# Give AI agents the ability to monitor your infrastructure

Levirge DevOps lets AI agents such as Claude check databases, logs, Docker
containers, model servers and message buses behind your firewalls, over MCP. A
small client, devops-client, connects outward, so nothing needs an inbound port
or VPN. Agents are read-only unless you grant a writer role, and every agent
action is audited.

**Status: private testing.** devops-client has no public download yet.
[Join the waitlist](/devops#cta).

## What an agent can check

- **Databases.** `get_sql_schema` and `run_query` on
  [PostgreSQL](/devops/plugins/postgresql.md), [MySQL / MariaDB](/devops/plugins/mysql.md)
  and [SQLite](/devops/plugins/sqlite.md); `run_mongo_query` on
  [MongoDB](/devops/plugins/mongodb.md); `influx_query` on
  [InfluxDB](/devops/plugins/influxdb.md); `redis_command` on
  [Redis / Valkey](/devops/plugins/redis.md).
- **Logs.** `log_tail`, `log_search` and `log_read` on
  [log files and directories](/devops/plugins/logs.md).
- **Containers.** `docker_list_containers`, `docker_get_logs` and
  `docker_get_stats` on [Docker](/devops/plugins/docker.md).
- **Message bus.** Streams, consumer lag and server health on
  [message bus servers](/devops/plugins/message-bus.md).
- **Model servers.** `llama_read` on [llama.cpp](/devops/plugins/llama.md):
  loaded model, slot use and token rates.

`find_connection` turns a name like "checkout" into the connection to use, and
suggests close matches when there's no exact one.

## Ready-made checks

Built-in prompts start an agent on a common job:

- `container_health_check`: resource use, unhealthy or stopped containers.
- `find_slow_queries`: slow queries in the logs above a threshold.
- `explain_query`: a query plan with `EXPLAIN ANALYZE`.
- `redis_memory_analysis`: keyspace spread, large keys, eviction, fragmentation.
- A message bus audit of streams and consumer lag.

## An investigation, step by step

Illustrative, with made-up names and results:

| Step | The agent calls | What it learns |
| --- | --- | --- |
| 1 | `find_connection` "checkout" | The checkout database, its log directory and the Docker host |
| 2 | `log_search` pattern: "timeout" | Timeouts started at 09:40, all on one endpoint |
| 3 | `docker_get_stats` checkout-api | CPU and memory are normal, so it isn't the container |
| 4 | `run_query` EXPLAIN ANALYZE … | A sequential scan on orders since an index was dropped |

## Give agents read-only access to critical infrastructure

Let an agent look at production without handing it the keys. It can read what
it's granted; changing anything needs a role you give it.
[Read-only database access, in detail](/devops/ai-database-access.md).

- **No inbound port or VPN.** devops-client connects outward.
- **Read-only unless you say otherwise.** Writes need a writer role, checked at
  Levirge DevOps and again at devops-client. A connection can also be set
  read-only.
- **Credentials stay on the host.** Passwords, connection strings, hosts and
  file paths aren't sent over the message bus.
- **Every agent action on record.** The audit log marks what an agent did apart
  from what a person did.
- **Errors that say what to do next.** A failed call returns a code and, where
  there is one, the next tool to call.

## Versus one MCP server per system

| | One server per system | Levirge DevOps |
| --- | --- | --- |
| Network | A way in to each network | Outbound-only, no inbound port or VPN |
| Connections | One per database, log host and Docker host | One MCP connection for all of them |
| Credentials | Copied to wherever each server runs | Stay on the host with devops-client |
| Audit | Varies by server | One log, agent actions marked |

## FAQ

### Can an AI agent monitor my infrastructure without a VPN or inbound port?

Yes. devops-client runs next to each system and connects outward through a
message bus to Levirge DevOps. The agent connects to Levirge DevOps over MCP,
so nothing behind your firewall needs an inbound port or VPN.

### Is it safe to give an AI agent access to production systems?

Agents are read-only unless you grant a writer role. Writes are checked twice,
at Levirge DevOps and again at devops-client, and a connection can also be set
read-only. Credentials stay on the host, and the audit log marks what an agent
did apart from what a person did.

### Can I give an AI agent read-only access to critical infrastructure?

Yes, and it is the default. An agent can read the databases, logs, containers
and message buses it is granted, but cannot change them unless you give it a
writer role. You can also set a connection read-only, which blocks writes for
everyone, agents included.

### Which AI agents work with Levirge DevOps?

Claude or any other MCP client. The agent gets one MCP connection that covers
every system where devops-client runs.

### What can the agent check?

PostgreSQL, MySQL/MariaDB, SQLite, MongoDB, InfluxDB and Redis/Valkey; log
files and log directories; Docker containers, logs and stats; message bus
streams and consumer lag; and llama.cpp model servers. Each plugin page lists
its MCP tools.

### How is this different from the Datadog or Grafana MCP servers?

Those let an agent query telemetry you have already sent to that platform.
Levirge DevOps reaches the systems themselves, so an agent can run a query,
tail a log file or read container stats directly. The two can sit side by
side.

### Can I install it today?

Not yet. Levirge DevOps is in private testing and devops-client has no public
download. [Join the waitlist](/devops#cta) for one email when it is ready to
install.