mcp-server-malcolm

English | 繁體中文

The first MCP server for Malcolm, the open-source network traffic analysis platform (Zeek + Suricata + Arkime + OpenSearch, with optional NetBox).

It gives any MCP-compatible AI agent structured access to Malcolm: search and aggregate network traffic, discover field names, query Suricata alerts, browse Arkime sessions, resolve NetBox assets, and check system health. Turn on the write classes and it can also create alerts, tag sessions, launch hunts, and upload PCAP.

Read-only until you opt in

With no configuration, this server exposes read tools only. It behaves like a read-only client, and nothing it does can change data in Malcolm.

The server splits write access into five classes, each behind its own environment flag and each off by default. It doesn't register a disabled class, so that class's tools never appear in list_tools() and can't be called. At startup it prints which classes are on:

[mcp-server-malcolm] write classes: alerting=off arkime-tag=off hunt-job=off pcap-upload=off arkime-view=off

Every write but one is additive: the exception is arkime_cancel_hunt, which stops a hunt job in progress rather than adding to it. None of them deletes data, removes a tag, or touches a user account — those stay out on purpose (see Non-goals).

Related MCP server: netbox-mcp

Why an MCP layer

Malcolm keeps all network metadata in one OpenSearch index (arkime_sessions3-*) with non-standard field names and its own filter syntax. An LLM asked to write raw OpenSearch DSL against that index gets it wrong more often than not. This server takes that job off the model:

• It exposes Malcolm's filter syntax instead of raw DSL.

• It provides field discovery so the model checks field names before it queries.

• It provides value enumeration so the model sees what values a field actually holds.

• It covers both field vocabularies. Arkime expressions take Arkime's own names (ip.src), the rest of Malcolm takes ECS names (source.ip), and Malcolm's own field list carries only the second set. arkime_field_search supplies the first.

• It wraps Suricata alert queries and handles the field mapping (suricata.alert.* vs rule.*).

• It adds NetBox asset context (IP-to-device, network segments).

The failure mode this is built against is a quiet one. Malcolm answers a query against a field it does not index with an empty result rather than an error, so a model that guesses a plausible-but-wrong name reads "no such traffic" and moves on. When a search comes back empty, this server checks the fields the query named and reports the name Malcolm actually stores the value under. That lookup runs only after a result set is already empty, so nothing is added to the model's context on queries that worked.

The write side follows the same idea. Rather than hand an agent the raw OpenSearch and NetBox passthroughs that Malcolm already leaves open to any authenticated user, this server exposes a small, named, audited set of write actions. More on that under Security model.

Read tools

These are always registered.

DSL core (backend-agnostic)

Plain OpenSearch DSL against the configured endpoint (Malcolm's /mapi/opensearch proxy). No Malcolm-specific query shape: point the base URL at any OpenSearch-compatible backend and they still work.

Core query

Field discovery (anti-hallucination)

These three malcolm_* tools cover the ECS names used by malcolm_search, malcolm_aggregate and the DSL tools. Anything going into an expression argument needs arkime_field_search instead: Arkime's parser accepts ip.src and rejects source.ip, and Malcolm's /mapi/fields does not list the expression names at all.

System health

Asset context (NetBox)

Arkime

Arkime saved objects and capture health

Arkime's connections.csv is deliberately not wrapped: on Arkime 6.6.0 it emits a nine-column header over seven-column rows, so every column after the second is mislabeled. arkime_connections answers the same question correctly.

File analysis

malcolm_file_scans reads Zeek's record of every file transfer it saw, which does not require the file extractor. Reaching the file itself does: malcolm_extract_file needs ZEEK_EXTRACTOR_MODE set and the extracted-files HTTP server on (FILESCAN_HTTP_SERVER_ENABLE), and a scan row only appears where Strelka is running. A carved file may be live malware, so the bytes never enter the MCP response.

