Skip to main content
Glama

mikrotik-mcp

mcp-name: io.github.StefanKnol/mikrotik-mcp

An MCP server for MikroTik RouterOS, over the binary API rather than by driving the CLI over SSH.

That distinction is the whole point. The CLI prints positional numbers, which are not a rule's identity. A server that lists rules by position and then writes by position either fails outright — where .id=3 matches nothing, so every write reports "not found" for rules that plainly exist — or, worse, succeeds against a different rule once the order has shifted. On a firewall that means deleting the wrong rule.

The binary API returns the real .id (*7, *1f) on every read. So:

  • every read returns id, and every write takes one back;

  • list results also carry position, which is display-only and refused for writes, with an error that explains why;

  • writes read back what they wrote, so a change can be verified rather than assumed;

  • remove_firewall_rule takes an optional confirm_comment and returns the rule it deleted.

Use it

uvx mikrotik-mcp

Configure through the environment — ps would show a password passed as a flag:

Variable

Default

Meaning

MIKROTIK_HOST

—

Router address. Required.

MIKROTIK_USERNAME

—

Required.

MIKROTIK_PASSWORD

—

MIKROTIK_PORT

8729

8729 for api-ssl, 8728 for plaintext.

MIKROTIK_TLS

true

MIKROTIK_TLS_FINGERPRINT

—

SHA-256 of the router certificate, to pin it.

MIKROTIK_TIMEOUT

10

In an MCP client's config:

{
  "mcpServers": {
    "mikrotik": {
      "command": "uvx",
      "args": ["mikrotik-mcp"],
      "env": {
        "MIKROTIK_HOST": "192.168.88.1",
        "MIKROTIK_USERNAME": "mcp-agent",
        "MIKROTIK_PASSWORD": "..."
      }
    }
  }
}

Related MCP server: MikroTik MCP

On the router

/ip service enable api-ssl
/user group add name=mcp policy=api,read,write,test
/user add name=mcp-agent group=mcp password=<strong-password>

The api policy is not optional, and its absence produces a login error identical to a wrong password — so if credentials look right and login still fails, check the group first.

Set MIKROTIK_TLS_FINGERPRINT if you can. MikroTik's API-SSL certificate is self-signed, so ordinary CA validation cannot succeed against a stock device; pinning is what makes the connection authenticated rather than merely encrypted.

Tools

34 of them, covering system info, interfaces, IP addressing, firewall filter (IPv4 and IPv6), firewall address lists, NAT, DHCP leases, DNS, routes and logs — plus ros_list, which reads any RouterOS path and so covers everything without a dedicated tool.

There is deliberately no generic write escape hatch.

Three things worth knowing before the first call:

  • Writes take an id (*7), never a position. Both appear in every list result. position is where a rule currently sits in evaluation order; it shifts whenever anything is added, removed or moved, and it is counted across the whole table rather than within the chain you filtered to. Passing one to a write is rejected rather than guessed at.

  • IPv4 and IPv6 are separate rule sets. The firewall tools take a family and default to ipv4, so a device can read as locked down while its IPv6 table is empty and therefore accepting everything. Listing IPv4 reports the IPv6 rule count for exactly this reason.

  • Address lists hold state no configuration export shows. Entries a rule adds at runtime carry a timeout and exist only in memory. Ask list_address_list_entries for a count_only first: a populated block list can hold tens of thousands.

With mcphub

mcphub puts MCP servers behind one sign-in and grants them out per account. One router is one backend, at its own /mcp/<slug>, registered in a client as its own connector.

Adding it

From the registry — the better default. Add from registry, search for mikrotik-mcp, and the hub builds the settings form from server.json and launches uvx mikrotik-mcp when you enable the backend. It runs in its own process and cannot read credentials the hub holds for anything else. Backends added this way start disabled: open it, look at the tool list, then enable.

As a plugin — install this package into the hub's own environment and mikrotik appears as a backend kind, with typed fields and a Test button that reports the router's identity and RouterOS version. Nicer to configure, but an in-process plugin can read every credential the hub holds. Use it when you build your own hub image.

