graylog-mcp

A read-only Model Context Protocol server for Graylog 4.x to 7.x. It lets an LLM investigate incidents on its own: search logs, follow one request across services, group errors with exact counts, and find when a problem started. Sensitive data is masked before any log line leaves the machine.

Everything site-specific (trace field names, redaction rules, timezone, application packages, error query) lives in configuration, not in code.

How it differs from other Graylog MCP servers

Quick start

One command (needs uv), run in the folder holding your .graylog-mcp.toml or anywhere for a fresh start:

uvx --from git+https://github.com/ntbang0901/graylog-mcp graylog-mcp start

It installs graylog-mcp, runs it in the background as one process for every Claude session, starts it again when you log in, connects Claude Code in the repositories of your groups and opens the admin page. Its Get set up card shows what is left; Finish setup does every step that needs no input in one click and then opens the form for what does (usually: your Graylog URL and a token).

Afterwards:

Prefer the terminal? graylog-mcp init is a guided setup that writes .graylog-mcp.toml and registers the server per repository (one process per session). See Setup helpers.

Or by hand:

Two environment variables are enough:

export GRAYLOG_URL=https://graylog.example.com
export GRAYLOG_TOKEN=<access token of a read-only user>
uvx graylog-mcp --check      # connects, prints the detected version and APIs, exits

Claude Code:

claude mcp add graylog --env GRAYLOG_URL=https://graylog.example.com --env GRAYLOG_TOKEN=... -- uvx graylog-mcp

Claude Desktop / any MCP client (mcpServers JSON):

{
  "mcpServers": {
    "graylog": {
      "command": "uvx",
      "args": ["graylog-mcp"],
      "env": {
        "GRAYLOG_URL": "https://graylog.example.com",
        "GRAYLOG_TOKEN": "...",
        "GRAYLOG_TIMEZONE": "Asia/Ho_Chi_Minh",
        "GRAYLOG_REDACTION_PACKS": "vn"
      }
    }
  }
}

The Graylog user

Create a dedicated user with the Reader role and read access to the streams the model may see, then generate an access token for it. The server only sends GET requests, plus POST to the four search endpoints that execute or check a query without saving it (/views/search/sync, /search/messages, /search/aggregate, /search/validate). Any other method or path is refused inside the client before a request is built.

Tools

All tools are annotated readOnlyHint and accept an optional instance. Tools that do several related things take a kind, view or what; an argument that does not apply to the one chosen is refused, not ignored.

Scan

• scan — "is anything wrong?" in one call. Runs every scan rule concurrently (crashes, resource exhaustion, error spikes, new error types, HTTP 5xx, connectivity, database, auth failures, plus your own) with exact counts against a baseline, and returns only what fired, most severe first, with its query, top groups and a sample. Ad hoc rules (checks) cover a specific request in the same call. See Scan rules.

Root cause analysis

• root_cause — "what broke first, and why?" in one call. Compares every service's errors, traffic and latency with a baseline window, pins each service's first error to the millisecond, detects deploys in the logs, infers the call graph from traces, and returns a ranked verdict with a timeline and evidence.
  • view="changes" — only the deploys and restarts found in the logs themselves: a new value of a version field (app_version, build, commit, ...), a host rollout (new sources replacing old ones), start/stop lines. No CI/CD integration needed.
  • view="service_map" — only which service calls which, inferred from sampled traces with no configuration: edges with traffic, error rate and p50/p95 latency.

Search

• search_logs — Lucene query, relative (15m, 2h) or absolute time (instance timezone or ISO 8601), streams by name or id, field selection, sort, paging; repeated lines grouped by default.
• get_message — one message with all fields, by index/id ref. Long values are cut and listed in cut_fields; name them in fields to read them in full. With context_seconds: the messages within ±N seconds of it instead, limited to the same source, the same streams or everything (scope).

Investigate

• aggregate — exact counts computed by Graylog:
  • kind="count" — number of matching messages.
  • kind="histogram" — counts over time with automatic interval, peak bucket and the onset of a spike.
  • kind="top" — top N values of the group_by field with exact counts.
  • kind="errors" — groups errors by exception, logger, source or any field: exact count, first and last seen, one sample message per group.
• trace_request — follows a correlation/request/trace id through the configured trace_fields on all streams (falls back to full text); returns a cross-service timeline, per-service steps with durations and the first error.
• compare_periods — two periods (e.g. before/after a deploy): which error groups are new, grew, disappeared or shrank, normalised per hour.

Discover

• describe — what="instances" (default: detected version and the API in use), "streams", "fields" (so the model writes queries with real names), "presets" or "scan_rules" (query, condition, severity, tags).
• run_preset — runs a named query you define in the config. A preset's tool keeps the names of the underlying operations: scan, root_cause, detect_changes, service_map, search_logs, count_logs, error_summary, log_histogram, top_values, trace_request or compare_periods.