Correlation and export

The 15 tools that build their own rows — the file, Arkime-inventory and Dashboards ones — declare a typed return, so a client receives structuredContent as well as the text. The rest pass an upstream response through verbatim and have no shape to declare.

Write tools (opt-in)

Each class is enabled by setting its flag to true. Nothing here runs unless you ask for it.

• alerting: malcolm_create_alert indexes an analyst- or agent-generated finding as an alert document you can see in Malcolm's dashboards. It uses /mapi/event, Malcolm's own purpose-built write endpoint, which is the template the other classes follow.

• arkime-tag: arkime_add_tags adds tags to sessions. It only adds; tag removal needs a higher Arkime role and its own safety design, so it's deferred.

• hunt-job: arkime_create_hunt launches a cross-PCAP packet search (expensive, so scope the query first). arkime_cancel_hunt stops one that is queued or running — not additive, since a cancelled scan can't resume. Job progress is read with arkime_hunt_status, which is a read tool now and stays available with this class off (see Arkime saved objects and capture health).

• pcap-upload: malcolm_upload_pcap sends a local capture file to Malcolm for ingestion, with a client-side size cap. The file must live inside MALCOLM_MCP_UPLOAD_DIR; if that staging directory is unset, uploads are refused, so the tool can never be steered into reading an arbitrary file off the host.

• arkime-view: arkime_create_view saves a named search expression and arkime_create_shortcut saves a named value list (IOC set) referenced in expressions as $name. Both are additive — they let an agent persist hunting knowledge for the human team, and neither deletes or overwrites.

Every write tool carries the MCP annotation readOnlyHint: false, so an MCP client can apply its own confirmation step before the call runs. destructiveHint is false on every additive write and true on the one exception, arkime_cancel_hunt, which stops in-progress work rather than adding to it.

Security model

Malcolm's default deployment already gives any authenticated user unrestricted write access to raw OpenSearch (/mapi/opensearch/*) and full NetBox CRUD (/mapi/netbox/*). Both are bare reverse-proxies with no HTTP-verb filtering; Malcolm's own read-only mode removes them rather than trying to filter them. In the common auth modes, "logged in" means admin-equivalent.

Turning on a write class here does not open a door that was otherwise shut. That door is already open at the platform level. This server adds a curated way through it:

• A small, named set of write actions instead of a raw passthrough.

• Off by default, enabled one class at a time.

• An audit line for every write attempt.

• MCP annotations so the client can require confirmation.

This server does not expose the raw OpenSearch and NetBox write passthroughs, behind a flag or otherwise. Curating that surface is what it is for.

Audit

Every write attempt emits one line of JSON, on success and on failure:

{"ts": "2026-07-06T09:12:44Z", "tool": "arkime_add_tags", "class": "arkime-tag", "target": "ids=240601-abc", "params": {"tags": "suspicious"}, "outcome": "ok"}

outcome is one of ok, http_4xx, http_5xx, or error:<type>. Long parameter values are truncated, and PCAP bytes are never logged. The sink is stderr by default; set MALCOLM_MCP_AUDIT_FILE to append to a file instead. Read tools are not audited.

Quick start

You don't write any code to use this. An MCP client (Claude Code, Claude Desktop, Cursor, …) launches the server as a subprocess and talks to it over stdio; your job is to tell the client how to launch it and which credentials to inject.

Every command in this chapter was run as printed, on Linux/aarch64 (kernel 6.14, Python 3.11.14 and 3.14.6) against a live Malcolm v26.07.1, and the error text is verbatim. Where something was reasoned from source rather than executed, or was left untested (x86_64 and macOS hosts, GUI MCP clients, four of the five write classes), it says so at that point.