What to set

Either route asks for the same things. As environment variables:

Variable

Default

MIKROTIK_HOST

—

Router address. Required.

MIKROTIK_USERNAME

—

Required. A dedicated user, not admin; its group needs the api policy.

MIKROTIK_PASSWORD

—

Required.

MIKROTIK_PORT

8729

8729 for api-ssl, 8728 for plaintext.

MIKROTIK_TLS

true

Turning it off sends the router password over the network in the clear.

MIKROTIK_TLS_FINGERPRINT

—

SHA-256 of the router certificate. Pinning it is what makes the TLS connection authenticated rather than merely encrypted.

MIKROTIK_TIMEOUT

10

Seconds.

No data directory: everything this server reads and writes lives on the router, so there is no local state to keep. Leave Give this server a data directory unticked.

Levels

The hub reads readOnlyHint and destructiveHint from each tool and enforces the grant's level from those alone — a tool above the level is left out of tools/list and refused if called anyway.

Level

Gets

viewer

17 tools

Reads the whole configuration and changes nothing.

user

+7 tools

Adds rules and entries, and disables or re-enables them.

admin

+10 tools

Deletes, edits in place, and reorders.

viewer — connectivity_check, get_dns_settings, get_firewall_rule, get_logs, get_nat_rule, list_address_list_entries, list_address_lists, list_dhcp_leases, list_dns_adlist, list_dns_static, list_firewall_rules, list_interfaces, list_ip_addresses, list_nat_rules, list_routes, ros_list, system_info.

user adds add_address_list_entry, add_dns_static, add_firewall_rule, add_nat_rule, make_lease_static, set_firewall_rule_enabled, set_nat_rule_enabled. Each of these writes, and none of them takes anything away: adding a rule is a write, not a destruction, and disabling one keeps the rule, its comment and its position so that re-enabling restores exactly what was there. Disabling is the reversible way to find out whether a rule is responsible for something, so it sits here rather than behind admin — the alternative would leave a user with no safe way to test at all.

admin adds the destructive ten:

Tool

What is lost

remove_firewall_rule

The rule. The returned copy is the only record.

remove_nat_rule

The rule.

remove_dns_static

The entry.

remove_address_list_entry

The entry. A dynamic one returns when its rule next matches; a static one does not.

update_firewall_rule

The previous values of whatever it sets or clears.

update_nat_rule

The same.

update_dns_static

The same.

move_firewall_rule

The previous order, which is the entire semantics of a firewall.

move_nat_rule

The same.

set_interface_enabled

Nothing on disk — but disabling the interface the request came in through severs the only route back, and nothing here can undo that remotely.

A test pins every one of these assignments by name, so a tool added later fails the suite until someone decides which level it belongs to.

Publishing to the MCP registry

server.json is the manifest for the official MCP registry, validated against the published schema. It declares the uvx mikrotik-mcp command and every MIKROTIK_* variable, marking which are required and which are secret — so a client that browses the registry can generate a correct settings form without knowing anything about this server.

Publishing has an order to it, because the registry verifies that whoever publishes an entry actually owns the package it points at.

1. The package must already be on PyPI. The registry fetches pypi.org/pypi/mikrotik-mcp/<version>/json and refuses an entry whose package does not exist. Tag a release and let CI publish it:

git tag v0.1.0 && git push --tags

That needs a PyPI Trusted Publisher first, and because this project does not exist on PyPI yet it has to be a pending publisher — the per-project Publishing tab only appears once a project exists, which is the chicken-and-egg this page solves:

pypi.org → Account settings → Publishing → Add a new pending publisher

Field

Value

PyPI Project Name

mikrotik-mcp

Owner

StefanKnol

Repository name

mikrotik-mcp

Workflow name

ci.yml

Environment name

pypi

The environment name matters: the publish job declares environment: pypi, and PyPI rejects the upload if they disagree.

