, feature requests) are always included. Groups: server (servers, Shield, cloud accounts), situs (sites, files, WordPress, catalog, S3 backups), pantau (monitoring), aplikasi (OpenClaw, Hermes), database, jaringan (load balancers, tunnels, VPN), k8s, armada (bulk deploy), pindah (platform moves). In Settings, when you create a token, tick the groups and the connect commands update. Tools outside the chosen groups are refused with a message naming their group.
Claude Code
claude mcp add --transport http sakapanel https://app.saka.work/mcp --header "Authorization: Bearer <token>"Other MCP clients
- MCP server address:
https://app.saka.work/mcp - Transport: Streamable HTTP
- Header:
Authorization: Bearer <token>
OpenClaw or Hermes (manage servers by Telegram chat)
Installed through Saka: open its app page, the Manage servers by chat card, choose the access, then press Connect to Saka. Saka creates a dedicated token, stores it on the app's server (not in the task history), and restarts the assistant. Press Disconnect to revoke it; deleting the app or releasing the server revokes it too. With Can make changes, every change the assistant makes is reported to your Telegram.
Installed yourself: create a token, then
openclaw mcp add saka --url https://app.saka.work/mcp --transport streamable-http --header "Authorization=Bearer <token>"mcp_servers:
saka:
url: https://app.saka.work/mcp
headers:
Authorization: "Bearer <token>"Anyone who can chat with that assistant can use its access to Saka. Make sure only you are approved in Chat access.
claude.ai, Claude Desktop, and ChatGPT (no token to copy)
- claude.ai or Claude Desktop: Settings, Connectors, Add custom connector. ChatGPT: Settings, Apps & Connectors, create a connector (developer mode).
- Enter
https://app.saka.work/mcpas the address, then press Connect. - A Saka page opens: sign in if asked, choose View only or Can make changes, then Allow. You're sent back to the AI app and the connector is active.
Each connector shows up as a token in Settings, API tokens (e.g. "Konektor Claude") and can be revoked any time. You're also notified on Telegram when a new connector connects.
Once connected, try asking: "How are my servers doing?" The agent will call server_daftar.
Available MCP tools
| Tool | What it does |
|---|---|
server_daftar | Lists servers with their latest status (connected, CPU, RAM, disk, apps). |
server_status | Live status of one server, requested from the agent right then. |
server_tambah | Creates a one-line install command for a new VPS; you run it on the server yourself. |
metrik_riwayat | CPU, RAM, disk, and load history (stored on that server, 30 days). |
server_bandwidth, server_kuota_bandwidth | Bandwidth in and out per day, per month, and per site, and the monthly quota for 80% and 100% alerts. See Bandwidth usage. |
resep_daftar | Catalog of ready-to-install apps and the fields they need. |
aplikasi_pasang | Installs an app from the catalog on a server (non-root container, local-only port). |
aplikasi_log | Recent app logs, read directly from the server (not stored by Saka). |
aplikasi_mulai_ulang | Restarts an app; data is not changed. |
aplikasi_hapus_siapkan | Step 1 of 2 for removing an app: a preview and a confirmation token (10 minutes). |
aplikasi_hapus | Step 2 of 2: removes the app with the token, only after you have approved the preview. |
aplikasi_akses_daftar | Access requests (pairing) waiting for approval, for example a new person messaging the OpenClaw bot. |
aplikasi_akses_setujui | Approves one access code, only after you recognize the sender. |
server_perbaiki | Readiness and server security fixes (keamanan-cek read-only, ssh-kunci, pembaruan-otomatis): swap, firewall-cek (port list, read-only), firewall, plus web-cek (the plan, read-only) and web (set up ports 80/443 when CasaOS, nginx, or Apache use them, with your approval; see Web ports). |
database_daftar | Lists database clusters with their state and members. |
database_buat | Creates a multi-location database cluster (PostgreSQL or MariaDB, 2 data + 1 witness). |
database_status | Cluster state, members, replication lag, and connection string. No password. |
database_sambungkan_aplikasi | Installs the Saka proxy on an app server so the app fails over automatically when a server goes down. |
panduan_deploy | Steps and rules to take an app built with OpenCode/Claude Code/Cursor from the user's computer online on a Saka server, from an empty VM to an HTTPS address. The AI calls it before starting. See From an empty VM to online. |
panduan_ha | Steps and rules to keep an app + database running when one VM or region dies. The AI calls it before starting. |
database_pindahkan_lama | Copies an existing database (PostgreSQL, or MySQL 8/MariaDB into a MariaDB cluster) into a cluster, then compares row counts. |
lb_daftar | Lists load balancers: balancers, targets, domains, state. |
lb_buat | Creates a load balancer with health checks and automatic HTTPS. |
lb_status | Health of each target server as seen from each balancer, and the HTTPS certificate for each domain. |