The server also sends the model instructions about Lucene syntax and a suggested investigation flow.

Scan rules

scan answers "is anything wrong?" or "scan X" with one call. Each rule is a Lucene query plus a condition, and fires when:

What keeps false alarms out:

• Share of traffic (per_traffic = true): the rule compares matches / traffic instead of matches per hour, so errors that triple because a sale tripled the traffic stay quiet, and a rise from 1% to 10% of a busy hour fires. Traffic is every message in the scan's scope, or traffic_query (e.g. path:/checkout for checkout errors). The built-in error, 5xx, connectivity and database rules use it.
• Seasonal baseline (baseline_shift = "1d" or "7d", baseline_periods = 3): the window is compared with the same window one day (week) earlier, three times; the period with the median rate is the reference, so the morning peak is compared with yesterday's morning peak, and one bad day among the three changes nothing. Periods without any data (older than the index retention) are dropped and the finding says so (baseline_note). By default the baseline is the period right before the window, as long as the window, or baseline = "24h".
• Significance: a growth fires only if it is unlikely to be chance, by an exact conditional binomial test of the window's count against the reference period's (the standard test comparing two Poisson rates). 3 errors then 7 (x2.3) stays quiet with a note saying so; 400 then 600 (x1.5, with growth = 1.4) fires. Each growth result carries its confidence; the rule's confidence (default 0.99) is the bar.

Every rule costs two exact counts, plus one per extra baseline period and for traffic; identical counts are made once per scan (rules share the error query, the traffic) and run concurrently (limits.scan_concurrency). Groups and one sample are fetched only for rules that fired or need them to decide, and a sample shown by one finding is not repeated by the next. The result has findings (fired, most severe first), quiet (checked and normal, with counts and trend) and skipped (could not be checked, e.g. a field these logs do not have).

Built-in rules: crash and resource_exhaustion (critical, any match); error_spike (share of errors x2), new_error_types (exception unseen in 24 h), http_5xx (share x2, only where a status field exists), connectivity (timeouts, refused/reset connections, DNS; share x3) and database (deadlocks, lock timeouts, too many connections; share x3), all high; auth_failures (rate x3, at least 20), medium.

[scan]
disable = ["auth_failures"]                 # built-in rules to turn off
exclude = 'logger_name:HealthCheck OR "GET /health"'   # noise dropped from every rule

[scan.rules.connectivity]                   # a built-in: only the keys you set change
min_count = 30

[scan.rules.card_declined]
description = "Bank declines"
query = 'service:payment AND message:"declined"'
severity = "high"                           # critical | high | medium | low
growth = 2.0
min_count = 20
per_traffic = true                          # compare declines / payment traffic, not declines per hour
traffic_query = "service:payment"
baseline_shift = "7d"                       # against the same hour of the last 3 weeks (payday, weekends)
group_by = "bank_code"                      # top groups shown with the finding
tags = ["payment"]

[scan.rules.ledger_mismatch]
query = '"ledger mismatch"'                 # no condition: any match fires
severity = "critical"
instances = ["payment/prod"]                # instances, groups or environments; everywhere when omitted

[scan.rules.consumer_lag]
query = "consumer_lag:>10000"
requires = ["consumer_lag"]                 # skipped (and said so) where the field does not exist

Other keys: errors_only = true ANDs the instance's error_query, exclude drops noise for one rule, baseline sets the rule's own baseline length, baseline_periods the number of shifted periods, confidence the significance bar. A rule's own baseline wins over the call's. The config is validated at startup like the rest (unknown keys, a rule that would fire on all traffic, new_groups without group_by).

Selecting what to run: scan(rules=["payment"]) takes rule names or tags, min_severity="high" leaves out lower rules before any request, and checks=[{"name": "declined", "query": "message:declined", "growth": 2}] adds rules written for the request at hand (same keys as above). A preset with tool = "scan" saves a scan the team runs often, and the scan MCP prompt (/mcp__graylog__scan in Claude Code) runs scan, verifies each finding and reports them as a table.

Writing rules that are fast and accurate:

• Match on fields when you have them (http_status:[500 TO 599], exception_class:...); otherwise quoted phrases ("Connection refused"). Avoid leading wildcards and regex: they scan every term in the index.
• Pick the condition from how often the event normally happens: never, then threshold = 0; always a little, then growth (the significance test handles noise; min_count only sets the smallest count worth a look).
• Anything that grows with traffic gets per_traffic = true; traffic with a daily or weekly curve gets baseline_shift.
• Drop known noise with exclude (per rule or under [scan]) rather than raising thresholds.
• Use a longer baseline (baseline = "24h") for rules on spiky or low-volume traffic, and requires for rules that depend on a field only some systems log.
• Give each rule a severity that matches who should be woken up, and tags that match how people ask ("payment", "security", "dependencies").

