tshark-mcp
A read-only MCP server that runs tshark -r on pcap files in allowed directories and turns them into diagnosed answers with frame numbers, aimed at both generic IP troubleshooting and mobile-core/telecom failures.
Orient on a capture:
pcap_info,capture_metadata(writer, interfaces, drops),capture_quality(can loss/timing conclusions be trusted?),incident_summary(ranked "what is wrong here" findings with next calls),capture_index(ring-buffer files in time order).IP/TCP health:
tcp_health(retransmissions split by cause, resets, windows, PMTUD black holes),tcp_connections(refused/filtered/prohibited, who closed),icmp_errors,conversations(top talkers per layer).Application protocols:
http_transactions(HTTP/1.x, HTTP/2, gRPC latency and status),dns_transactions(unanswered queries, rcodes, retries),tls_handshakes(SNI vs chosen version/cipher, alerts, certificates).VoIP/IMS:
sip_dialogs,sip_call_detail,media_path_check(one-way audio, 488, NAT/SDP mismatches),rtp_streams(loss, jitter, RTCP, MOS estimate),ims_registration,sms_over_ims.Mobile core:
diameter_transactions/_sessions/_avps,gtp_sessions/gtp_ies,trace_subscriber(one timeline across Diameter, GTP, PFCP, MAP, SIP),roaming_attach_check,compare_captures.Reference without a capture:
cause_lookup(Diameter/GTP/NAS/SIP/Q.850/HTTP cause codes with spec refs),find_fields,check_filter.Drill down and search:
filter_packetswith arbitrary fields,packet_detailfor one frame,expert_infofor malformed/undecoded packets,find_identifieracross files or a directory.Workflow helpers:
open_capturehandles reused across calls (with caching and sticky filter/window).Safety: read-only, no shell or live capture, only allowed directories, hard caps on size/packets/output/time, plus optional redaction of IMSIs, phone numbers and IPs.
Enables LLM agents to analyse existing packet captures through tshark, Wireshark's command-line packet analyser, with read-only tools for inspecting protocols such as Diameter, GTP, SIP, SCTP/M3UA/SCCP/TCAP/MAP, and tracing subscriber activity across a capture.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@tshark-mcp trace IMSI 310150123456789 in /pcaps/attach-failure.pcap"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
tshark-mcp: Wireshark analysis for AI assistants
Drop a pcap, ask your AI assistant what went wrong, get an answer with frame numbers.
tshark-mcp is a read-only MCP server that runs tshark -r on captures you allow and turns
them into answers: failed connections, slow servers, DNS and TLS errors, one-way audio, and
the telecom failures generic tools miss (Diameter, GTP, SIP/IMS, SIGTRAN, 5G).
claude mcp add tshark-mcp -- uvx --from tshark-mcp-server tshark-mcp --allow-dir /path/to/pcapsNeeds uv and tshark (comes with Wireshark). Cursor, Claude Desktop, VS Code, Docker and the Claude Code plugin: the guide.

Why this one
Answers, not packet dumps. Ask "what is wrong in this capture?" and
incident_summaryruns every check the capture's protocols allow, ranks the findings and names the next call. Every claim carries frame numbers you can open in Wireshark.Built by a mobile core engineer. It pairs requests with answers across Diameter, GTP, PFCP, SIP, RTP and MAP, names the cause codes with their 3GPP reference, and follows one subscriber across all of them.
Safe to hand to an agent. No shell, no live capture, no writes: only
tshark -ron files inside directories you allow, under hard caps on size, packets, output and time. Optional redaction hides IMSIs, phone numbers and IPs.Checked against public captures. 102 public sample captures, 1,064 tool calls; every wrong answer found was fixed and is re-tested (results).
Related MCP server: Wireshark-MCP
Learn more
Guide: install, example questions, tool list, prompts, safety, limits
MIT licence, Copyright (c) 2026 Karl Hulse.
Available Tools
34 toolscapture_indexCapture index (ring-buffer files in time order)ARead-onlyIdempotent
Index a directory of capture files (a ring buffer, a day of rotated files) in time
order without decoding a packet: per file the first and last packet time, duration,
packets, bytes, interfaces and link type, the hole or overlap against the previous
file, and with start/end (same forms as on every tool) covers_window: which files
to open for an incident window. summary gives the span, totals, holes (gaps longer
than min_gap_s seconds between files, default 1), overlaps, ordering (whether
name order is time order: a wrapped ring or renamed files break it) and for a window
window_files and window_covered.
More detail: resource schema://tools/capture_index.
Inputs: path (required), end, start, offset, +3 more. Returns: path, count, error, files, +16 more
capture_metadataCapture file metadata (writer, interfaces, drops)ARead-onlyIdempotent
What the capture file says about itself, read from the pcapng blocks without
decoding a packet: did the capture box drop packets, which snaplen and capture filter
were used, which tool and OS wrote the file, which interfaces are in it. file:
format, version, byte order, compression, section count. writer: the section
header's application, OS, hardware and comment. interfaces[]: id, name,
description, link type, snaplen, timestamp resolution and offset, capture filter, and
from the Interface Statistics Block (tshark has no field for it) statistics.received,
dropped_by_interface, dropped_by_os, accepted_by_filter, delivered with their
timestamps; plus packets, bytes, first/last packet, packets cut to the snaplen,
timestamps going backwards, and when the first packet came relative to interface 0.
drops_total sums the drop counters (absent when no statistics block was written).
packet_comments: count and the first five. issues in plain words (drops, a tiny
snaplen, a capture filter, no statistics block, concatenated sections). A legacy pcap
returns its header with metadata_available: false. Bounded by the file size cap and
the run timeout; strings come from the file's writer and are untrusted.