2. The README must carry the ownership token. The registry looks for mcp-name: io.github.StefanKnol/mikrotik-mcp in the PyPI description, which is this file — it is at the top, on its own line. That is what proves the person publishing the registry entry controls the PyPI package.

3. The registry entry publishes itself. The registry job in ci.yml runs after the PyPI upload on the same tag, authenticating with mcp-publisher login github-oidc. GitHub signs a short-lived OIDC token, the registry verifies it and grants the io.github.StefanKnol/* namespace from the repository owner. There is no secret to store and no device-flow login to sit through on each release.

It waits for PyPI to actually serve the new version first. The registry fetches pypi.org/pypi/mikrotik-mcp/<version>/json and refuses an entry whose package it cannot see, and that endpoint lags the upload by some seconds — long enough that publishing straight afterwards races it.

Re-running it is safe: a version already in the registry is the outcome the job exists to produce, so it reports that and passes rather than failing on the registry's duplicate refusal.

To publish by hand — recovering a failed run, or a version released before this job existed:

curl -sL https://github.com/modelcontextprotocol/registry/releases/latest/download/mcp-publisher_linux_amd64.tar.gz | tar xz mcp-publisher
./mcp-publisher login github && ./mcp-publisher publish

login github opens the device flow and proves you are StefanKnol, which authorises the same namespace the OIDC path gets without asking.

The version appears in four places — pyproject.toml, __version__, server.json, and again inside that file's packages entry. Rather than edit them by hand:

uv run python scripts/release.py 0.2.0 --tag

It sets all four, verifies them, commits and tags. The registry treats each version as its own row and checks PyPI has the package at exactly that version, so the nested packages[].version matters as much as the top-level one — and it is the one that gets missed, because the manifest still validates without it. A test enforces that they agree.

Development

uv sync --extra dev
uv run pytest

License

MIT

Available Tools

24 tools
add_dns_staticC

Add a static DNS entry.

Inputs: ttl, name, cname, regexp, +2 more. Returns: result

add_firewall_ruleA

Add a firewall filter rule.

Order decides everything in a RouterOS firewall: the first matching rule wins. A new rule appended after a final drop is dead code, so set place_before unless the end of the chain is genuinely what you want.

Inputs: chain (required), action (required), log, comment, +15 more. Returns: result

add_nat_ruleA

Add a NAT rule, such as a port forward.

A port forward also needs the forward chain to permit the traffic; check list_firewall_rules(chain='forward') for a rule accepting connection-nat-state=dstnat before concluding it is broken.

Inputs: chain (required), action (required), comment, disabled, +9 more. Returns: result

connectivity_checkA
Read-onlyIdempotent

Confirm the hub can reach and authenticate to this device.

Inputs: no named parameters. Returns: result

get_firewall_ruleA
Read-onlyIdempotent

Every field of one firewall rule, including counters.

Inputs: rule_id (required). Returns: result

get_logsB
Read-onlyIdempotent

Read the device log, most recent last.

Inputs: limit, topic, search. Returns: result

list_dhcp_leasesA
Read-onlyIdempotent

List DHCP leases — effectively the inventory of what is on the network.

Inputs: search, active_only. Returns: result

list_dns_staticA
Read-onlyIdempotent

List static DNS entries and regex overrides.

Inputs: no named parameters. Returns: result

list_firewall_rulesA
Read-onlyIdempotent

List firewall filter rules in evaluation order.

Each rule carries both an id and a position. Use id for any later change: position is only where the rule currently sits, and it shifts as soon as anything is added, removed or moved.

Inputs: chain, verbose, include_dynamic. Returns: result

list_interfacesA
Read-onlyIdempotent

List every interface with its type, MAC, MTU and running state.

Inputs: type_filter, running_only. Returns: result

list_ip_addressesA
Read-onlyIdempotent

List configured IPv4 addresses and the interfaces they sit on.

Inputs: no named parameters. Returns: result

list_nat_rulesA
Read-onlyIdempotent

List NAT rules, including port forwards, in evaluation order.

Inputs: chain, verbose. Returns: result

list_routesC
Read-onlyIdempotent

List the IPv4 routing table.