Searches that find nothing

search_logs and aggregate with kind="count" do not stop at "no matches". Four decisions, each explained in the result:

1. Safe rewrite, before the search runs, only for a query that cannot match as written: a field this Graylog does not have but with one clear counterpart (Service:payment -> service:payment; the instance's configured name for it, exception: -> exception_class: from group_fields; the one field ending in that word, status:500 -> http_status:500; or an alias learned below), or a level word on a numeric level field (level:ERROR -> level:3). The result carries rewritten: {from, ran, why}.

2. Counted suggestions, after a search found nothing: a misspelled field (close names and known synonyms such as app -> service), the level in the other form, another letter case (keyword fields are case-sensitive), a prefix wildcard (OrderNotFound -> OrderNotFound*), lower-case and / or (Lucene searches them as words), a bare id (req-7f3a9c, a UUID) also looked for in the trace_fields, outside the repository's focus, a wider range (with last_seen, the time of the newest match). Each is counted exactly by Graylog (in parallel, bounded); the ones that find something come back in suggestions with their count and reason, best first, and the hint names the one to run.

3. Learning from use. A suggestion counts as offered; it counts as used when the model then runs it and finds something. A fix's weight is the mean of Beta(1 + used, 1 + offered - used), so it needs evidence and loses weight when it is ignored. A fix used at least twice with weight >= 0.75 applies by itself on that instance: a field alias becomes a safe rewrite, and a case or prefix fix reruns the empty search. Focus and range are only ever suggested, since they change the question.

4. Diagnosis, after a search found nothing, in diagnosis: the counts of each top-level AND clause, the ones that match nothing on their own (or that each matches but not together), the values the field of such a clause really has in the range with the close ones (service:paymnt -> did you mean payment; numbers are never guessed), and the dead clauses that match on their own further back.

Learned fixes live in ~/.config/graylog-mcp/learned.json, per instance; the admin page (Activity > What graylog-mcp learned) shows each one with its counts, weight and whether it applies by itself, and forgets one on request. GRAYLOG_MCP_LEARN=off stops learning and applying learned fixes; the deterministic fixes stay.

Ask in plain language

For people who do not write Lucene or know field names. ask takes a question, works out what it means, finds the error it describes and follows that error's trace across services:

ask("Why did checkout fail with a bank connection error on prod this morning?")
ask("Thanh toán bị lỗi timeout trên prod trong 1 giờ qua, đi qua những service nào?")
ask("How many errors were there yesterday?")

It is off by default. Bring your own key: run graylog-mcp login --ask (it asks for the key without echoing it, tests it with one request and saves it in ~/.config/graylog-mcp/secrets.toml), or set TYPESAFE_API_KEY in the environment, then restart the server. Without a key the tool is not even listed and nothing leaves the machine. --model NAME and --base-url https://... on the same command save your choice of model and endpoint (a key for another account or a proxy of your own); the environment variables GRAYLOG_MCP_ASK_MODEL and GRAYLOG_MCP_ASK_BASE_URL do the same, and [ask] model in a config file wins over both. The endpoint must speak TypeSafe's System One API (POST /v1/systemone). graylog-mcp logout --ask forgets the saved key, model and endpoint. The default model is TypeSafe's Jev, reached over HTTPS at https://api.typesafe.ai. It only chooses: the answer is built from the existing tools and fixed sentences, and starts with a line such as "Understood: prod, last hour, errors only" so the asker can correct it.

What is sent to TypeSafe: the question, the names and descriptions of your configured instances, and up to max_candidates (30) error messages, each redacted like every other output and cut to send_chars (250) characters. Never raw logs, field values the redactor masks, or anything from the Graylog token. Graylog itself is still only read (GET and the read-only searches).

When the question is too vague ("Is the system ok?") ask returns a clarify text instead of guessing; when nothing in the period matches it says so with the best score; when two errors match about equally it lists them. An error without a trace id gets a hint instead of a flow. cost in the result counts the requests made.

[ask]
# api_key_env = "TYPESAFE_API_KEY"   # name of the variable, never the key
# language = "en"                    # "en" or "vi": the language of the summary text
# min_confidence = 0.6               # below this, ask asks back
# relevance = 0.5                    # an error group scoring lower is "not what you described"
# code_patterns = ['ORD-\d+']        # lets "what happened to order ORD-123?" find the code
# model, base_url, max_candidates, send_chars

Limits: the error groups come from the newest 500 error messages of the period, and the traces of one error are looked up in one query of at most 500 messages. The error query and trace and service fields are the instance's (error_query, trace_fields, service_fields). A repository's own .graylog-mcp.toml may set the other [ask] keys but not base_url or api_key_env, which decide where the key and the error text go: set those in your user or a central config.