Inputs: path (required), redact. Returns: file, note, path, error, +10 more
capture_qualityCapture qualityARead-onlyIdempotent
Can loss and timing conclusions be drawn from this capture? Call it before
reading retransmissions or missing packets as network faults. It reports copies of one packet (with the
likely cause), capture drops versus network loss, truncated and oversized frames, checksum
offload, one-sided flows and timestamp problems, and verdict.loss_conclusions
(trustworthy, use with care or not trustworthy) with the reasons.
More detail: resource schema://tools/capture_quality.
Inputs: path (required), end, start, redact. Returns: last, path, error, first, +15 more
cause_lookupCause and result code lookupARead-onlyIdempotent
What a telecom cause or result code means, with the spec reference to quote in a
ticket. No capture is read. protocol is one of diameter, gtpv2, gtpv1, pfcp,
nas_emm, nas_esm, nas_5gmm, nas_5gsm, s1ap, ngap, sip, q850, isup, camel_release,
map, tcap_abort, sccp_return, sgsap, rp, cp, tp_status, http_sbi, dns, tls, icmp,
icmpv6, http, nas. code: 78, "0x4e"; s1ap/ngap
"radioNetwork:21"; icmp "3:13"; http_sbi a cause ("USER_NOT_FOUND"). For Diameter,
application (s6a, s13, gx, gy, rx, cx, sh, swx, s6b, s6c, sgd, t6a, or base)
picks the Experimental-Result-Code meaning: 5001 or 5004 differ between the base
Result-Code and an application. message (a SIP method such as REGISTER or INVITE)
adds the note for that message. Returns name (from tshark's own value strings,
name_source says so, otherwise the server's table), meaning, ref,
typical_causes, related codes and check_next (which tool to run on the
capture); explained: false means only the name is known. Give query instead of
code to list the codes whose name or meaning contains every word ("apn",
"roaming").
Inputs: protocol (required), code, limit, query, +7 more. Returns: avp, ref, also, code, +33 more
check_filterCheck a display filterARead-onlyIdempotent
Check a Wireshark display filter without reading any capture: tshark compiles it
against an empty capture and its own error comes back in errors, with error_at
giving the offset and length of the part it points at (e.g. "nosuch.field" is not a valid protocol or protocol field, "0x100000000" too big for this field). When
dftest is installed beside tshark, its warnings (deprecated names such as "ssl" or
"bootp") come back in filter_warnings; warnings_checked says whether it ran.
fields lists the fields and protocols the filter names, with their types. The
server's own lint (no ;, backquote or $ outside quotes, ||/&& rather than a
single |/&, no control characters) is applied first. Use it before passing a
filter to another tool.
Inputs: filter (required), redact. Returns: error, valid, errors, fields, +6 more
compare_capturesCompare two capturesARead-onlyIdempotent
Compare a working and a failing capture of the same procedure. Builds the message
ladder of each (the trace_subscriber timeline for imsi/msisdn; give imsi_b/
msisdn_b when capture B is a different subscriber; without any identity, every
Diameter, GTPv2-C, MAP and SIP transaction of the file), drops keepalives (CER, DWR,
Echo), merges consecutive identical retries into one step with repeat, aligns the
two ladders by protocol and message, and reports first_divergence (the first step
that is missing, replaced or different), every aligned step with its op (same,
changed, replaced, only_a, only_b) and, for changed steps, the differences: result,
pairing, the key Diameter AVPs (flags, Cancellation-Type, subscription APNs, ...) and
the GTP IEs (APN, RAT, cause, F-TEID interfaces, PAA allocated); identities are
ignored. timing lists the largest round-trip and offset gaps. Both captures share
one time budget. limit caps the ladder rows returned (default 100).
Inputs: path_a (required), path_b (required), end, imsi, +11 more. Returns: a, b, count, error, +17 more
conversationsConversations (top talkers)ARead-onlyIdempotent
List conversations (top talkers) at one layer: who talked to whom and how much.
Use it to find the busy or unexpected peers; for whether TCP connections succeeded use
tcp_connections, for media streams rtp_streams. layer is one of ip, ipv6, tcp,
udp, sctp or eth. Each row has both endpoints, frames and bytes in each direction,
totals, relative start time and duration. filter is an optional display filter
that restricts which packets are counted, e.g. gtp or sctp.port == 3868.
Sorted by total frames, capped at limit.
Inputs: path (required), end, layer, limit, +7 more. Returns: count, error, layer, offset, +14 more
diameter_avpsDiameter AVP detailARead-onlyIdempotent
The AVPs of Diameter requests and answers, paired as in diameter_transactions.
