Skip to main content

Connectors

A connector groups everything Akmatori uses for one system: the tools the agent can call, the webhook that sends its alerts, the messaging channels where results go, or a remote MCP server. Open Connectors to see yours on the Yours tab, or open the Discover tab to add one.

What a connector can do

Tools

The agent can query and act on the system: metrics, logs, hosts, tickets, databases.

Alerts

The system sends alerts to an Akmatori webhook; each alert can open or join an incident.

Messaging

Channels where alerts and results are posted, and (Slack) where the agent listens.

One connector can do more than one of these. The Zabbix connector holds the Zabbix API access that the agent queries and the Zabbix webhook that opens incidents. A connector page shows a tab for each thing the connector can do: Tools (Server for a custom MCP connector), Channels and Alerts. Messaging and alert connectors also have Formatting, and every connector has Settings.

Built-in connectors

SSH

Run commands on your servers over SSH with key-based authentication, optionally through a jump host. Hosts are read-only (diagnostic commands only) unless you allow write commands; you can also allow single commands and deny others.

Parallel executionAllowed / denied commandsOutput capture

Zabbix

Query the Zabbix API for hosts, problems, items, triggers and history, and receive Zabbix alerts through a webhook. Raw API requests are read-only.

Host lookupMetrics queryProblem historyWebhook

VictoriaMetrics

Query VictoriaMetrics with PromQL and MetricsQL: instant and range queries, series and label values. Raw API requests are read-only; admin and write endpoints are refused.

PromQL/MetricsQLRange queriesLabel discovery

Grafana

Search dashboards, inspect panels, query data sources (including Prometheus and Loki), read alert rules, create silences and annotations, and receive Grafana alerts through a webhook.

Dashboard lookupPrometheus/Loki queriesAlert rulesWebhook

PagerDuty

Look up incidents, alerts, notes, services, on-call and escalation policies, and recent changes. The agent can also acknowledge, resolve, reassign and annotate incidents and send Events API events. Receives PagerDuty webhooks.

Incident lookupOn-call contextIncident actionsWebhook

Kubernetes

Read-only diagnostics: namespaces, pods and their logs, events, deployments, StatefulSets, DaemonSets, Jobs, CronJobs, nodes, services, ConfigMaps and Ingresses.

Pod inspectionLog retrievalEvent analysis

PostgreSQL

Run read-only SQL (SELECT only) and inspect tables, indexes, locks, active queries, replication and database stats.

Read-only SQLSchema inspectionLock analysis

ClickHouse

Run read-only queries, inspect databases and tables, and review the query log, running queries, merges, parts, replication and cluster info.

Read-only SQLQuery logsCluster metadata

NetBox

Look up devices, interfaces, sites, racks, IP addresses, prefixes, VLANs, circuits, virtual machines and tenants. Requests to other NetBox API paths are read-only.

Device inventoryIPAM lookupCircuit context

Catchpoint

Read Catchpoint alerts, tests, performance metrics, errors, monitoring nodes and Internet Weather outages. The agent can also acknowledge alerts and run instant tests.

Synthetic testsInternet outagesAlert actions

Jira

Search issues and read comments, transitions, changelogs, projects and users. Writes (comments, transitions, new and updated issues) stay off until you turn on Allow write operations in the connector settings.

JQL searchIssue workflowComment updates

Gitea

Review pull requests before they merge: read the pull request, its changed files, the diff of one file or all files, any file at a branch or commit, comments and reviews, and search pull requests. Give a pull request link or repo and number; you can limit the agent to a list of repositories. Posting a comment stays off until you allow it in the connector settings. Works with Forgejo too.

Pre-merge reviewPer-file diffsRepository allowlist

Prometheus Alertmanager

Receive Alertmanager alerts through a webhook. A resolved notification marks the matching alert as resolved.

Alerts onlyWebhookResolve handling

Datadog

Receive Datadog monitor alerts through a webhook.

Alerts onlyWebhook

Slack

Post alerts and results, listen to channels for alerts, and answer mentions in threads.