Root cause analysis

Example verdict, from the scenario in the integration suite (a bad deploy of payment), identical on Graylog 4.3, 5.0, 5.2, 6.1 and 7.0:

Most likely origin: payment (confidence high). errors began at 08:42:49.756, 182.5s after payment deployed 1.3.9 -> 1.4.0 (hosts pay-1, pay-2 -> pay-3, pay-4). Also affected: gateway errors (+50ms), bank-adapter traffic drop.

How it gets there:

1. Signals per service. Two pivots (errors, and traffic with average latency) split by service over the incident window plus a baseline window right before it.
2. Onsets. For each service, the first interval that is clearly abnormal against its own baseline (median and MAD, so one noisy minute in the baseline does not hide anything) and stays abnormal: error rise, traffic drop or spike, latency rise. Error onsets are then refined to the exact first error message.
3. Changes. Version fields, host rollouts and start/stop lines (see view="changes") in the window and the 24 hours before it. A service that merely went silent is reported as a traffic drop, not a change.
4. Call graph. Traces sampled per service (and from failing requests) give caller -> callee edges.
5. Ranking. Earliest onset (exact timestamps break ties inside the first interval), a change shortly before the onset, callers failing after it, and a dependency failing before it (which points further down) all move the score. The result lists the reasons for each candidate, a merged timeline, the first error message with its ref, and the next calls to verify the hypothesis.

It is a ranked hypothesis with its evidence, not a certainty. Field names come from config: service_fields, trace_fields, version_fields, latency_fields, change_query.

Version detection

At startup the server reads GET /api/system (falling back to GET /api/ when the token may not read system info), picks the APIs below and caches the result; describe shows the choice. An API that answers 404/405/501 is dropped and the next one is used.

Pivot requests use "field": "x" on 4.x and "fields": [...] from 5.0. Histograms always use a views pivot with a time grouping. Query syntax errors and zero-result queries are checked with POST /search/validate so the model gets the actual problem ("incomplete query", "unknown field: sevrity") instead of a bare "all shards failed". Graylog 7 ships its own MCP endpoint; this server is still useful there for masking and compact output.

Every tool, each API path forced on its own (universal, views, Scripting API), the read-only guarantee and the no-leak check pass against real containers of:

Behaviour observed on real servers and handled: 5.x+ pivots return an (Empty Value) bucket for documents without the grouped field (reported separately as without_field); the Scripting API returns "-" for missing fields and no message index; invalid queries come back as HTTP 500 from universal search on 4.x-5.x; universal search is still present in 7.0.

Output shaping

• Masking (before anything else):

  • always on: sensitive field names (password, token, secret, authorization, cookie, ...), emails, Bearer/Basic credentials, JWTs, key=value / "key": "value" secrets, credentials in URLs, PEM private keys, card numbers that pass the Luhn check;
  • packs enabled in config: vn (phone numbers, 12-digit CCCD, optional 9-digit CMND), us (SSN), eu (IBAN with checksum), uk (NINO), in (Aadhaar, PAN);
  • a ready-made profile (below), then your own packs, sensitive_fields and regexes on top, plus an allow-list and excluded field names to avoid false positives.

  profile	Adds to the core rules
  basic (default)	nothing
  pii	field names for personal names, phone, birth date, home/billing/shipping address, street
  finance	field names for account, card, CVV, IBAN, routing and tax ids; the eu pack (IBAN)
  strict	pii + finance + every country pack + IPv4 addresses (noisy: 1.2.3.4 version numbers are masked)

  [redaction]
  profile = "pii"
  packs = ["vn"]                    # added to the profile's packs
  sensitive_fields = ["order_note"] # added to the profile's field names
  exclude_fields = ["phone"]        # always wins over the profile
  

  Grouping keys (aggregate with kind="top" or "errors") are masked too.

• Stack traces (Java/Kotlin, Python, .NET, Go, Node): exception lines, Caused by blocks and frames from app_packages are kept; the rest becomes … N frames. Without app_packages, the first N frames are kept (the last N for Python).

• Repeated lines: numbers, UUIDs, hex, IPs and timestamps are normalised and identical templates grouped with count, first, last and one sample.

• Size: per-value and per-call character caps, with truncated and next_offset.

• Time: shown in the configured timezone with its offset. Each line carries ref: index/id for follow-up calls.

• Errors: clear messages for 401, 403, query syntax errors (with position), timeouts, unknown streams (with suggestions) and unsupported versions.