Select with session_id, imsi (User-Name digits), command_code (316 ULR, 317 CLR,
318 AIR, 319 IDR, 321 PUR, 272 CCR, ...) and/or frame (a request or answer frame
number); the whole session is fetched so answers are kept. detail="key" (default)
gives the AVPs read first: Result-Code and Experimental-Result, ULR/ULA/CLR/PUR/PUA/
IDR/NOR flags with bit names, Cancellation-Type, RAT-Type, Visited-PLMN-Id, AIR
vectors, Gx CC-Request-Type, Event-Trigger, rule reports and QoS, and a digest of
Subscription-Data (MSISDN, status, ODB, access restrictions, AMBR, one line per APN
with PDN type, QCI, ARP, AMBR and PGW), plus a per-application digest under gy,
rx, cx, sh or s13 whose failed codes carry cause_lookup's explained.
detail="full" gives every AVP as a nested tree, at most max_avps leaf AVPs per
message (server cap 2000); avps_dropped counts the rest. Authentication material is
never returned. limit caps transactions (default 20). filter is ANDed with diameter.
More detail: resource schema://tools/diameter_avps.
Inputs: path (required), end, imsi, frame, +12 more. Returns: count, error, detail, offset, +12 more
diameter_sessionsDiameter sessionsARead-onlyIdempotent
Group Diameter traffic by Session-Id. Each session gives the application, origin
and destination hosts and realms, User-Name (IMSI), first/last timestamp and duration,
the request and answer command codes seen, every Result-Code or Experimental-Result-Code,
counts of unanswered and failed transactions (with unanswered_near_capture_end for
those the capture stopped too soon after to tell, as in diameter_transactions), the
worst round-trip, and the list of transactions (request paired with its answer by
Hop-by-Hop id, with RTT). Narrow with
session_id (exact), imsi (User-Name digits) or command_code (e.g. 316 for ULR):
matching messages select the sessions, then every message of those sessions is
fetched so answers are not lost. filter is an extra display filter ANDed with
diameter. Messages without a Session-Id (DWR, CER) are counted, not listed.
Inputs: path (required), end, imsi, limit, +9 more. Returns: count, error, offset, window, +12 more
diameter_transactionsDiameter transactionsARead-onlyIdempotent
Pair Diameter requests with their answers (Hop-by-Hop + End-to-End id, checked
against Session-Id). Use it for which requests failed or went unanswered and how slow
the answers were; group them per Session-Id with diameter_sessions, read the AVPs
behind a transaction with diameter_avps, follow one subscriber across protocols with
trace_subscriber. Each transaction gives command code and name (ULR/ULA, AIR/AIA,
CCR/CCA, ...), application, origin/destination host and realm, User-Name (IMSI),
Result-Code or Experimental-Result-Code with its name, round-trip ms and
unanswered: true for requests with no answer. An unanswered request whose capture
stops before its answer was due (the last packet on its interface comes before the
slowest answer to the same command in the capture plus 0.2 s, at most 2 s) also gets
near_capture_end and a note; summary.unanswered_near_capture_end counts those.
filter is an optional display filter ANDed with diameter, e.g. diameter.cmd.code == 316.
Inputs: path (required), end, limit, start, +6 more. Returns: count, error, offset, window, +11 more
dns_transactionsDNS transactionsARead-onlyIdempotent
DNS queries paired with their responses (by transport, client and server address
and port, DNS id and question name), including the queries that never got one,
which tshark's own dns.time cannot show. Each transaction gives name, type (A,
AAAA, NAPTR, SRV, HTTPS, ...), client, server, rcode and its name (NOERROR,
NXDOMAIN, SERVFAIL, REFUSED, ...), answers and addresses, round-trip ms,
unanswered (with near_capture_end and a note when the capture on its interface
stops before the answer was due: the resolver's slowest answer plus 0.2 s, at most 2 s;
counted in summary.unanswered_near_capture_end),
retransmissions (the same query resent with the same id and port), retry_of
(a new query for a name whose last query was unanswered or failed),
truncated_tc and tcp_fallback_frame (a TC answer followed by the same query
over TCP). servers gives per resolver queries, answered, unanswered share,
rcodes and latency min/p50/p90/p99/max; names lists the names that failed, were
retried or were slow (over 1 s); issues says it in plain words. Copies of one
packet (a capture on several interfaces or GTP legs) are counted once. DNS inside
GTP-U tunnels is read from the inner packet. name keeps that name and the names
under it (case-insensitive); filter is ANDed with dns. Transactions with a
problem come first, then the rest in capture order.
Inputs: path (required), end, name, limit, +7 more. Returns: count, error, names, issues, +14 more
expert_infoExpert info (warnings, malformed packets)ARead-onlyIdempotent
tshark's own expert information (-z expert): the warnings, errors and
malformed-packet flags its dissectors raised while decoding, grouped by severity,
group (Malformed, Undecoded, Sequence, Protocol, Checksum, ...), protocol and summary,
with the count, the first frame and up to five sample frames of each (an item tshark
reports only to its statistics, such as TCP "exceeds MSS", has no frames). severity is
the lowest level kept: error, warn (default), note or chat. filter is a display
filter that restricts which packets are examined, e.g. diameter or
sctp.port == 3868. summary gives counts by severity and protocol and the number
of malformed frames. Use it to catch decode problems (a malformed Diameter AVP, an
unknown vendor, a GTP IE tshark could not parse) before trusting the protocol
tools, and for the TCP sequence problems tshark already flags.