1. Install

PyPI currently carries an older build than this repository. The published mcp-server-malcolm 0.9.0 was cut from an earlier commit; this tree's pyproject.toml also says 0.9.0, so the version number gives you no signal that the two differ. pip install mcp-server-malcolm and bare uvx mcp-server-malcolm install the published release, not the code documented here. Until the next release, install from source to get this tree.

pip install mcp-server-malcolm      # published release

From a checkout:

git clone https://github.com/nagameTW/mcp-server-malcolm.git
cd mcp-server-malcolm
pip install -e .

Or build a wheel and install it into a clean virtualenv, which is the path everything below was verified through:

$ uv build --out-dir /tmp/mcp-malcolm-deploy/dist
Successfully built /tmp/mcp-malcolm-deploy/dist/mcp_server_malcolm-0.9.0.tar.gz
Successfully built /tmp/mcp-malcolm-deploy/dist/mcp_server_malcolm-0.9.0-py3-none-any.whl

$ python3 -m venv /tmp/mcp-malcolm-deploy/venv
$ /tmp/mcp-malcolm-deploy/venv/bin/pip install \
    /tmp/mcp-malcolm-deploy/dist/mcp_server_malcolm-0.9.0-py3-none-any.whl

That pulls 32 packages, most of them from mcp>=2,<3 (resolved to mcp 2.0.0). The wheel itself is py3-none-any, pure Python; the compiled dependencies (cryptography, pydantic-core, rpds-py, cffi) all installed from prebuilt manylinux_*_aarch64 wheels here, nothing compiled from source. PyPI publishes the same wheels for x86_64 and macOS, but no install was run on either, so treat those as unverified.

To run this branch without installing it anywhere permanent, point uvx or pipx at the checkout:

uvx --from /path/to/mcp-server-malcolm mcp-server-malcolm
pipx run --spec /path/to/mcp-server-malcolm mcp-server-malcolm

Check the install by starting the server with stdin closed. It prints its write-class banner, reaches EOF, and exits 0:

$ timeout 3 mcp-server-malcolm < /dev/null
[mcp-server-malcolm] write classes: alerting=off arkime-tag=off hunt-job=off pcap-upload=off arkime-view=off
$ echo $?
0

Nothing has to be configured for the process to start. Connection settings are read at startup but not used until a tool calls Malcolm, so a wrong URL or password surfaces as a failing tool call, not a failed launch.

2. Register it with your client

Claude Code — one command, no config file to find:

claude mcp add malcolm \
  -e MALCOLM_URL=https://malcolm.example \
  -e MALCOLM_USERNAME=analyst \
  -e MALCOLM_PASSWORD='your-password' \
  -e MALCOLM_SSL_VERIFY=/path/to/malcolm-ca.crt \
  -- mcp-server-malcolm

Everything after -- is the launch command; each -e is an environment variable injected into it. claude mcp add --help gives the signature as claude mcp add [options] <name> <commandOrUrl> [args...], with -e, --env <env...> and -s, --scope <scope>.

Registering, health-checking and removing a server, run end to end:

$ claude mcp add malcolm-deploy-test -s local \
    -e MALCOLM_URL=https://malcolm.example \
    -e MALCOLM_USERNAME=analyst \
    -e MALCOLM_PASSWORD='your-password' \
    -e MALCOLM_SSL_VERIFY=false \
    -- /tmp/mcp-malcolm-deploy/venv/bin/mcp-server-malcolm
Added stdio MCP server malcolm-deploy-test with command: … to local config

$ claude mcp list
Checking MCP server health…
malcolm-deploy-test: /tmp/mcp-malcolm-deploy/venv/bin/mcp-server-malcolm  - ✔ Connected

$ claude mcp remove malcolm-deploy-test -s local
Removed MCP server malcolm-deploy-test from local config

Pick where the entry is stored with -s:

claude mcp get malcolm prints the registered command and environment. Note that it prints MALCOLM_PASSWORD in cleartext, unmasked, so don't run it where the terminal is being recorded or shared.