• Load on Graylog: each instance takes at most max_concurrency requests at a time (default 16), shared by every tool call and, with the shared server, every session. Connection errors, 429 and 502/503/504 are retried retries times (default 2) with exponential backoff, honouring Retry-After.

Configuration

Environment variables:

A TOML file (--config or GRAYLOG_MCP_CONFIG, else .graylog-mcp.toml in the current directory or a parent up to the repository root, else ~/.config/graylog-mcp/config.toml) covers the rest: several instances, token or basic auth, TLS verification and CA bundle, proxy, timezone, trace_fields, error_query (because level is a syslog number or a string depending on how logs are shipped), app_packages, redaction packs and custom patterns, presets and limits. Secrets are never written in the file; it names the environment variable that holds them (token_env, password_env). The file is validated at startup and any mistake (unknown key, bad regex, unknown timezone, missing secret) stops the server with a precise message.

See examples/config.toml for every option.

One repository, several environments

Commit a .graylog-mcp.toml at the root of your application repository with one instance per environment. The server finds it from the current directory or any parent up to the repository root, so everyone working in the repo gets the same setup. Tokens stay out of the file: each instance names its environment variable.

Only use this with repositories you trust. The file is applied as it is, with no approval step, and with the shared server, opening a client in the repository is enough. Its url, proxy, verify_tls and *_env settings decide where a saved secret is sent and which one, include can pull in other files, [redaction] can switch masking off, repos decides which folders get a .mcp.json, and [http] can open the server to the network. Read a file you did not write before you run start, init or install there, or open a client in that repository.

# .graylog-mcp.toml
default_instance = "staging"          # what tools use when no environment is named
timezone = "Asia/Ho_Chi_Minh"

[redaction]
packs = ["vn"]

[investigation]
trace_fields = ["traceId", "X-Request-ID"]

[instances.dev]
url = "https://graylog-dev.example.com"
token_env = "GRAYLOG_DEV_TOKEN"
description = "Development cluster, noisy, debug logs on"

[instances.staging]
url = "https://graylog-staging.example.com"
token_env = "GRAYLOG_STAGING_TOKEN"
description = "Staging, deployed on every merge to main"

[instances.prod]
url = "https://graylog.example.com"
token_env = "GRAYLOG_PROD_TOKEN"
description = "Production"
error_query = "level:<=3 AND NOT logger_name:healthcheck"   # any key can differ per environment

Register the server once for the project (Claude Code reads .mcp.json at the repo root and expands ${VAR} from each developer's environment):

{
  "mcpServers": {
    "graylog": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/ntbang0901/graylog-mcp", "graylog-mcp"],
      "env": {
        "GRAYLOG_DEV_TOKEN": "${GRAYLOG_DEV_TOKEN}",
        "GRAYLOG_STAGING_TOKEN": "${GRAYLOG_STAGING_TOKEN}",
        "GRAYLOG_PROD_TOKEN": "${GRAYLOG_PROD_TOKEN}"
      }
    }
  }
}

Then ask in plain words: "why is checkout failing on staging?", "compare errors on prod before and after 14:00". Every tool takes instance, describe shows each environment with its description and detected Graylog version (environments may run different versions), and results always name the instance they come from. A developer who has no token for an environment simply gets a clear error for that one; the others keep working.

Clients that do not start the server inside the repository (Claude Desktop) need the path explicitly: GRAYLOG_MCP_CONFIG=/path/to/repo/.graylog-mcp.toml.

Setup helpers and admin UI

graylog-mcp ui opens a page on 127.0.0.1 with:

• Dashboard: the setup steps (one Finish setup button), log volume and errors over the last 24 hours or 7 days (read from Graylog when the page opens, nothing stored), every environment's health with a sparkline or its problem and a button to fix it (Add token, Fix), the tools the model used, and the searches that found nothing;
• Activity: every tool call the model made, per hour: found something / found nothing / error, per-tool counts and latency (median, p95), and the recent calls with what was asked; Re-run opens one in the Playground;
• Environments: one card per group with its environments, default environment, repositories and their focus; descriptions and defaults save as you change them. The environment form tests the connection and saves in one click (Save anyway if Graylog cannot be reached), names the environment from the URL (graylog-uat.acme.com gives uat) and saves the token or password to your per-user secrets file;
• Log fields: settings per scope, with detection from your logs in the same page: suggestions next to the value in effect, applied to the scope you are editing;
• Redaction: toggle country packs, add custom patterns and allow-list entries, and see live which rules mask your sample text;
• Playground: run any tool against any environment and see the exact output, its size and an approximate token count;
• Other clients: snippets and one-click install for Claude Desktop, Cursor, VS Code and Claude Code per project;
• Config file: edit the TOML with validation and an automatic backup.