Inputs: path (required), end, limit, start, +7 more. Returns: count, error, items, offset, +13 more
filter_packetsFilter packetsARead-onlyIdempotent
Return packets matching a Wireshark display filter as rows of fields. Default
fields: frame.number, frame.time_epoch, ip.src, ip.dst, _ws.col.protocol, _ws.col.info.
Pass fields (tshark field names such as "diameter.cmd.code") to choose your own;
unknown names are rejected. Multiple occurrences of a field are joined with commas.
Use this for anything the dedicated protocol tools do not cover.
Inputs: path (required), display_filter (required), end, limit, +7 more. Returns: count, error, fields, offset, +12 more
find_fieldsFind tshark fieldsARead-onlyIdempotent
Search the fields and protocols this tshark knows (tshark -G fields) to build a
display filter or a fields list for filter_packets. Every word of query must
appear in the field's abbreviation, name or description (case-insensitive), e.g.
"result code", "teid", "Session-Id" or "retransmission". protocol keeps only the
fields of one protocol (its abbreviation, e.g. "diameter", "gtpv2", "tcp"). Rows give
the field abbreviation, its name, type (FT_UINT32, FT_STRING, ... or "protocol"), its
protocol and description; the exact abbreviation comes first, then abbreviations
that start with or contain the query, then names, then descriptions. No capture is
read.
Inputs: query (required), limit, offset, output, +4 more. Returns: count, error, query, fields, +12 more
find_identifierFind identifier across capturesARead-onlyIdempotent
Find which capture files mention an identifier. path is one capture or a
directory inside the allowed roots (its .pcap/.pcapng/.cap files are scanned, not
subdirectories). identifier is an IMSI, MSISDN, IPv4/IPv6 address, SIP Call-ID or
Diameter Session-Id; kind is auto, imsi, msisdn, ip, call_id or session_id (auto
searches digits as both IMSI and MSISDN). Returns one row per file with hits: hit
count, first and last timestamp and a per-protocol breakdown, sorted by hits. At most
max_files files are opened (capped by the server), each under the usual packet,
size and time limits, and the whole call shares one time budget; files that could
not be read are listed under errors. Use it to pick the file before pcap_info.
Inputs: path (required), identifier (required), end, kind, +4 more. Returns: kind, note, count, error, +23 more
gtp_iesGTPv2-C information elementsARead-onlyIdempotent
The information elements of paired GTPv2-C requests and responses: the GTP
counterpart of diameter_avps. Select with imsi, teid (header or F-TEID, decimal or
0x hex), frame (a request or response frame) and/or message (type number or name,
e.g. "Create Session Request"; a request name also selects its response); both halves
of each transaction are kept. detail="key" (default) gives request_ies and
response_ies with the IEs read first: IMSI, MSISDN, MEI, APN, selection mode, PDN
type, PAA, APN restriction, serving network, RAT, ULI (TAI, ECGI, CGI, SAI, RAI with
MCC-MNC), UE time zone, indication flags set, charging characteristics and ID,
recovery, APN-AMBR, F-TEIDs by interface, PCO option names (never PAP/CHAP contents),
bearer contexts (EBI, QCI, ARP, MBR/GBR, F-TEIDs, cause), the cause with its
offending_ie and source (CS bit), and the RAN/NAS cause a Delete Session carries;
other IEs are listed by name. A rejected or partly accepted response adds
likely_causes built from the request (cause 78 and the APN sent, 73 with IPv4v6
requested, 92 with PAP in the PCO, 64 and the Recovery IE, ...). detail="full" gives
every IE as a tree, at most max_ies IEs per message (server cap 2000).
filter is ANDed with gtpv2; limit caps transactions (default 20).
Inputs: path (required), end, imsi, teid, +12 more. Returns: count, error, detail, offset, +12 more
gtp_sessionsGTPv2-C sessionsARead-onlyIdempotent
Pair GTPv2-C requests with responses by sequence number and peer pair. A response
sent by the request's target towards another address (a NAT or proxy boundary) is
paired too and its path is given as response_path. Reports
message names (Create Session, Modify Bearer, Delete Session, Echo, ...), IMSI, MSISDN,
APN, F-TEIDs with interface names, cause code and name, allocated UE address (PAA),
round-trip ms and unanswered: true. Narrow with imsi (digits), apn (case-insensitive; a network
identifier such as internet also matches internet.mnc001.mcc001.gprs) or teid (decimal or 0x hex; matches the header TEID or any F-TEID):
the matching requests select sequence numbers, then both halves of each transaction
are fetched so responses are kept. filter is a raw display filter ANDed with gtpv2,
e.g. gtpv2.message_type == 32; it applies literally to each packet.
More detail: resource schema://tools/gtp_sessions.
Inputs: path (required), apn, end, imsi, +9 more. Returns: count, error, offset, window, +12 more
http_transactionsHTTP transactionsARead-onlyIdempotent
Which HTTP endpoints are slow or failing, what status codes came back, how
big were the responses, what did gRPC return? HTTP/1.x requests paired with their
responses by tshark (pipelining included) and HTTP/2 / gRPC followed per stream.
view="requests" (default) gives each request with status, sizes, response_ms and gRPC
status; view="endpoints" aggregates per host, method and path; summary names issues.