For a project-scope entry, keep the secret in each person's shell rather than in the file:

{
  "mcpServers": {
    "malcolm": {
      "command": "mcp-server-malcolm",
      "env": { "MALCOLM_PASSWORD": "${MALCOLM_PASSWORD}" }
    }
  }
}

Other MCP clients — no equivalent CLI, so edit the client's own JSON config. The block is the same shape:

{
  "mcpServers": {
    "malcolm": {
      "command": "mcp-server-malcolm",
      "env": {
        "MALCOLM_URL": "https://malcolm.example",
        "MALCOLM_USERNAME": "analyst",
        "MALCOLM_PASSWORD": "your-password",
        "MALCOLM_SSL_VERIFY": "/path/to/malcolm-ca.crt"
      }
    }
  }
}

That exact block was verified by driving its command and env fields through the MCP Python SDK's own stdio_client and ClientSession, which is what a generic client does with them. No GUI client was launched here: Claude Desktop reads claude_desktop_config.json and other clients vary, per their own docs, which this project has not independently confirmed.

If mcp-server-malcolm isn't on the PATH your client sees (common with a virtualenv), give the absolute path to the executable instead: /path/to/.venv/bin/mcp-server-malcolm.

What the client sees on connection. With no write flags set, an initialize plus tools/list returns:

protocol_version: 2025-11-25
server_info:      name='mcp-server-malcolm' version='0.9.0'
capabilities:     prompts, resources (subscribe=false), tools — all list_changed=false
instructions:     3624 characters
tools:            51
prompts:          1  — hunt_workflow
resources:        2  — malcolm://fields/malcolm, malcolm://fields/arkime

The Arkime resource served 724,261 characters covering 4,051 expression fields on this deployment. One caveat for anyone scripting against the SDK: mcp 2.x uses snake_case attributes (protocol_version, server_info, is_error), not the camelCase of the wire protocol and the 1.x SDK. A script written against serverInfo/isError raises AttributeError.

3. Connection settings

Defaults below are what MalcolmClient.from_env reads (client.py:294-304).

The https://localhost and true defaults were confirmed by running with the variable unset and observing the request that came out. The admin/admin credential defaults come from reading client.py:296-297: unsetting all three connection variables produced a 401 against https://localhost, which proves the URL default and proves the credentials are wrong for that lab, not that they are literally admin. The 30-second timeout is likewise a source read — an attempt to time it against a non-routable address returned in about 5 seconds, because the OS-level connect failure fired first, so the 30-second path was never exercised.

On TLS: verification is on by default, and Malcolm ships self-signed certs. Pointing MALCOLM_SSL_VERIFY at Malcolm's CA bundle only works if Malcolm's server certificate carries a subjectAltName matching the hostname you connect to — the certificate generated by Malcolm's own setup has no SAN extension, so verification fails against it even with the right CA. For a remote Malcolm, install a certificate with a correct SAN. MALCOLM_SSL_VERIFY="false" disables verification entirely and is only acceptable against an isolated localhost lab; over a network it would send credentials and query results down an unauthenticated channel.

When the first call fails

Three failures account for nearly every first run. All three were reproduced with malcolm_ping; the text is verbatim.

Self-signed certificate with MALCOLM_SSL_VERIFY unset. This is the most likely one, because the default is verify-on and a stock Malcolm's certificate does not pass verification:

Error executing tool malcolm_ping: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: unable to get local issuer certificate (_ssl.c:1082)

The _ssl.c line number tracks your interpreter, not this server: the same failure reads _ssl.c:1016 on Python 3.11, which is what the Docker image and the CI floor both run.

The message never names MALCOLM_SSL_VERIFY, so it is easy to read as a broken install. Fix it by installing a certificate with a correct SAN and pointing MALCOLM_SSL_VERIFY at its CA bundle, or, on an isolated lab only, by setting MALCOLM_SSL_VERIFY=false.