situs_daftar | Lists sites (WordPress, PHP, static) across all servers, with their domains and state. |
situs_buat | Creates a WordPress, PHP, or static site with automatic HTTPS, a local or cluster database, or moves one from cPanel (from a backup file, or by signing in to the old cPanel; the cPanel password is only passed to your server). The agent must confirm the plan with you first. |
situs_status | Site state: HTTPS and DNS for each domain, size, list of backups, SFTP account. No passwords. |
situs_cadangkan | Makes a site backup now (files + database), stored on the server. |
situs_alamat | Gives an existing site an automatic https://<name>-app.hack.id address (HTTPS included, no DNS setup). New sites just leave the domain empty. |
situs_unggah | Uploads images, fonts, PDFs, or a zip (extracted in place) to a site, up to 10 MB per call. With panduan_deploy, an agent can put a static site folder online without Git, photos included. |
situs_uji_pulih | Tests that the latest backup really restores: opened in a temporary container with no internet access, database imported, site loaded once. The live site is not changed. See Backup restore test. |
situs_atur_php | Changes a site's PHP settings (like cPanel's MultiPHP INI Editor): upload limit, memory, execution time, max_input_vars, time zone, show errors. The site restarts for a few seconds. |
situs_pindah_salin | First step of moving to another server: copies the site straight to the target server (optionally with another PHP version and a WordPress update with automatic rollback). The old site keeps running. |
situs_ubah_siapkan | Step 1 of 2 for actions that replace a site's content: pulihkan (restore), hapus (delete), pindah_sekarang (move now: the old site shows a maintenance page briefly and the site goes live on the new server), pindah_selesai (finish: delete on the old server once DNS points to the new one), or pindah_batal (cancel the move). Returns a preview and a confirmation token (10 minutes). |
situs_ubah | Step 2 of 2: runs that action with the token, only after you have approved the preview. |
fitur_usulkan | Sends a feature request to the Saka team when you need something Saka cannot do yet. The agent must show you the draft title and description, and send it only after you agree. |
katalog_daftar, situs_perbarui_aplikasi | Ready-made apps: list the one-click apps (install with situs_buat type katalog) and update to the latest version with a backup first. See Ready-made apps. |
situs_deploy, situs_rollback, situs_env, situs_email_uji | Apps from Git: deploy the latest commit (zero downtime), roll back to a previous release, set environment variables (random secrets are generated on the server without being shown), and send a test email with the app's SMTP settings. See Apps from Git. |
tunnel_daftar, tunnel_buat, tunnel_hapus | Free public addresses <name>-app.hack.id for a site or web app on a server without a public IP (home server). Deleting needs the name as confirmation. See Home server & tunnel. |
vpn | Team WireGuard VPN: list, set up on a server, add a device (split or full; the configuration with its secret key is given only once), revoke, remove. |
riwayat_aksi | Account activity log: who changed what, AI agents through MCP included. Read-only. |
situs_periksa_aplikasi | Check a Git app: secrets in the repo, missing env variables, paths open without login. Read-only. |
cloud_akun_daftar, cloud_pilihan, server_buat, cloud_server_daftar | Create new servers from a DigitalOcean, Vultr, or IDCloudHost account you have connected: see locations and sizes, create a server (Saka installs itself), then follow it until it connects. Account tokens are only added in the panel, never over chat. See Create servers from a cloud account. |
server_ganti_nama, server_port, server_log | Rename a server's display name, list open ports with their programs, and read the tail of the system log. Read directly from the server, no shell. |
aplikasi_telegram | Add a Telegram bot to an installed Hermes or OpenClaw, or change the bot token. App data is kept; if anything fails, the app is put back as it was. |
aplikasi_pindah, dokploy_pindah | Move to Saka: OpenClaw and Hermes to another VPS, and apps and sites from Dokploy, Coolify, aaPanel, and CyberPanel (platform), including standalone databases into a cluster (klaster). See Moving to Saka. |
situs_git_database | One-click database for a Git app (PostgreSQL, MySQL, MariaDB, MongoDB, Redis). See Databases for Git apps. |
database_cadangan_status, database_pulihkan, database_salinan_tabel, database_salinan_ambil, database_salinan_hapus | Database backups to your S3: see the status and the time range you can restore, restore to a point in time as a separate copy, compare its tables with the database now, bring back the tables you need (to a separate place, or replacing them with your approval), and delete the copy. See Backups & restore. |