PostListenMentions

Telegram

Post alerts and results to Telegram chats and forum topics through a bot. The agent can also send its own messages to these chats during a run (telegram.send_message).

PostAgent messagesHTML renderingTopics

Discord

Post alerts and results to Discord channels through a bot, without pinging anyone.

Post onlyMarkdown

Zulip

Post alerts and results to Zulip channels (streams) through a bot, under a topic you choose. Mentions are made silent.

Post onlyMarkdownTopics

The built-in tools events, runbooks, memory, incidents and proposals are part of Akmatori itself and need no connector. Declarative HTTP connectors that you define through the API also appear in Discover.

Adding a connector

1

Open Discover

Go to Connectors → Discover (or + Add → Browse connectors) and pick the system.

2

Configure the first part

Enter the API URL and credentials for a tool, create the alert webhook for an alert-only system such as Alertmanager, or enter the bot credentials for Slack, Telegram, Discord or Zulip. When you save, Akmatori creates the connector and opens its page.

json
// Zabbix connector → Tools → settings
{
  "zabbix_url": "https://zabbix.example.com/api_jsonrpc.php",
  "zabbix_token": "••••••••"
}
3

Add the rest on the connector page

Add the webhook on the Alerts tab and copy its URL into the source system, or add channels on the Channels tab. To run a second environment (for example a staging Zabbix), add the same connector again: each environment is its own connector.

4

The agent can use it at once

Every enabled connector tool is available to every incident run. You do not assign tools to skills. Akmatori lists the tools for the agent in the generated connected-tools skill. A cron job can use connector tools only when Allow connector tools is on. To keep a tool from the agent, disable its tool instance.

After an upgrade from a version without connectors, every existing tool, alert source and messaging integration is its own connector. To put a Zabbix webhook next to its Zabbix API access, open the connector's Settings tab and use Merge another Zabbix connector into this one. Webhook URLs and tool names do not change when you merge or move parts.

Messaging connectors

Slack posts alerts and results, listens to channels for alerts, and answers mentions (see Slack). Telegram, Discord and Zulip only post alerts and results through a bot; they do not listen to channels. The Channels tab of a messaging connector holds its channels, its credentials and a Test connection button.

  • Telegram: create a bot with @BotFather, add it to the chat, and use the chat ID (for example -1001234567890 or @channel; add :topic_id for a forum topic).
  • Discord: the bot needs Send Messages and Read Message History; use the channel ID. Messages never ping anyone.
  • Zulip: create a generic bot and enter the Zulip site URL, the bot's email and its API key. Use the channel (stream) name or ID. Messages go to the topic you set on the connector (default Akmatori); a reply goes to the topic of the message it answers. Mentions are made silent, so they do not notify anyone.
  • To send an alert source's notifications to Telegram, Discord or Zulip, choose that channel as its notification channel. The alert is posted there and the investigation result follows as a reply.

Agent messages to Telegram

During a run the agent can send its own message to a Telegram chat with the telegram.send_message tool, for example to warn the on-call team before the final answer. It uses the bot and the chats of the Telegram connector, so nothing is set up twice. Only enabled chats that can post are accepted; without a chat name the default chat is used. The agent can send at most 10 messages per run, as plain text.

SSH: allowed and denied commands

SSH hosts are read-only by default: the agent can run diagnostic commands such as cat, journalctl or systemctl status, and nothing that changes the host. On the Tools tab of an SSH connector (advanced settings) you can add two lists, one pattern per line. They apply to every host of the connector, ad-hoc hosts too.