Wrong password. Clear and actionable — the status and the URL are both in the message:

Error executing tool malcolm_ping: Client error '401 Unauthorized' for url 'https://malcolm.example/mapi/ping'
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/401

Unreachable host (typo'd MALCOLM_URL, firewalled port, wrong scheme):

Error executing tool malcolm_ping: ConnectTimeout for https://192.0.2.99:9999/mapi/ping

The two unreachable cases read differently. A port that actively refuses the connection answers All connection attempts failed; a host that simply never replies raises ConnectTimeout, whose str() httpx leaves empty — up to 1.0.1 that reached the caller as a bare Error executing tool malcolm_ping: with nothing after the colon. The exception name and target are filled in now. Credentials embedded in MALCOLM_URL are stripped from that URL before it is shown.

4. Enabling write tools (optional)

All five write classes are off unless you set their flag, so the default install is read-only. Add the flags to the same -e / env block:

-e MALCOLM_MCP_ENABLE_ALERTING=true
-e MALCOLM_MCP_AUDIT_FILE=/var/log/malcolm-mcp-audit.jsonl

A disabled class is not registered rather than hidden, so the change shows up in tools/list. Counting the tools an MCP session sees, with nothing else changed:

env unmodified                          tool count: 51
MALCOLM_MCP_ENABLE_ARKIME_TAGS=true     tool count: 52   (new: arkime_add_tags)

Only arkime_tags was toggled and counted this way. The other four classes route through the identical if cfg.<flag>: gate in tools/__init__.py::register_write_tools, so the same behaviour follows from the code, but it was not separately measured.

A flag counts as on only for the exact string true, case-insensitive (config.py:15); anything else, including 1 and yes, leaves the class off. The startup banner is the check.

The full flag list is in Configuration reference.

Running it by hand

Only useful for troubleshooting. A stdio MCP server has no interactive interface: started from a terminal it sits silently waiting for JSON-RPC on stdin, which is what a working server looks like. It does print the enabled write classes to stderr on startup, so this confirms the flags took effect. Both entry points behave identically:

$ mcp-server-malcolm
[mcp-server-malcolm] write classes: alerting=off arkime-tag=off hunt-job=off pcap-upload=off arkime-view=off

$ python -m mcp_server_malcolm
[mcp-server-malcolm] write classes: alerting=off arkime-tag=off hunt-job=off pcap-upload=off arkime-view=off

Running it in a container

The repository ships a Dockerfile that installs the package from the build context and runs it as a non-root user:

$ docker build -t mcp-server-malcolm:local -f Dockerfile .
$ docker run --rm --entrypoint id mcp-server-malcolm:local
uid=10001(app) gid=10001(app) groups=10001(app)

The build took 23.7s cold and produced a 205MB image on python:3.11-slim (Debian 13 trixie). It is single-architecture: a plain docker build on this aarch64 host produced linux/arm64 only, which will not run on an x86_64 host without emulation. A multi-architecture image would need docker buildx build --platform linux/amd64,linux/arm64; that was not attempted, and this Dockerfile does not produce one.

Point the client at docker run -i --rm … as the launch command:

docker run -i --rm --network host \
  -e MALCOLM_URL -e MALCOLM_USERNAME -e MALCOLM_PASSWORD -e MALCOLM_SSL_VERIFY \
  mcp-server-malcolm:local

-e VAR with no =value inherits the value from the invoking shell, so the password is never part of the command string and never lands in ps output or shell history. It is still readable afterwards through docker inspect, as below.

How the container reaches Malcolm decides whether anything works at all:

Both working modes completed a full MCP session against the live Malcolm: initialize, tools/list returning 51 tools, and two tool calls (malcolm_ping → pong, count → 202,531 conn sessions). MALCOLM_SSL_VERIFY at its true default failed inside the container for the same self-signed-certificate reason it fails outside one.