situs_proses, situs_cron_jalankan, situs_log, situs_pratinjau, situs_lokasi_kedua | More for Git apps: workers and cron, run a cron job now, read logs (app, build, worker, web), pull request previews with their own database per PR, and a copy in a second location so the app keeps running when one server or region goes down. |
armada_daftar, armada_buat, armada_status, armada_deploy, armada_env | The same Git app on many servers: group it, release in stages (one canary server first, then waves, stopping by itself if many fail), and change shared variables. See Bulk deploy. |
cadangan_s3_masa_simpan | Changes how many days site backups are kept in S3 (1 to 3650) so the bucket does not pile up; the newest backup of each site is always kept. The agent must confirm with you first when shortening. See Backup to S3. |
pantau_luar, laporan_bulanan | Monitor URLs/APIs outside Saka (list, add with a keyword or HTTP status, remove), and the account's monthly report (availability, bandwidth, security). See Monitor addresses outside Saka. |
pantau_daftar | Internet monitoring results for every site, tunnel, and load balancer (reachable or not, since when, 30-day uptime), plus your public status pages. |
server_cloud, server_snapshot, server_ubah_ukuran, server_hapus_cloud | The server's VM in your cloud account: see provider, size, and snapshots; take a snapshot; resize (with your approval); delete the VM at the provider and in Saka (confirmed with the server name). See Snapshot, resize, and delete. |
k8s_daftar, k8s_status, k8s_buat, k8s_adopsi, k8s_aplikasi, k8s_terapkan, k8s_hapus, k8s_log, k8s_cadangan | Kubernetes: list clusters and their state (nodes, workloads, failing pods), create a cluster from empty servers or adopt an existing k3s cluster, deploy apps with HTTPS without YAML, apply YAML, delete resources, read logs, and take etcd snapshots. Actions that change the cluster need your approval. |
email_dns_cek | Check a sender domain's email DNS (SPF, DKIM, DMARC, MX) and suggest the records the user adds at their DNS provider. Read-only. See Outgoing email. |
berkas_daftar, berkas_baca, berkas_ubah | Site file manager: list a folder and the trash, read text files, and write, create folders, rename, delete to trash, restore, and unzip (with your approval). |
wp_keamanan, wp_kuatkan, wp_perbarui, wp_nonaktifkan_plugin, wp_pasang_ulang, wp_staging | WordPress: security score, known vulnerabilities (CVE) for installed versions, a hardening list fixable row by row, security check (outdated versions, plugins closed for security issues, files changed by intruders), reinstall original files, updates with backup & automatic rollback or on staging first, plugin deactivation, and staging create, push, and delete. Actions that change the live site need your approval. See WordPress security & staging. |
perisai_status, perisai_atur, perisai_lepas | Saka Shield: see blocked IPs and why and WAF findings; turn it on/off, change exceptions, rate limit, country blocks, and the OWASP WAF mode (with your approval); lift a block. See Saka Shield. |
tugas_pantau_daftar, tugas_pantau_buat, tugas_pantau_hapus | Scheduled job monitoring via ping addresses (heartbeat): list with status, create a monitor (your AI agent can also add the curl command to your crontab), and delete. See Scheduled job monitoring. |
tugas_status | State of long-running tasks (for example an app install). |
peringatan_aktif | Problems happening right now across all servers (disconnected, disk full, app stopped). |
database_tambah_node | Add a new data server to a running PostgreSQL cluster as a replica (S3 backup configuration is copied too). |
database_jadikan_utama | Planned switchover of the primary role to another replica through Patroni. |
database_diagnosa | Read-only evidence of a cluster node's state and a guess at the problem. |
database_ubah_mode | Switch the replication mode between synchronous and asynchronous. |
database_akses | Allow or revoke IP addresses that may reach the database port. |
jaringan_periksa | Pre-check servers and UDP/TCP between servers before creating a cluster, with round-trip time. |
dns_failover_daftar | List one-click DNS switches and which record is currently active. |
dns_failover_alihkan | Switch the A record to the standby server, or back to the primary. |
dns_failover_gladi | Failover drill: really switch, measure resolver convergence, check the standby, then switch back. |
s3_tujuan_uji | Test a saved S3 destination before using it for backups. |
pantau_uji_alarm | Send a test message through the alarm channel (Telegram or email) to confirm alerts arrive. |
Egress monitor. The aplikasi_egress tool (and GET /api/v1/server/{id}/aplikasi/{nama}/egress) lists the internet addresses an app container (OpenClaw, Hermes and so on) connects to: IP, port, reverse name when available, and when each was first and last seen. It only observes and blocks nothing, so it cannot make a bot stop working. It reads the kernel connection table and changes no firewall rules. Destinations not seen for 30 days are dropped. Connections to private and loopback addresses are not recorded.