Filters: host (the name and its subdomains), path_prefix, method, status (404 or a
class such as "5xx"), min_ms (response_ms at least); filter is a display filter ANDed
with http || http2. No decryption: HTTPS needs TLS keys tshark already has.
More detail: resource schema://tools/http_transactions.
Inputs: path (required), end, host, view, +12 more. Returns: view, count, error, offset, +15 more
icmp_errorsICMP errorsARead-onlyIdempotent
Every ICMP and ICMPv6 error in the capture: who sends unreachables, about which
flows, is a firewall rejecting, is fragmentation needed somewhere, is TTL exceeded a
traceroute or a routing loop? Rows group the errors per sender, kind and quoted
flow (source, destination, protocol; the destination ports are listed): kind is
net_unreachable, host_unreachable, protocol_unreachable, port_unreachable,
frag_needed (with mtu), packet_too_big (ICMPv6, with mtu), admin_prohibited,
ttl_exceeded, reassembly_timeout, source_quench, redirect, parameter_problem,
no_route, address_unreachable or source_route_failed; with count, first/last frame,
the quoted IP ID and quoted_flow_seen (the frame of the original packet when the
capture holds it, IPv4 only). Also by_sender, echo (requests and replies) and issues. sender is an IP
address; about an address or address:port of the quoted flow (either end); kind keeps one
kind; filter is a display filter ANDed with the ICMP error and echo types.
More detail: resource schema://tools/icmp_errors.
Inputs: path (required), end, kind, about, +9 more. Returns: echo, count, error, errors, +15 more
ims_registrationIMS registrationARead-onlyIdempotent
IMS registration per UE: whether each UE is registered on IMS and, if not, why.
Use it for "why can't this UE register" or "is it registered for SMS/voice"; for calls
use sip_dialogs, for SMS over IMS sms_over_ims. Built from every
SIP REGISTER transaction in the capture (public identity in the To header); the legs of one REGISTER through the
CSCFs (same Call-ID and CSeq) count once, copies of a packet are told apart from
retransmissions. Each row has the REGISTER transactions, Contact feature tags (smsip,
MMTel), implicit IMPUs, third-party REGISTERs and a verdict (registered, expired,
deregistered, challenge-loop, failed, in_progress, unknown) with reasons. impu keeps
users whose IMPU, IMPI or P-Associated-URI contain it; imsi keeps IMSI-based identities;
filter is ANDed with the REGISTER filter.
More detail: resource schema://tools/ims_registration.
Inputs: path (required), end, impu, imsi, +9 more. Returns: count, error, ipsec, offset, +13 more
incident_summaryIncident summary (what is wrong in this capture)ARead-onlyIdempotent
Start here on any capture nobody has looked at yet: "what is wrong here?". Runs
pcap_info, capture_quality, conversations and the protocol tools the capture carries
under one time budget and returns findings ranked by severity (high, medium, low):
one plain sentence each, evidence (tool, arguments, frames), confidence and the
next_call that shows the detail. Also capture (span, packets, protocols, trust
verdict, what the capture cannot show), top_talkers (5), timeline, not_checked
(passes skipped and why) and not_analysed (protocols present that no pass reads, each
with its next_call). focus: auto (default), web, dns, tls, voip or mobile;
budget_s caps the whole call (default and maximum: the server's run timeout).
audience="support" adds explanation: the same findings in plain words for a support
agent or a ticket (what it means, who to involve, frames, ticket_lines ready to paste).
More detail: resource schema://tools/incident_summary.
Inputs: path (required), end, focus, start, +3 more. Returns: path, error, focus, passes, +11 more
media_path_checkMedia path check (one-way audio, 488)ARead-onlyIdempotent
Check one call's media path: the SDP offer and answer (the last completed
exchange, so a re-INVITE wins) against the RTP actually captured in each direction.
directions gives, for offerer to answerer and back, the expected source and
destination from the SDP, the streams seen and a status: as negotiated, arrives from
another source (NAT or relay), sent elsewhere, or missing. issues explain one-way
or no audio and failed calls: no_rtp, one_way_media, source_differs_from_sdp,
nat_not_latched (a private SDP address whose far end keeps sending to it),
rtp_to_wrong_address, private_sdp_address, media_direction (sendonly, recvonly,
inactive), media_rejected (port 0), payload_type_not_negotiated, codec_mismatch
(no codec in common between the offer and the answer, or the SDP of a 488) and
no_answer. verdict is the most telling issue in one sentence. RTP to or from the
call's SDP addresses is read; heuristic: true also decodes RTP that no SDP in the
capture set up.
Inputs: path (required), call_id (required), end, start, +2 more. Returns: error, found, offer, answer, +10 more
open_captureOpen a capture: a handle for the next callsARead-onlyIdempotent
Pin one capture for an investigation of several calls: returns a handle ("cap:..."),
accepted by every capture tool in place of path (also path_a/path_b). A filter and a
start/end window given here apply to every later call on the handle (a call's own
start/end replaces the window; its own filter is ANDed). Later calls on the same file
reuse tshark's output from a memory cache when they repeat a run (cache in a result counts
the reused runs). A handle stops working when the file changes (stale_handle) or the server
restarts (unknown_handle): call open_capture again.