On credentials: docker inspect <container> --format '{{json .Config.Env}}' prints MALCOLM_PASSWORD in cleartext, and does so regardless of whether the value was passed as -e VAR, -e VAR=value, or --env-file — Docker stores the resolved environment in the container's metadata either way. Anyone with Docker daemon or socket access can read the Malcolm password back out for as long as the container object exists. There is no MALCOLM_PASSWORD_FILE-style secrets-file input; client.py:297 reads the environment variable and nothing else. Keeping containers ephemeral (--rm, one per client session, which is the model a stdio server already implies) shortens the window without closing it.

MALCOLM_MCP_ENABLE_PCAP_UPLOAD is the one feature that needs a bind mount, since malcolm_upload_pcap reads a file that must already sit inside MALCOLM_MCP_UPLOAD_DIR. The host directory mounted there has to be readable by uid 10001 inside the container. That requirement is read from tools/write/pcap_upload.py, not exercised — the write classes stayed off throughout this testing.

Usage

Python (direct import)

MalcolmClient is usable on its own, with no MCP client, no server process and no mcp transport in the loop. mcp_server_malcolm's __all__ is ["MalcolmClient", "__version__"], and that one class carries 62 public methods covering the entire read surface. Everything in this section was run against a live Malcolm v26.07.1 from a wheel built from this tree.

Build a client either from the environment or from explicit arguments:

import asyncio
from mcp_server_malcolm import MalcolmClient

async def main():
    client = MalcolmClient.from_env()          # reads MALCOLM_URL, MALCOLM_USERNAME, …
    # or:
    # client = MalcolmClient(
    #     base_url="https://malcolm.example",
    #     username="analyst",
    #     password="…",
    #     ssl_verify=False,
    # )
    try:
        # Malcolm filter dict
        hits = await client.search(
            filters={"event.dataset": "conn"},
            limit=5,
        )

        # Top values of a field, over one 24-hour window
        agg = await client.aggregate(
            fields="destination.port",
            filters={"event.dataset": "conn"},
            limit=5,
            time_from="1714003200",
            time_to="1714089600",
        )

        # Check a field name before trusting a query that returned nothing
        ok = await client.resolve_field("http.useragent")
        bad = await client.resolve_field("http.user_agent")

        # Arkime expression syntax, epoch-second window
        sessions = await client.arkime_sessions(
            expression="protocols==dns",
            limit=3,
            time_from="1714003200",
            time_to="1714089600",
        )
    finally:
        await client.close()

asyncio.run(main())

What those calls actually returned:

search()     keys: ['filter', 'range', 'results']
aggregate()  {'destination.port': {'buckets': [{'doc_count': 82147, 'key': 53},
                                               {'doc_count': 31271, 'key': 80},
                                               {'doc_count': 27911, 'key': 8080}, …]}}
resolve_field('http.useragent')   {'exists': True,  'field': 'http.useragent', 'type': 'string'}
resolve_field('http.user_agent')  {'exists': False, 'field': 'http.user_agent',
                                   'suggestion': 'http.useragent', 'type': 'string'}
arkime_sessions()  recordsTotal: 6030807  recordsFiltered: 310414
                   first session id: 3@240425:240425-zT5pQlD2hY2Gwyzziep8Vg

search() returns {"filter", "range", "results"} on this Malcolm version, with the hits under results and no top-level total key. Read the payload you actually get rather than assuming a shape.

Closing the client is the caller's job. MalcolmClient has close() but no __aenter__/__aexit__, so async with MalcolmClient(...) raises:

TypeError: 'mcp_server_malcolm.client.MalcolmClient' object does not support
the asynchronous context manager protocol (missed __aexit__ method)

Use try/finally as above. Dropping the last reference without closing leaves a real open socket to Malcolm, reclaimed only when the garbage collector gets to it, and the warning is silent under normal interpreter settings — it appears only under python -W always -X dev:

ResourceWarning: unclosed <socket.socket fd=6, family=2, type=1, proto=6, …>

Errors. Three exceptions, all with MalcolmToolError as their base:

from mcp_server_malcolm.errors import MalcolmToolError, ToolInputError, UpstreamError

