Repository navigation
Releases
vkteams-mcp 0.4.0
Many replicas, JSON logs, metrics, and a chart.
- A shared state store. The counters that cannot travel sealed move out of each replica's memory into Redis or
Valkey (standalone withVKTEAMS_MCP_STATE_URL, or behind Sentinel withVKTEAMS_MCP_STATE_SENTINELSand
VKTEAMS_MCP_STATE_SENTINEL_MASTER;_STATE_DB,_STATE_TLS,_STATE_USERNAME,_STATE_PREFIX; passwords only from
files,_STATE_PASSWORD_FILEand_STATE_SENTINEL_PASSWORD_FILE, with the key file's checks). Used authorization
codes (a code redeemed on one replica was accepted again on another), sign-in code requests per address and per
network, wrong codes per sign-in step, the send allowance and the upload locks. Sessions, drafts and links stay sealed
and are never stored. Keys are HMACs of the value under a sub-key of the sealing key: no login, address or IP in the
store; every key has a TTL, and the scripts carry#!luaso a store out of memory refuses them instead of letting a limit through (Redis or Valkey 7 or later). Nothing set is memory, right for one replica, with a warning at start. - A store that cannot be reached fails closed. Sign-in, spending a code and sending refuse with "try again shortly";
reading goes on. A refused token request answers 503 and leaves the code unspent./readyzdoes not depend on the store. - JSON logs, all of them. Every line of every process, the init container and a failed start included, is one
JSON object (ts,level,logger,event,msg); a quote or a traceback can no longer break a record, a traceback
is frames and exception types without messages, and a record that looks like content is dropped and counted.
LOG_LEVEL. - Metrics on
VKTEAMS_MCP_METRICS_PORT(9090), never the public port, with/healthzand/readyzthere too:
mcp_http_requests_in_flightand the family's other series (requests, tools, VK Teams calls, sign-in, refusals,
links, the store,mcp_build_info), all labelledserver="vkteams-mcp". VKTEAMS_MCP_TRUSTED_PROXIES. Behind an ingress the per-network sign-in limit counted the whole company as one
address; now only the listed proxies may say who the client is, and without the setting that limit is off (the
per-address limit stays).VKTEAMS_MCP_CA_BUNDLEfor a VK Teams server behind an internal CA.- The chart
oci://ghcr.io/widmo-ru/charts/vkteams-mcpreplacesdeploy/kubernetes/anddeploy/helm/: the key
through the same init container (nowpython -m vkteams_mcp.stage, JSON log, also staging the store's passwords), a
rolling update, a PodDisruptionBudget, soft spreading, probes on the metrics port, explicit ingress paths, an optional
ServiceMonitor (off by default), no HPA by default. - The key is stable across deploys. 0.3.x drew a new random key on every Helm render, so every deploy signed
everyone out. The key is now a Secret you own (scripts/new-sealing-key.shmakes one), and a deploy keeps everyone
signed in. Replacing it is a deliberate rotation: everyone signs in again once. - Breaking:
VKTEAMS_MCP_SECRETis refused, as 0.3.0 announced. The server does not start while the variable is
set, whether or not a key file exists, and the error names the variable. To migrate: put the key in a Secret (the one
you already run with, if you want nobody signed out:kubectl get secret … -o jsonpath='{.data.key}'holds it, or
write the value of the variable under the fieldkey), deploy with the chart (or add the init container
python -m vkteams_mcp.stage /secret/key:/run/vkteams-mcp/key), and remove the variable from the Deployment,envFrom
and any secret manager that sets it. The same key value in the file keeps everyone signed in. Steps in
deploy/README.en.md. - The image's bases are pinned by digest and replaceable (
BASE_IMAGE,UV_IMAGE). The image is taggedx.y.zandx.y; there is nolatesttag any more: pin a version. - A number that is not a number in a header or a setting is refused, not a crash.
Proof is in CI: the state store's contract suite on real Valkey and Redis (standalone and Sentinel), and the chart in
kind with three replicas, Valkey with Sentinel and a fake VK Teams (scripts/chart-e2e.sh).