Egress limit. Based on that monitoring you can restrict where an app may send data. The aplikasi_egress_batas tool (GET /api/v1/server/{id}/aplikasi/{nama}/egress/batas) reads the mode, allow list, observed destinations and blocked ones. The aplikasi_egress_atur tool (PUT on the same path with {mode, izin, uji_menit}) changes them; it is on the Telegram-approval list when you turn that on. Three modes per app: pantau (monitor, the default, no firewall rules), simulasi (the app keeps working, Saka only records what WOULD be blocked) and batasi (internet connections outside the allow list are refused immediately). The allow list holds domains (full names, no wildcards, resolved to IPs every minute with old IPs kept 10 more minutes), IP addresses, or IP blocks (at least /8), with optional port and protocol. Private addresses, replies to incoming connections and the server DNS are always allowed; IPv6 is not covered. Saka checks from inside the container network that the rules really apply and cancels restrict mode by itself if they do not. With uji_menit (10 recommended) restrict reverts to simulate by itself unless confirmed again. Rules live in the DOCKER-USER chain and are restored automatically if the firewall is flushed or the container is recreated (within 5 seconds). Safe flow: simulate for a few days, complete the list from what was recorded, then restrict with a 10 minute trial. Domain names: Saka passively reads the DNS replies (UDP/53) the container receives and shows the name the app asked for (the domain field) next to the IP in the seen, allowed and blocked lists; the Allow domain button adds it to the allow list. DNS over TCP, DoH or DoT is not visible, so those entries stay IP only. Ready-made presets (Telegram, WhatsApp, Discord, Slack, OpenAI, Anthropic, Google Gemini, OpenRouter, DeepSeek, Groq, Mistral, xAI, Together, Perplexity, GitHub, npm and pip): list them with aplikasi_egress_preset (GET /api/v1/egress/preset) and use them through the preset argument (ids) of aplikasi_egress_atur or in the PUT body. Wildcards are not supported, so WhatsApp media (rotating CDN server names) needs names added from what simulate mode records. In the panel: the Restrict the addresses it may contact card under the internet addresses card, with preset buttons.
API route list. GET /api/v1/rute (and the MCP tool rute_api) returns every REST route in a machine-readable form: method, path, path parameters, group, and access kind (public, login or admin; admin routes only show to admins). Filter with ?kelompok=situs or ?q=egress. The list is generated from the code and guarded by a test so it cannot fall behind. It does not include request body schemas yet; for the body shape, see the matching MCP tool through tools/list.
Every error includes kode (code), arti (meaning), and langkah (next step), so the agent can fix things itself or explain to you what to do. MCP tool descriptions and responses are in Indonesian; your AI agent can still reply to you in your own language.
Security principles
- No root, no terminal. The agent can only call the tools above. The agent on your server also only knows a fixed set of commands, not arbitrary shell commands.
- Deleting always takes two steps.
aplikasi_hapus_siapkanreturns a preview and a token. The agent must show you the preview and ask for approval; only then isaplikasi_hapuscalled with the token anddisetujui_pengguna=true. Restoring, deleting, and moving a site follow the same pattern (situs_ubah_siapkan, thensitus_ubah). The token is single-use and valid for 10 minutes. Deleting a database cluster or load balancer, or releasing a server, is not available through MCP; that can only be done from the panel or the API with name confirmation. - The firewall must be checked first. Before turning on the firewall, the agent must call
firewall-cekand show you the ports currently in use. SSH always stays open. - Passwords are not handed over. Database, SFTP, and WordPress admin passwords are never sent through MCP; you view them yourself in the panel. AI keys and bot tokens you provide are only passed on to your server and are not stored by Saka.
- Feature requests only with your approval.
fitur_usulkanis sent only after you approve the draft request. - Access approval stays with you.
aplikasi_akses_setujuimay only be called after you recognize the sender. - Logged. Every task that changes a server is recorded in the server history, marked "mcp".
AI directly on the server vs AI through Saka
Giving an AI agent root SSH access is fast, but one wrong command, or one web page that tricks the agent, can delete your data and expose your server.
| AI directly on the server (root SSH) | AI through Saka (MCP) | |
|---|---|---|
| Access | Root shell: can run anything | Only the tools provided: status, install apps, sites, create databases, load balancers, fixes |
| Keys & passwords | SSH keys, database passwords, and AI keys are readable by the agent | Never handed to the agent; you view them yourself in the panel |
| Dangerous actions | Run immediately, with no preview | Removing an app always takes two steps; the firewall must show the ports in use first |
| Tricked through chat or the web | Injected commands run with root privileges | At most it calls limited tools that still need your approval |
| Audit trail | Not recorded anywhere | Every task on the server is recorded in the history with its result |
| Server dies suddenly | The agent dies with it; nobody tells you | Saka monitors from outside and sends a Telegram notification |
AI costs are paid directly to your AI provider. Saka does not charge for MCP.