Observed messages:

ToolInputError:  invalid field name: '../../arkime/api/hunts' — expected a Malcolm
                 field name such as 'source.ip' (letters, digits and _ . - @ [ ])
UpstreamError:   status=404  message=Client error '404 Not Found' for url
                 'https://malcolm.example/dashboards/api/saved_objects/search/00000000-…'
UpstreamError:   status=None message=

That last one is the library-side view of the empty error message described under When the first call fails: str(exc) is the empty string. It was traced past this project's code to httpx.ConnectTimeout.__str__(), which is empty for the same host when called through a bare httpx.AsyncClient. Branch on exc.status is None to detect an unreachable host; the message will not tell you anything.

Rate limiting holds, it does not raise. MALCOLM_MAX_CONCURRENCY (default 8) and MALCOLM_MAX_REQUESTS_PER_MINUTE (default 600) also exist as the constructor arguments max_concurrency and max_requests_per_minute. With max_requests_per_minute=2, three sequential ping() calls completed at:

request 1  t+0.0s
request 2  t+0.0s
request 3  t+60.1s

with [malcolm] rate cap reached, holding https://…/mapi/ping for 60.0s on the DEBUG log. There is no rate-limit exception: past the cap, a call is indistinguishable from a slow one, so a per-call timeout tighter than the window will fire on what is really a queued request. The 60-second window itself is a module constant (_RATE_WINDOW_SECONDS), not configurable — only the request count inside it is. A non-positive constructor argument is rejected outright:

ValueError: max_concurrency must be a positive integer, got 0
ValueError: max_requests_per_minute must be a positive integer, got 0

The environment path is deliberately more forgiving: an absent, blank or unparseable value is logged and replaced with the default (client.py:68-82), so a typo in a deployment's env file cannot switch the limiter off.

There is no supported write path for a library user. All seven write primitives are private — _write_event, _write_arkime_tags, _write_arkime_view, _write_arkime_shortcut, _write_arkime_hunt, _write_arkime_hunt_cancel, _write_upload_pcap — and each is called from exactly one place under tools/write/, which a seam test enforces. No public method indexes an alert, tags a session, creates or cancels a hunt, saves a view or shortcut, or uploads a PCAP. Calling an underscore method directly performs the mutation while skipping the audit record that tools/write/_common.py::run_write writes around every write reached through the MCP layer, so writes belong to the tool layer, and the direct-import path is a read client.

Malcolm filter syntax

Malcolm uses a simple JSON filter syntax, not OpenSearch DSL:

# Exact match
{"event.dataset": "conn"}

# Multiple values (OR)
{"network.direction": ["inbound", "outbound"]}

# Negation
{"!network.transport": "icmp"}

# Field must exist (not null)
{"!related.password": null}

# Combined (AND)
{"event.dataset": "dns", "source.ip": "192.0.2.77"}

Values match exactly. Malcolm compiles this dict to an OpenSearch terms query, so there is no wildcard: {"rule.name": "*MALWARE*"} looks for a signature literally named *MALWARE* and quietly finds nothing. For substring matching either enumerate the values first with malcolm_field_values and pass the ones you want as a list, or use search_dsl and write the wildcard query yourself. malcolm_alerts does that enumeration for you on its signature and category arguments.

Examples

Search DNS queries to a suspicious domain

malcolm_search(
  filters='{"event.dataset": "dns", "zeek.dns.query": "ntp.ubuntu.com"}',
  limit=20,
  time_from="7 days ago"
)

Aggregate top talkers by protocol

malcolm_aggregate(
  fields="source.ip,destination.ip,network.protocol",
  filters='{"network.direction": ["inbound", "outbound"]}',
  limit=20
)

Verify field names before querying

malcolm_field_search(prefix="zeek.dns")
malcolm_field_values(field="event.dataset")
malcolm_field_profile(field="zeek.ssl.server_name")