  • Allowed commands also run on read-only hosts. A pattern must match the whole command: systemctl restart nginx allows only that command, and a * as the last word allows any arguments (journalctl -u nginx *). A * inside a word matches any characters (docker restart app-*).
  • Denied commands are blocked on every host, also on hosts with write commands allowed. A pattern matches the start of a command, after sudo and the program path: reboot also blocks sudo /sbin/reboot now.
  • Each part of a chain (;, &&, ||, |) is checked on its own. An allowed command can never contain $, a backtick, <, >, &, brackets or a backslash, so systemctl restart nginx; reboot and systemctl restart nginx $(reboot) stay blocked.
  • The agent sees both lists in its tool notes, and a blocked command tells it which commands are allowed.
SSH connector → Tools → advanced settings
# Allowed commands (also on read-only hosts)
systemctl restart nginx
journalctl -u nginx *
sudo systemctl reload haproxy

# Denied commands (every host, also with write commands allowed)
reboot
shutdown
rm -rf *
systemctl stop *
Every allowed pattern lets the agent run that command on every host of the connector without approval. Denied commands are a guard rail, not a sandbox: on a host with write commands allowed, a shell (sh -c "...") can still hide a command from the list.

Custom connectors (remote MCP servers)

Connect any remote MCP server by URL. Choose + Add → Add custom connector (MCP), enter a name and the MCP server URL (Streamable HTTP, for example https://mcp.example.com/mcp), choose the authentication (None or one of the modes below), and press Test connection to see the server's tools before you save.

OAuth (sign in)

Akmatori finds the server's authorization server, uses your client ID or registers itself, and sends you to sign in. Tokens refresh in the background. Set the external URL in Settings → General first: the redirect URL is built from it.

Bearer token

Sent as Authorization: Bearer … on every request.

API key header

Any header name with a secret value, for example X-API-Key.

  • The agent calls the server's tools as ext.<name>.<tool> (for example ext.github.search_issues). In a tool name, every character other than a letter, digit, - or _ (dots too) becomes an underscore.
  • Like any other connector, its tools are available to every incident run, and to cron jobs that allow connector tools, as soon as the connector is enabled.
  • On the Server tab you can change the URL and authentication, sign in again, and turn single tools off.
  • Only remote servers are supported. Local (stdio) MCP servers are not: they would run commands inside the gateway container.
Use only MCP servers from developers you trust. Akmatori does not control which tools a server offers and cannot check that they work as intended or that they will not change. The agent can call every tool you leave on.
bash
# Add a remote MCP server through the API (same as the dialog)
curl -X POST https://akmatori.example.com/api/mcp-servers \
  -H "Authorization: Bearer $AKMATORI_TOKEN" -H "Content-Type: application/json" \
  -d '{"name":"GitHub","url":"https://mcp.example.com/mcp",
       "auth_config":{"type":"bearer","token":"..."},
       "disabled_tools":["delete_repo"]}'

# Test it through the gateway (same network policy as real calls)
curl -X POST https://akmatori.example.com/api/mcp-servers/1/test \
  -H "Authorization: Bearer $AKMATORI_TOKEN"

Response formatting

The Formatting tab of a messaging or alert connector shows which formatting rules can reshape the answers for each of its channels and webhooks. Rule for this opens a new rule already scoped to that channel or source. Answers are stored as Markdown and converted for each platform: Slack gets mrkdwn cut to Slack's size limit, Telegram gets HTML split at 4096 characters, Discord gets Markdown split at 2000 characters, and Zulip gets Markdown split at 10,000 characters.

Security

The AI agent never sees credentials. The MCP Gateway reads them at the moment it runs a tool, and only for the tools the run is allowed to use.
The API never returns stored secrets: responses show masked values. Leave a secret out on update to keep the stored value.
Each run gets a tool allowlist from its kind: incident runs get every enabled connector tool, cron jobs get them only when Allow connector tools is on, and proposal chats get no connector tools. The gateway enforces the allowlist for built-in, connector and MCP tools.
Remote MCP and OAuth requests refuse loopback, link-local and cloud-metadata addresses and never follow redirects. Set MCP_BLOCK_PRIVATE_NETWORKS=true to also refuse private networks.
Only the tool schema list of the gateway is reachable through the public proxy; tool calls and gateway admin endpoints stay on the internal network.
The run log of each incident or request shows the agent's tool calls with their arguments and output.