Image: ghcr.io/widmo-ru/vkteams-mcp:0.4.0 (linux/amd64, linux/arm64)
Chart: helm upgrade --install vkteams-mcp oci://ghcr.io/widmo-ru/charts/vkteams-mcp --version 0.4.0 -f values.yaml
vkteams-mcp 0.3.0
The sealing key leaves the environment.
- The key arrives as a file,
VKTEAMS_MCP_KEY_FILE(default/run/vkteams-mcp/key), mode 0400 and owned by the
server's user. In Kubernetes an init container copies it there from a Secret into an emptyDir with
medium: Memory; the server's container gets no Secret volume and no variable, sokubectl exec … envshows
nothing. The file stays, so a container restarted inside the pod (an OOM kill, a crash, a liveness probe)
keeps everyone signed in. Without the file the server does not start and says why; a file the group or
anyone else can read is refused. - The process makes itself non-dumpable (
PR_SET_DUMPABLE=0) on Linux: another process of the same user cannot
read its memory through/proc/<pid>/memor ptrace. What remains iskubectl execinto the server's
container reading the file as the same user: that is closed by the cluster's RBAC onpods/exec. - The deploy pipeline writes a new random key into the Secret on every deploy:
deploy/kubernetes/new-key.sh
for any CI,deploy/helm/secret.yamlfor Helm and werf. Every deploy signs everyone out once: old tokens
get 401invalid_tokenwithresource_metadata, the old registration getsinvalid_client, and the client
signs in again. A pod restarted or rescheduled without a deploy reads the same Secret and keeps everyone signed in.
Replicas are fine;strategy: Recreatekeeps two keys from serving during a rollout. VKTEAMS_MCP_SECRETis deprecated. In 0.3.x it is still used when there is no key file, with a warning
in the log, so an upgrade with unchanged manifests neither crash-loops nor signs anyone out; when a key file
is there the variable is ignored. The next minor release refuses to start while it is set. Either way it is
dropped from the process's environment.- The settings no longer print the key in their
repr. - The image has
/run/vkteams-mcp, owned by the server's user.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.3.0 (linux/amd64, linux/arm64)
Verified live on 2026-11-19 against a real VK Teams server, on this image in kind: no key in env or logs, key file -r-------- 10001, /proc/1/environ refused, no Secret volume in the server's container; a signed-in session survived a container restart with the same access token; after a new key the old access token got 401 invalid_token with resource_metadata, the old refresh token 401 invalid_client, and a second sign-in worked. How to repeat it: deploy/README.en.md.
vkteams-mcp 0.2.6
From the first production report.
- One-to-one chats no longer vanish from the chat list. An older session's event stream replayed the
group list without dialog states; the server now opens a new session whenever the stream carries no
dialog states, not only when it is empty. - A chat can be named by a colleague's name even when it is not in the chat list yet: the directory is
searched, asfind_peopledoes, and several matches are offered rather than guessed. "Избранное",
"Favorites" and "Saved messages" name the person's own chat. - The image is built on Alpine and has no HIGH or CRITICAL findings in Trivy; the Debian base carried 101,
none with a fix from Debian. CI now fails on any such finding. - The image is published as a plain Docker manifest list with linux/amd64 and linux/arm64 and nothing else.
The provenance entries of earlier releases made a Nexus proxy offer only arm64.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.6 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.5
The open QA remarks.
- Search across all chats reports
chats_matchedinstead of atotalthat counted chats, not messages,
and stops offering a cursor once every matching chat has come. - Voting checks the poll first: a closed poll and an option the poll lacks are refused by name, before the
action allowance is spent. find_peoplesays when there are more people than shown, marks bots askind: bot, refuses an empty
query and says how many chats a person shares when it lists only ten.list_chatsfollows the event stream past the session's first batch, solast_at, unread counts and last
messages show what happened since the session began (live:last_atstayed hours behind).- Sending a draft that went out before says so:
already_sentwith a note. VK Teams answers such a resend with
the first message's id and time, even if that message was deleted, and posts nothing new. - A call needs a title. A damaged cursor, garbled or with an altered signature, is called damaged.
get_chatwithinclude_memberson a
one-to-one chat says who its members are instead of ignoring the request.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.5 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.4
From the sixth QA round, a full regression of 158 scenarios with no failures.