Inputs: active_only. Returns: result

make_lease_staticA
Idempotent

Pin a dynamic DHCP lease so the device keeps its address.

Inputs: lease_id (required), comment. Returns: result

move_firewall_ruleA
Destructive

Reorder a firewall rule relative to another rule.

Both ends are given by id rather than position, so the move means the same thing even if the list shifted since you last read it.

Inputs: rule_id (required), before_rule_id. Returns: result

remove_dns_staticA
DestructiveIdempotent

Delete a static DNS entry.

Inputs: entry_id (required). Returns: result

remove_firewall_ruleA
DestructiveIdempotent

Delete a firewall rule permanently.

There is no undo. Disabling a rule (set_firewall_rule_enabled) is almost always the better first move, and the returned copy of the deleted rule is the only record you will have afterwards.

Inputs: rule_id (required), confirm_comment. Returns: result

remove_nat_ruleA
DestructiveIdempotent

Delete a NAT rule permanently. Returns the rule that was removed.

Inputs: rule_id (required). Returns: result

ros_listA
Read-onlyIdempotent

Read any RouterOS configuration path.

The escape hatch for everything without a dedicated tool. Read-only: it lists items at a path and cannot modify the device.

Inputs: path (required). Returns: result

set_firewall_rule_enabledA
DestructiveIdempotent

Enable or disable one firewall rule, leaving its position intact.

Disabling is the reversible way to test whether a rule is responsible for something — prefer it over removing the rule.

Inputs: enabled (required), rule_id (required). Returns: result

set_interface_enabledA
DestructiveIdempotent

Bring an interface up or down.

Disabling the interface you are reaching the router through will cut this connection, and nothing here can undo that remotely.

Inputs: enabled (required), interface_id (required). Returns: result

set_nat_rule_enabledB
DestructiveIdempotent

Enable or disable one NAT rule.

Inputs: enabled (required), rule_id (required). Returns: result

system_infoA
Read-onlyIdempotent

Identity, RouterOS version, model, uptime, CPU and memory for the device.

Inputs: no named parameters. Returns: result

update_firewall_ruleA
DestructiveIdempotent

Change fields on an existing firewall rule.

Only the arguments you pass are modified; omitted ones are left alone. Returns the rule as it stands afterwards.

Inputs: rule_id (required), log, chain, action, +14 more. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 24 tools

    Disambiguation5/5

    Each tool targets a distinct resource/action pair, from list_interfaces to move_firewall_rule, so an agent can reliably select the right operation. ros_list is broad but is explicitly framed as a read-only escape hatch, not a competing path.

    Naming Consistency4/5

    Most tools follow a clean list_/add_/remove_/set_*_enabled pattern, and the firewall group is especially consistent. A few stragglers like system_info, ros_list, and connectivity_check break the verb-first convention without causing real confusion.

    Tool Count4/5

    24 tools is at the upper edge of reasonable, but the server spans many distinct MikroTik subsystems: interfaces, firewall, NAT, DHCP, DNS, routes, and logs. Each tool earns a place, and there is little obvious redundancy beyond the intentional ros_list fallback.

    Completeness3/5

    Firewall lifecycle is well covered with create, read, update, delete, move, and enable/disable. However, NAT and DNS entries lack update operations, routes are read-only, and ros_list cannot write, so some configuration changes require workarounds or are impossible.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to interact with MikroTik RouterOS devices through API and SSH connections, supporting network monitoring, configuration management, and diagnostics across multiple routers with automatic connection fallback.
      3
      MIT
    • A
      license
      C
      quality
      A
      maintenance
      Provides a bridge between AI assistants and MikroTik RouterOS devices, enabling natural language management of VLANs, firewall rules, DNS settings, and more.
      239
      2,499 PyPI
      348
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables management of MikroTik routers running RouterOS 6 and 7 via SSH, Telnet, or API with automatic command adaptation. Provides over 46 MCP tools for device management, firewall, DHCP, VPN, configuration profiles, and more.
      3
      MIT