More detail: resource schema://tools/open_capture.
Inputs: path (required), end, start, filter, +1 more. Returns: end, path, error, mtime, +9 more
packet_detailOne packet in detailARead-onlyIdempotent
The full dissection of one frame as tshark decodes it: every protocol layer with
its fields, nested as in Wireshark's packet details pane, with tshark's names for
enumerated values. Use it to check a row from another tool against the packet: rows
carry frame numbers (frame, request_frame, answer_frame, first_frame, ...).
layers keeps only these protocols, e.g. ["diameter"] or ["ip", "gtpv2"] (the frame
layer is always kept). At most max_fields leaf fields are returned (default 300,
server cap 2000); when the cap bites, the innermost layers are kept first and
fields_dropped counts the rest. Values longer than 512 characters are cut.
Inputs: path (required), frame (required), layers, redact, +1 more. Returns: path, time, error, frame, +8 more
pcap_infoCapture summaryARead-onlyIdempotent
Summarise a capture file: packet count, duration, first/last timestamp, size,
link type and the top protocol-hierarchy entries. Call this first to learn what a
capture contains before choosing a protocol tool. path must be inside an allowed
directory.
Inputs: path (required), end, start, redact. Returns: path, error, ipsec, limits, +16 more
roaming_attach_checkRoaming attach checkARead-onlyIdempotent
Why can this roamer not attach or get a bearer. Collects the subscriber's
attach and PDN steps (S6a AIR/AIA and ULR/ULA, MAP sendAuthenticationInfo,
updateLocation and updateGprsLocation for 2G/3G, GTPv2-C Create Session and Modify
Bearer, Gx CCR-I/CCA-I, PFCP Session Establishment) and groups them into attempts
in time order. Each attempt gives visited_plmn (ULR Visited-PLMN-Id, GTPv2 Serving
Network or the MAP peer GT), roaming, access (RAT, TAI, ECGI), steps with
result, code, ref and frames, the ULA subscription digest (carried from an
earlier attempt when this one has no ULA), the Create Session request (APN, PDN
type, selection mode, RAT, serving network, flags, APN-AMBR), checks (APN
subscribed, PDN type allowed, home routing or local breakout from the APN operator
identifier against VPLMN-Dynamic-Address-Allowed, RAT against
Access-Restriction-Data, ODB, AMBR), first_failure and a one-sentence verdict.
home_plmn ("MCC-MNC") overrides the home network read from the IMSI (needed for
multi-IMSI SIMs and MVNO ranges; otherwise the MNC length follows the MCC).
Content-free: no authentication vectors, SMS or USSD text are read or returned.
Inputs: imsi (required), path (required), end, limit, +8 more. Returns: imsi, count, error, offset, +16 more
rtp_streamsRTP streams and call qualityARead-onlyIdempotent
RTP streams with quality figures, linked to their SIP call. Per stream (SSRC plus
source and destination address:port): the Call-ID whose SDP advertised the
destination (or source) address and port (linked_by), payload type and codec
from the SDP rtpmap (or the static payload-type table), packets, expected, lost and
lost_pct, duplicates, sequence_errors, out_of_order, max/mean interarrival delta,
RFC 3550 max/mean jitter in ms, duration, the reverse stream's SSRC or one_way,
RTCP reception reports about the stream (fraction lost, cumulative lost, jitter,
round trip where the LSR matches a captured SR) and mos_estimate: a simplified
G.107 E-model score from loss, jitter, delay and codec. It is an ESTIMATE from
network metrics only (method and inputs are returned), never a measured MOS.
rtcp_xr lists RTCP XR VoIP metrics the endpoints reported themselves.
call_id keeps the streams of one call (RTP to or from its SDP addresses is read,
so a NATed source is still found); ssrc (decimal or 0x hex) keeps one stream.
tshark decodes RTP from the SDP in the capture; for a capture without the SIP
signalling set heuristic: true (heuristic RTP/RTCP over UDP, may misread other
UDP traffic as RTP). filter is ANDed with the RTP filter.
Inputs: path (required), end, ssrc, limit, +9 more. Returns: calls, count, error, notes, +16 more
sip_call_detailSIP call detailARead-onlyIdempotent
Everything about one SIP call (INVITE dialog) by Call-ID: the message ladder
(messages carrying SDP are marked), From/To/Request-URI/P-Asserted-Identity, the
outcome (answered, failed , cancelled, no final response),
final_response with its reason phrase, timing: post-dial delay (INVITE to the
first 180/183, or to the failure response when there was none, as RFC 6076 defines
Session Request Delay), ringing, answer and setup (ACK) times, early media (SDP in
a 18x) and the call duration from 200 OK to BYE; retransmissions per message;
every Reason header with its Q.850 cause named and every Warning header; who ended
the call (ended_by: BYE or CANCEL from the caller or callee side, told apart by
the From tag); and media: each SDP with its offer/answer role, address, port,
codecs, direction and ptime, the codecs offer and answer share and the negotiated
one. For the RTP itself use rtp_streams with the same call_id.