- A sealed link, token or draft has one spelling: the spare bits of its last character must be zero, so
four spellings of one link no longer all work. - A reaction that is not one emoji is refused as such, not as a missing reaction.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.4 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.3
Fixes from the fifth QA round.
- Control characters, private-use and unassigned code points count as invisible: a poll answer, a message
or an edit made only of them is refused. Live, a poll answer of U+0001 alone was sent and kept. Emoji
blocks stay visible, so an emoji newer than the server's Unicode database is not taken for empty. - A reply to someone else's message names the quoted author among the draft's mentions and counts them
towards the 20-people cap: VK Teams notifies them (live: 20 mentions and a quote recorded 21).
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.3 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.2
Fixes from the fourth QA round.
- A mention written with spaces inside,
@[ address ], counts towards the 20-people cap in drafts and
edits and is listed in the draft preview. VK Teams treats it as a mention (live: an edit with 21 of them
notified 21 people). - An edit to invisible characters only is refused, as a draft already was. Hangul fillers (U+3164 and
its kin) and the blank Braille pattern (U+2800) now count as invisible in drafts, polls and edits. - A bot button's print is keyed by the server's secret. The unkeyed 24-bit print let a bot find another
button with the same print in under a minute and have it pressed. Its fields are encoded one by one,
so moving a line from the label into the data changes the print. @someone@example.comin an edit becomes a mention, as in a draft, and counts towards the cap.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.2 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.1
- A client the server no longer knows (registered with 0.1, or before the secret changed) used to get
raw JSON from/authorize. The browser now shows a page in the person's language that says to clear
the connector's authorization in the AI app and connect again. Redirects are unchanged.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.1 (linux/amd64, linux/arm64)
vkteams-mcp 0.2.0
The server keeps nothing: any number of replicas sharing one secret serve the same clients.
- The person's VK Teams session travels sealed (AES-GCM) inside their OAuth tokens, which the client
refreshes itself by the standard. The client's registration is itsclient_id; the authorization code,
the sign-in code step and file links carry what they need. No database, no/datavolume,
VKTEAMS_MCP_DATABASEis gone. save_draftreturns a sealeddraft;send_drafttakes it, and with attachments thefileseach upload
printed. Sending the same draft again posts nothing twice: VK Teams merges a resend with the same
localMsgId(live), and scheduled drafts are checked against the queue, which VK Teams does not merge.- Disconnecting ends the session in VK Teams. A single issued token can no longer be revoked before it
expires (access tokens live an hour), a refresh token is not single-use, and the rate allowance is counted
per replica. A scheduled poll or quote sent again is queued again. - A test runs a sign-in, a draft and an upload across fresh replicas; CI starts two containers and checks
that one accepts what the other issued.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.2.0 (linux/amd64, linux/arm64)
vkteams-mcp 0.1.0
The first public release.
- 12 tools by default (14 with group administration): chats, messages, people, files, call links and
bots, about 4,100 tokens of catalogue. - Streamable HTTP with OAuth 2.1: client registration, PKCE, rotating refresh tokens. A sign-in page in
Russian and English where the person types the code VK Teams e-mails them; the model never sees it. - Two-step send, a confirmation flag on every action others see, one rate allowance per person, at most
20 people per action, mentions included. - Message ids travel as strings: live ids exceed 2^53, which a JavaScript client rounds.
- A bot's buttons are shown by number and a print; the server presses with the bot's own data and
refuses a button the bot changed after it was read. - Files move through signed links that expire in minutes; downloads pass a byte range through.
- Checked on a live VK Teams server (API 139) in three QA rounds; every defect found is fixed and held by
a test.
Image: ghcr.io/widmo-ru/vkteams-mcp:0.1.0 (linux/amd64, linux/arm64)