mcp-syslog-crunchtools

MCP server for the logs collected by crunchtools/syslog.

Built for RT #1460, to close a specific gap: Hermes gets paged by Nagios and can restart a service, but it cannot read the service's logs — so every remediation is a blind restart. This turns that into an informed one.

Capabilities

The triage loop

nagios_current_problems_tool          → what is broken
syslog_search_tool(source=…,          → why it broke
                   severity="ERR",
                   since="15m")
syslog_context_tool(timestamp=…)      → what else was happening at that moment
nagios_schedule_check_tool            → confirm the fix

syslog_context_tool without a source is the one that earns its keep. It spans every source at once, which is how "the app died" gets connected to "the database container OOMed four seconds earlier".

Design notes

Every result is bounded, and says when it is. Logs are unbounded and this output lands in a model's context window. Each tool caps its results, caps how many lines it will scan, and annotates the answer when either limit is hit:

[!] Stopped after the 2,000,000-line scan limit, so this result is INCOMPLETE
    and an empty or short result does not mean nothing happened.

That annotation is load-bearing. A caller that cannot distinguish "no errors occurred" from "I stopped looking" will draw the wrong conclusion from an empty result — and this server exists to inform remediation decisions.

Time filtering is cheap. The collector puts the date in the filename, so a ten-minute query opens one file rather than reading ninety days of history.

Source names are untrusted. They come from a model and are used to build a filesystem path. Each is resolved and then confirmed to still be inside the log root, which catches traversal, absolute paths, and symlinks pointing out of the tree — see tests/test_security.py.

Severity is "at least this severe". severity="ERR" returns ERR, CRIT, ALERT and EMERG. An unrecognised severity is kept rather than dropped, on the grounds that hiding a line you do not understand is worse than showing it — but it is not counted as an error in syslog_stats_tool, or a healthy service would report a 57% error rate.

Severity is not badness. Podman records anything a container writes to stderr at priority err, and plenty of services log routine INFO there. One chatty httpx-based service in this fleet sits around 65% "ERR" while being entirely healthy:

PRIORITY=3 | 2026-08-23 15:55:24 INFO  httpx: HTTP Request: GET https://... "200 OK"

The collector is reporting the journal faithfully; the journal is reporting the file descriptor. Read the message, and prefer a change in error rate to its absolute value. This caveat is in the server's MCP instructions too, so an agent querying it is told the same thing.

Two line formats are parsed. The collector emitted five fields before 2026-08-23 and six after, and the old lines stay in retention for 90 days. Which layout a line uses is decided by where a real severity sits, not by counting fields.

Log format

The collector writes six space-delimited fields:

2026-08-23T15:41:52+00:00 crunchtools.com crunchtools.com httpd ERR AH00169: caught SIGTERM
└─ timestamp ───────────┘ └─ host ──────┘ └─ source ────┘ └prog┘ └sev┘ └─ message ────────┘

source is the log stream — normally a container name. program is the process inside it, which matters for systemd containers where httpd, php-fpm and mariadb all file under one service name.

Configuration

No credentials — the server reads files off a read-only bind mount.

Running

podman run -d --name mcp-syslog \
  --network crunchtools \
  -p 127.0.0.1:8027:8027 \
  -v /path/to/syslog/data/logs:/logs:ro \
  quay.io/crunchtools/mcp-syslog:latest \
  --transport streamable-http --host 0.0.0.0 --port 8027

Mount :ro. This server never needs to write, and a read-only mount means a bug here cannot destroy the forensic record it exists to protect.

Development

uv sync
uv run ruff check src tests
uv run mypy src
uv run pytest -v