# Before writing an Arkime expression, look the name up in Arkime's own
# vocabulary. Lines come back as "exp | db | type | group", e.g.
# "ip.src | srcIp | ip | general".
arkime_field_search(keyword="src")

Which of those columns a parameter wants is decided per parameter, not per tool. Measured on Malcolm v26.07.1:

• exp (ip.src, port.dst, protocols) — every expression argument, plus the field lists of arkime_unique, arkime_multiunique and arkime_spigraphhierarchy. Those three reject a db name outright: srcIp,dstIp returned the body "Unknown expression srcIp" under HTTP 200 from multiunique and HTTP 403 from spigraphhierarchy.

• db (srcIp, dstPort, node) — arkime_connections' src_field and dst_field, and nothing else. srcIp/dstIp returned a 10-node graph there; ip.src/dstIp returned HTTP 403 and srcIp/port.dst HTTP 500.

• The storage path — what arkime_spigraph's field and arkime_spiview's spi take. It is the same string as the db column for 4,034 of this deployment's 4,051 fields; the other seventeen print a camelCase db alias and store under a dotted name (srcIp is source.ip, dstPort is destination.port, totBytes is network.bytes), and those two parameters want the dotted one.

A dotted storage path is also accepted wherever the exp column is: destination.port returned the same 10,000 unique lines as port.dst, while dstPort returned none. It is the one spelling that answers on every route.

The HTTP responses quoted above are what Arkime itself answers. You will not see them through these tools: the server now recognises a db name on an exp parameter and refuses it before the request leaves, with a message naming the counterpart — 'srcIp' is an Arkime db name; this parameter takes expression names … Did you mean 'ip.src'?. They are quoted because they are why the guard exists.

Create an alert (alerting class enabled)

malcolm_create_alert(
  title="Periodic beacon to 192.0.2.77",
  severity=2,
  description="60s-interval C2 candidate",
  source_ip="192.0.2.10",
  dest_ip="192.0.2.77"
)

Tag sessions for review (arkime-tag class enabled)

arkime_add_tags(session_ids="240601-abc,240601-def", tags="review,beacon")

Launch a hunt (hunt-job class enabled)

arkime_create_hunt(
  name="beacon-bytes",
  search="deadbeef",
  search_type="hex",
  total_sessions=42,
  start_time=1717200000,
  stop_time=1717203600,
  expression="ip==192.0.2.77"
)

Configuration reference

Verifying against your own Malcolm

Every tool here reshapes what Malcolm returns — trimming, renaming, and in a few places correcting it. scripts/api_parity_check.py proves that reshaping never changes the facts: for each of the 51 tools it asks the same question twice, once through a real MCP stdio session and once with a direct HTTP call to Malcolm, and compares the values that carry meaning.

MALCOLM_URL=https://malcolm.example \
MALCOLM_USERNAME=... MALCOLM_PASSWORD=... \
PARITY_TIME_FROM=<epoch-seconds> PARITY_TIME_TO=<epoch-seconds> \
uv run --with mcp python scripts/api_parity_check.py

The time window matters: Arkime defaults to a recent one, so point it at a period your deployment actually holds. The script exits non-zero if any tool disagrees with the API, and also if a tool is exposed but has no comparison written for it — so adding a tool without a parity check fails the run.

Malcolm API endpoints used

These endpoint paths and body shapes match Malcolm 26.07.1 and Arkime 6.6.0. Both drift between releases, so re-check against your own version if a write tool returns an unexpected error.

Non-goals

Version 1 leaves these out on purpose:

• Destructive writes (Arkime session delete, tag removal, user management).

• Raw OpenSearch write or raw NetBox CRUD passthrough, behind a flag or otherwise.

• The streamable-http transport (stdio only).

Requirements

• Python 3.11+

• A Malcolm instance with API access

• Network connectivity to Malcolm (HTTPS)

License

MIT © nagameTW