Tool calls are recorded by the MCP server (stdio or shared) in ~/.config/graylog-mcp/usage.jsonl: tool, environment, duration, number of results, size, and the query or id asked for, redacted with your redaction rules before it is written. It stays on your machine, is trimmed past 4 MB, and GRAYLOG_MCP_USAGE=off in the server's environment turns it off.

Ctrl/⌘+K jumps to any page, environment, tool or action; unsaved changes show a save bar (Ctrl/⌘+S) and are not lost by leaving the page. Light or dark follows the OS until you pick one (the Follow OS theme checkbox goes back to it), with a phone layout; slate greys, teal for actions and data, amber for warnings, coral for failures.

The UI listens on loopback only, checks the Host header, and every API call needs the random token in the link printed at startup. It writes the config file and client configs when you ask. A token or password entered in the form is saved, when you tick "Save it on this machine", to the same per-user secrets file as graylog-mcp login, never to the config file.

*_env settings hold the name of a variable (GRAYLOG_PROD_TOKEN), never the secret: a value that is not a valid variable name is refused, and never echoed back, since it is most likely a secret typed into the wrong field.

Groups and your own environments

Large organisations often run one Graylog per system and per environment: ERP, CXP, PAYMENT... each with its own dev/uat/prod (or sandbox, dr, prod-eu...). Environment names are yours; nothing assumes dev/staging/prod, and each group can have a different set.

# graylog-org.toml: one file for the whole company, e.g. in a shared platform repository
timezone = "Asia/Ho_Chi_Minh"
default_environment = "uat"            # used when a question names only the system

[environments.prod]                    # declare environments once; every group inherits these settings
description = "Production"
error_query = "level:<=2"
ca_bundle = "/etc/ssl/corp-ca.pem"
[environments.uat]
description = "User acceptance"

[groups.erp]
description = "ERP"
trace_fields = ["correlationId"]       # anything set on a group applies to all its environments
[groups.erp.environments.uat]
url = "https://graylog-erp-uat.corp"
token_env = "GRAYLOG_ERP_UAT_TOKEN"
[groups.erp.environments.prod]
url = "https://graylog-erp.corp"
token_env = "GRAYLOG_ERP_PROD_TOKEN"

[groups.payment]
description = "Payment platform"
default_environment = "sandbox"
[groups.payment.environments.sandbox]
url = "https://graylog-pay-sbx.corp"
token_env = "GRAYLOG_PAYMENT_SANDBOX_TOKEN"
[groups.payment.environments.prod]
url = "https://graylog-pay.corp"
token_env = "GRAYLOG_PAYMENT_PROD_TOKEN"

Settings are layered: [environments.<env>] < [groups.<group>] < [groups.<group>.environments.<env>]. Each pair becomes an instance named <group>/<environment>.

A service repository then only says which group it belongs to:

# payment-api/.graylog-mcp.toml
include = "../platform/graylog-org.toml"   # relative to this file; a list is allowed
default_group = "payment"
only_groups = "payment"                    # load only this group here (a list, or "*" for every group)

How the model (and you) pick an instance, in every tool's instance argument:

describe returns the groups with their environments and descriptions, graylog-mcp init asks for groups first, and the admin UI shows environments by group (instances from an included file are marked and edited in that file).

Repositories of a group

List the repositories each group serves, as local folders (paths relative to the config file, ~ allowed) or git remotes (git@gitlab.corp:acme/payment-api.git, acme/payment-api or just payment-api):

[groups.payment]
repos = ["~/code/payment-api", "../payment-worker", "acme/payment-gateway"]

When the server runs inside one of them (a folder or any subfolder, or a clone whose origin matches), it loads only that group: "errors on prod" means that system's production, and the other groups' Graylog servers are not reachable from that repository. Asking for one returns an error that says why and how to enable it. describe shows the repository and the scope; graylog-mcp doctor shows which repository was recognised.

To reach more groups from a repository, set only_groups = ["payment", "erp"] (or "*" for all) in its .graylog-mcp.toml, or GRAYLOG_MCP_GROUPS=payment,erp in the MCP client's environment. only_groups also works without repositories, e.g. in a personal config. The admin UI always shows every group.

Manage them in the admin UI (Environments, in each group's card: add a local folder, see its git remote and whether it is set up, "Set up" writes .graylog-mcp.toml with an include of the shared file and only_groups, so the scope holds on every machine whatever its folder layout, and registers Claude Code there) or from the command line:

graylog-mcp repo --config ../platform/graylog-org.toml add payment   # current folder; also sets it up
graylog-mcp repo list
graylog-mcp repo remove payment ~/code/payment-api

Focus: search this repository's service by default

Inside a repository, "errors in the last hour" means that service's errors. search_logs, aggregate, compare_periods and root_cause with view="changes" add a filter on the service field and say so in their result (focus). Other services and streams are searched only when asked for: streams=["*"] (everything), explicit streams, or a query that names a service field (application:"other-service"). trace_request and the other root_cause views always span every service.