Inputs: path (required), call_id (required), end, start, +2 more. Returns: call, error, found, ipsec, +2 more
sip_dialogsSIP dialogsARead-onlyIdempotent
Group SIP messages by Call-ID: the list of calls and other dialogs and how each
ended. Use it to find a call; then sip_call_detail for one call's timings and SDP,
ims_registration for REGISTER flows, sms_over_ims for SIP MESSAGE delivery. Each dialog lists the method/response sequence with
timestamps, From, To, Request-URI, P-Asserted-Identity, the method chain, the final
status of the initial request, time to that final response and the dialog duration;
outcome_status is the final status after any 401/407 challenge was retried (a
REGISTER answered 401 then 200 succeeded), attempts counts transactions of the
initial method and resent_messages counts retransmissions;
no_final_response: true marks requests that never completed. Narrow with call_id
(exact), number (digits found in From, To, Request-URI or P-Asserted-Identity) or
method (INVITE, REGISTER, MESSAGE, ...): matching messages select the dialogs, then
every message of those dialogs is fetched so responses are kept. filter is a raw
display filter ANDed with sip; it applies literally to each packet.
Inputs: path (required), end, limit, start, +10 more. Returns: count, error, ipsec, offset, +13 more
sms_over_imsSMS over IMSBRead-onlyIdempotent
SMS over IMS (TS 24.341): SIP MESSAGE carrying application/vnd.3gpp.sms, one row
per RP-DATA or RP-SMMA. Each row gives direction (MO: RP-DATA from the UE with an
SMS-SUBMIT; MT: RP-DATA to the UE with an SMS-DELIVER or SMS-STATUS-REPORT), the RP
message reference, the SMSC address (smsc_digits), the TPDU (type, TP-MR,
TP-OA/TP-DA, status report requested, concatenation part, TP-Status of a status
report; never the text), the SIP response to the MESSAGE with its frame and delay,
and ack: the RP-ACK or RP-ERROR that answers it (paired by RP-MR and the UE's
identity, including the identities its REGISTER associated), with RP cause named,
the TPDU it carries and the SIP response to the MESSAGE that carried it. An MO
submit that asked for a status report gets status_report. ue_registration is the
UE's IMS registration from this capture and whether its Contact had +g.3gpp.smsip. number keeps rows naming
those digits, imsi rows whose identities contain it; filter is ANDed with the MESSAGE
filter.
More detail: resource schema://tools/sms_over_ims.
Inputs: path (required), end, imsi, limit, +9 more. Returns: count, error, ipsec, offset, +13 more
tcp_connectionsTCP connections (refused, filtered, who closed)ARead-onlyIdempotent
Can this host reach that port, and if not, is it refused, filtered or
prohibited? One row per TCP connection attempt (a SYN with a new initial sequence
number is a new attempt, the same one a retry; mirrored copies are dropped first)
with outcome: connected, refused (the server side answered the SYN with a reset:
nothing listens, or a firewall rejects with a reset), filtered (no answer at all: a
firewall drops it or the host is down; near_capture_end when the capture stopped
less than 3 s after the last SYN), prohibited (ICMP administratively prohibited
quoting the SYN), unreachable (another ICMP unreachable or TTL exceeded),
reset_during_handshake, handshake_incomplete (SYN-ACK never acknowledged), one_sided
(only one direction captured; not a failure) or seen_midway. Failed attempts come first; view="destinations" groups them per server
address and port with issues. server, client (IP addresses) and port (the server
port) narrow the attempts; outcome keeps one outcome; filter is a display filter ANDed
with tcp || icmp errors. For loss, windows and MTU inside a connection use tcp_health.
More detail: resource schema://tools/tcp_connections.
Inputs: path (required), end, port, view, +11 more. Returns: view, count, error, offset, +12 more
tcp_healthTCP stream healthARead-onlyIdempotent
Health of every TCP stream, with copies of the same packet (mirror, SPAN,
several interfaces) removed first, so they are not counted as retransmissions the
way tshark's own analysis counts them. Per stream: the handshake (SYN retries,
SYN to SYN-ACK and SYN-ACK to ACK times as seen from the capture point, MSS,
window scale, SACK); per direction the data segments and retransmissions split by
where the loss happened (lost_before_capture: a gap later filled by a
retransmission; retransmitted_after_capture: the capture saw the original;
missed_by_capture: acknowledged by the receiver but absent from the file, so
not network loss), reordering, duplicate ACKs, zero windows, keep-alives;
resets with likely_from endpoint or middlebox, judged by the reset's TTL and
source MAC against that side's other packets, and the idle time before it; ICMP
fragmentation-needed / packet-too-big about the stream (a mirrored copy of an ICMP
error counts once, summary.icmp_copies_dropped); and MTU problems against
mtu (an MSS too large for the path, DF packets above it, full-size segments
retransmitted and never acknowledged while small ones pass: a PMTUD black hole).
issues says it in plain words; streams with issues come first. stream is
tshark's tcp.stream index; filter is a display filter ANDed with tcp.