Each repository sets its focus in its own .graylog-mcp.toml:

[focus]
service = "catalog-mdm-service"   # value of the service field; a list for several; false = no service filter
streams = ["MDM"]               # optional: streams searched by default
# field = "application"         # optional: the service field (default: the first of service_fields present)

Without [focus], the service is guessed from the repository name (folder or git remote) by matching it against the values of the service fields (service, application, ...) over the last 24 hours: catalog-mdm finds catalog-mdm-service. No clear match, no filter; describe shows what was picked. Set it from the admin UI (Environments > the repository's Focus), with graylog-mcp repo focus catalog-mdm-service --streams MDM inside the repository, or for one MCP client with GRAYLOG_MCP_SERVICE=<name> (off disables it).

Settings by scope

The admin UI's Log fields page edits field names, queries and connection options for a chosen scope: global, one environment (for every group), one group, or a single instance. Each field shows the value currently in effect; leaving it empty inherits it.

One process for every session

With stdio, the client starts a server process for every session: ten Claude Code sessions are ten Python processes, each with its own connections and version detection. graylog-mcp start runs one server for all of them instead, and still answers each session with the configuration of the repository it works in (its .graylog-mcp.toml, group, focus), as if it had been started there.

What start does, so you know what is on your machine:

• A stable copy: uv tool install (in ~/.local/share/uv/tools), not uvx's temporary cache. update reinstalls it from the same source; the version shown everywhere includes the git commit.

• The server: graylog-mcp serve --shared --admin --port 18742 --config <file> on 127.0.0.1 (the port is fixed so no client configuration goes stale: if another program uses it, start stops with an error; --port picks another one, later commands follow it), started at login and started again if it crashes:

  	Started at login by	Restarted after a crash by
  macOS	a LaunchAgent (~/Library/LaunchAgents/io.github.ntbang0901.graylog-mcp.plist)	launchd
  Linux with systemd	a user unit (~/.config/systemd/user/graylog-mcp.service)	systemd
  Linux desktop without systemd	XDG autostart (~/.config/autostart/graylog-mcp.desktop)	graylog-mcp keepalive
  Windows	a Task Scheduler task graylog-mcp, no console window	graylog-mcp keepalive

  Elsewhere (WSL or a server without systemd, no desktop session) it runs until you log out; on WSL, turning on systemd ([boot] systemd=true in /etc/wsl.conf) is enough. Log: ~/Library/Logs/graylog-mcp.log or ~/.config/graylog-mcp/server.log.

• The admin page at http://127.0.0.1:18742/admin/, with the access token kept in ~/.config/graylog-mcp/admin-token (owner only); graylog-mcp ui opens it.

• Claude Code, connected repository by repository: start writes the shared server into the .mcp.json of each repository of your groups, creating the file when it has none (the other servers in it stay, the previous file is kept as .mcp.json.bak): {"mcpServers": {"graylog": {"type": "http", "url": "http://127.0.0.1:18742/mcp", "headers": {"X-Graylog-MCP-Repo": "${PWD:-}"}}}}. Claude Code fills in the folder it runs in; the server walks up to the repository, and asks you once to approve a project's server. A repository whose .mcp.json still starts graylog-mcp itself gets a private override (Claude Code's local scope) instead: the file stays as it is for your teammates. graylog-mcp start --global registers it once for every project (user scope) instead. Without it, start removes a registration for every project that an earlier version made, so that Claude Code does not report the server as defined in several scopes with different endpoints.

• A different port: whenever the server starts (by start, at login, after update), what an earlier port registered follows it: Claude Code's registrations and overrides, the URL of a .mcp.json entry on another local port (rewritten in place, .bak kept) and the Cursor and VS Code entries of your groups' repositories. status lists what still points elsewhere.

How the server picks the configuration of a session:

• The repository's own .graylog-mcp.toml; without one, the --config file, with the group matched from its repos. start keeps one such file wherever you run it: the one given to start --config last time, else ~/.config/graylog-mcp/config.toml. graylog-mcp ui opened in a folder without a config edits that file too.
• Edits (admin page, login, repo focus) apply within seconds, without a restart.
• One Graylog client per instance, shared by every repository using it: one connection pool, one version detection, one stream and field cache.
• Tokens come from the server's environment or graylog-mcp login (the admin page saves them there too).
• Other clients: Cursor and VS Code send ${workspaceFolder} (graylog-mcp install cursor --shared), any client can add ?repo=<folder> to the URL. Claude Desktop already runs one server for all its chats.

Shared HTTP server and Docker

stdio is the default. For one server shared by a team, use streamable HTTP with its own bearer token:

GRAYLOG_MCP_HTTP_TOKEN=$(openssl rand -hex 32) graylog-mcp --transport streamable-http --host 0.0.0.0 --port 8000

The server refuses to listen on a non-loopback address without a token (unless --allow-no-auth is given, e.g. behind an authenticating proxy). /healthz is open; the MCP endpoint is /mcp.

docker build -t graylog-mcp .
docker run -p 8000:8000 \
  -e GRAYLOG_URL=https://graylog.example.com -e GRAYLOG_TOKEN=... \
  -e GRAYLOG_MCP_HTTP_TOKEN=... graylog-mcp

A server for the team on a VM

One serve --shared process on a VM can serve every developer and every repository. The VM cannot read the developers' folders, so a client names its repository by identity: its git remote (github.com/acme/payment-api, acme/payment-api, any git URL) or, failing that, its folder name. Groups list their repositories by remote or name, and [repos."<repository>"] holds each one's settings (what a repository's own .graylog-mcp.toml holds on a developer's machine):

# /etc/graylog-mcp/org.toml on the VM, kept in a git repository and deployed by the platform team
[groups.payment]
repos = ["acme/payment-api", "acme/payment-worker"]
[groups.payment.environments.prod]
url = "https://graylog-pay.example.com"
token_env = "GRAYLOG_PAYMENT_PROD_TOKEN"

[repos."acme/payment-api"]                  # focus, default_group/environment/instance, only_groups, scan, presets
focus = { service = "payment" }

[http]
auth_token_env = "GRAYLOG_MCP_HTTP_TOKEN"   # every developer sends it
# repo_paths = false                        # default off a loopback address: folders clients name are never read

graylog-mcp serve --shared --host 0.0.0.0 --port 8000 --config /etc/graylog-mcp/org.toml

The config is read again within seconds of a change: deploy the file, no restart. The admin page writes files on the machine it runs on, so it only listens on the loopback address; reach it with an SSH tunnel (ssh -L 8000:127.0.0.1:8000 vm, then http://127.0.0.1:8000/admin/ with serve --admin).

On each developer's machine, nothing runs but Claude Code:

How a connection names its repository: in a project's .mcp.json, Claude Code runs graylog-mcp headers (its headersHelper) in the project folder when it connects; it prints the repository's git remote. Without graylog-mcp installed, the remote written by connect applies. Registered for every project, there is no helper (it would run outside the project): the server identifies the repository by the name of the folder Claude Code runs in (${PWD}), so list repositories by name in repos, or use connect --project for the ones listed by remote.

Testing

uv sync
uv run pytest              # unit + contract tests (no network), coverage >= 85%
uv run ruff check && uv run ruff format --check && uv run mypy

• Unit tests: masking (including values that must not be masked), stack traces in five languages, line grouping, time parsing, pivot building and parsing, config validation, HTTP auth.

• Contract tests: every tool against an in-memory Graylog (tests/fake_graylog.py, served through httpx.MockTransport) that reproduces the response shapes of 4.3, 5.0, 5.2, 6.1 and 7.0, with planted sensitive values that must never reach the output, an assertion that no request could change state, and a size budget per call.

• Integration tests: tests/integration/docker-compose.yml starts Graylog 4.3, 5.0, 5.2, 6.1 and 7.0, each with MongoDB and OpenSearch; seed.py ships the same GELF dataset, then every tool runs against it:

  tests/integration/run.sh v50      # or v43 v52 v61 v70
  

  CI runs the whole matrix weekly, on demand, and on pull requests that change the backends.

CI on every push: ruff, mypy, tests on Linux (Python 3.11-3.13), Windows and macOS, wheel build with metadata check and a smoke test in a clean environment, and a Docker build with an HTTP smoke test. CodeQL scans the code and workflows; Dependabot keeps dependencies and actions current. See CONTRIBUTING.md and SECURITY.md.

Project layout

src/graylog_mcp/
  config.py     environment + TOML, validated at startup (fail fast)
  client.py     async httpx client: auth, TLS, proxy, read-only guard, HTTP error mapping
  backends/     version detection and API selection: universal.py, views.py, scripting.py
  timerange.py  time parsing, display, interval selection
  redact.py     core rules, country packs, Luhn / IBAN checks
  shaping.py    field selection, truncation, stack traces, line grouping, output budget
  tools.py      tool implementations
  rca.py        root cause analysis: onsets, change detection, service map, ranking
  ask/          the optional `ask` tool: Jev client, question understanding, error groups and flow, fixed texts
  operations.py every operation in one table: what presets, the admin page and usage know by name
  toolspecs/    the tools a model sees, each declared once: its arguments and the operations it runs
  server.py     model instructions, and the MCP tools built from those declarations

License

MIT