Inputs: path (required), end, mtu, limit, +8 more. Returns: count, error, offset, window, +13 more
tls_handshakesTLS handshakesARead-onlyIdempotent
TLS handshakes per connection, read from the clear part of the handshake. No
decryption and no key material. For TCP and QUIC: the client's offer (SNI, ALPN,
TLS versions, number of cipher suites without GREASE, fallback SCSV) against the
server's choice (version, cipher suite, ALPN when in clear), HelloRetryRequest,
time from SYN to ClientHello and from ClientHello to ServerHello, alerts with
level, name and who sent them (QUIC CONNECTION_CLOSE carrying a TLS alert too),
and for TLS 1.2 and older the leaf certificate (subject, issuer, SAN, notBefore,
notAfter, chain length). outcome is established, alert, no_server_hello,
reset_by_server, closed_by_server, client_gave_up, client_rejected_likely (TLS
1.3: the client sent one encrypted alert-sized record instead of Finished,
usually a certificate it would not accept), unreadable (the server answered but
tshark could not read a ServerHello), incomplete, one_sided (only one direction
captured: not a failure), or server_hello for QUIC (the rest is encrypted). server_name keeps that SNI and the names under it;
filter is ANDed with the base filter. Problem connections come first.
More detail: resource schema://tools/tls_handshakes.
Inputs: path (required), end, limit, start, +7 more. Returns: count, error, issues, offset, +12 more
trace_subscriberSubscriber traceARead-onlyIdempotent
Build one time-ordered timeline for a subscriber across Diameter (User-Name),
GTPv2-C and GTPv1-C (IMSI/MSISDN IEs, then following the control-plane TEIDs), PFCP
(sessions naming the IMSI or a UE address the GTP rows allocated, followed by SEID), MAP (IMSI/MSISDN)
and SIP (From/To/P-Asserted-Identity containing the MSISDN). Each row carries the
protocol, message name, result (Result-Code, GTP cause, SIP final status), peers and
the request/answer pairing including unanswered: true. Give imsi, msisdn or
both as digit strings. The whole call shares one time budget (the per-run timeout);
when it runs out the timeline found so far is returned with truncated and the pass
that stopped marked timed_out. GTP rows carry the proxy-chain fields of
gtp_sessions (chain, leg_without_response, near_capture_end, duplicate_of_frame, note);
unanswered Diameter and GTP rows the capture stopped too soon after get near_capture_end,
counted in unanswered_near_capture_end.
Start here for "why did subscriber X fail" questions.
Inputs: path (required), end, imsi, limit, +8 more. Returns: count, error, ipsec, offset, +16 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 34 tools
Nearly every tool has a clearly bounded job, and the descriptions actively route the agent ("for calls use sip_dialogs, for SMS over IMS sms_over_ims"; tcp_connections vs tcp_health; diameter_transactions vs diameter_sessions vs diameter_avps). The only real overlap is between the capture-overview family (pcap_info, capture_metadata, capture_quality, capture_index, incident_summary), where incident_summary explicitly says "start here" and calls the others, so an agent may hesitate over which entry point to use but not misselect a protocol tool.
All names are snake_case and consistently descriptive: protocol tools are named resource-first (diameter_sessions, gtp_ies, dns_transactions, tls_handshakes) and action tools verb-first (filter_packets, find_fields, check_filter, open_capture, compare_captures). The mix of noun-phrase and verb_noun patterns is a minor deviation from a single convention, but it is predictable and never ambiguous.
34 tools is above the comfortable range, but the surface is genuinely broad (Diameter, GTPv2-C, SIP/IMS, RTP, DNS, TLS, HTTP plus capture hygiene and cross-protocol correlation), and each tool appears to cover a distinct protocol+view. The meta-tools (incident_summary, trace_subscriber, roaming_attach_check, compare_captures) and open_capture handle caching help keep the set navigable, but it is still at the heavy end.
Coverage is unusually complete: capture discovery and indexing, quality/trust assessment, raw filter and per-frame fallbacks, filter authoring helpers, cause decoding, and deep per-protocol transaction/IE tools with cross-protocol subscriber tracing and A/B comparison. Remaining gaps are minor — no follow-stream payload extraction or export-objects tool, no dedicated PFCP or MAP transaction tool (only reachable through trace_subscriber/roaming_attach_check), and no decryption path for TLS/HTTP.
Maintenance
Related MCP Connectors
LLM agent security scan: injection, jailbreak, tool abuse probes plus guardrail audit.
Safety tools for AI agents: safe page reader, prompt-injection, secrets, PII and MCP server scans
Read-only smart-contract security intelligence for autonomous agents.
Read-only tools over the Safer Agentic AI framework: 239 patterns + 15 heuristics.
Related MCP Servers
- FlicenseNot gradedqualityNot gradedmaintenanceEnables Wireshark-like PCAP analysis through tshark, providing tools for filtering packets, extracting protocol fields, drilling down into frames, session tracking, and timeline analysis for troubleshooting 5G, IMS/SIP, and network protocol issues.-
- AlicenseCqualityDmaintenanceEnables LLMs to perform automated network forensics, threat hunting, and traffic analysis using Wireshark/TShark.344MIT
- FlicenseCqualityDmaintenanceEnables LLMs to capture, analyze, and summarize network traffic using Wireshark CLI tools, supporting live capture, pcap analysis, and LLM-oriented summaries.76-
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to interact with tshark and related command-line tools for live packet capture, PCAP analysis, filtering, stream following, and capture file merging.1MIT