yandex-tracker-mcp-lite

Unofficial MCP server for Yandex Tracker that keeps your agent's context for your work: 9 tools, about 1.2k tokens of tool definitions. The official server costs about 24.8k tokens before you ask anything, the most popular community server about 20.7k.

Лёгкий неофициальный MCP-сервер для Яндекс Трекера: 9 инструментов, около 1,2 тыс. токенов на описания инструментов вместо 24,8 тыс. у официального и 20,7 тыс. у aikts, ответы в 1,4–30 раз короче, всё остальное API доступно через один инструмент call_api. Настройка одной командой: uvx yandex-tracker-mcp-lite setup.

uvx yandex-tracker-mcp-lite setup

Measured

Same organisation, same questions, measured with bench/measure.py on 2026-11-08 against the official server at mcp.tracker.yandex.net and aikts/yandex-tracker-mcp (commit 51a0c0d, default settings and with TRACKER_ENTITIES_ENABLED). Character counts are exact; tokens are estimated at 3.3 characters per token.

Tool definitions. "Seen by the model" is name + description + input schema, the part a client sends to the model in every session. The full tools/list also carries output schemas and annotations, which clients keep to themselves; for aikts those output schemas are 128k of its 201k characters.

	tools	tools/list chars	seen by the model, chars	≈ tokens
official	71	84,684	81,761	24,776
aikts	55	200,864	68,226	20,675
aikts + entities	94	468,529	144,825	43,886
lite	9	4,355	4,044	1,225

Answers to the same questions, characters returned (aikts needs two calls for an issue with comments):

question	official	aikts	lite
who am I	107	206	74
read an issue	707	2,561	511
issue + last 5 comments	1,454	4,000	691
search a queue (12 issues)	3,537	27,785	918
available transitions	1,022	210	31

Clients that load MCP tools on demand (Claude Code's tool search) pay less of the definitions up front; Codex, Cursor and Claude Desktop pay all of it in every session. The answers are smaller everywhere.

Set up in two minutes

You need: a Tracker organisation you belong to, uv (for uvx), and a browser where you are signed in to Yandex with the account you use in Tracker.

The quick way

Run this in your own terminal (not through an agent's shell: the token would end up in its transcript):

uvx yandex-tracker-mcp-lite setup

It walks you through the steps below, opens the right pages, reads the token with a hidden prompt, detects whether your organisation is Yandex Cloud or Yandex 360, checks access, explains what is wrong if it fails, and adds the server to Claude Code. For other clients it prints ready configs (the token masked unless you pass --show-token).

By hand

1. Create an OAuth app at oauth.yandex.ru/client/new:

• choose «Для доступа к API или отладки» (API access or debugging). "No access to user data" there refers to the Yandex ID profile, not to Tracker;
• name it anything, give a contact email;
• in «Доступ к данным» type «трекер» and add «Чтение из трекера» (tracker:read) and «Запись в трекер» (tracker:write);
• «Создать приложение», copy the ClientID. You don't need the client secret.

2. Issue a token. Open this link with your ClientID, check that the account in the corner is the one you use in Tracker, press «Разрешить» and copy the token (y0_…, valid for a year):

https://oauth.yandex.ru/authorize?response_type=token&client_id=<ClientID>

3. Find the organisation id. Tracker → Администрирование → Организации, field «Идентификатор». The card's title tells you which variable to use:

card title	variable
Yandex Cloud Organization	TRACKER_CLOUD_ORG_ID
Яндекс 360 для бизнеса	TRACKER_ORG_ID

4. Check it works (replace the two values, keep the quotes):

export TRACKER_TOKEN='y0_...'
export TRACKER_CLOUD_ORG_ID='bpf...'        # or TRACKER_ORG_ID for Yandex 360

curl -s https://api.tracker.yandex.net/v3/myself \
  -H "Authorization: OAuth $TRACKER_TOKEN" \
  -H "X-Cloud-Org-ID: $TRACKER_CLOUD_ORG_ID"   # X-Org-ID for Yandex 360

You should see your login. If not, see Troubleshooting.

5. Add the server. In the same shell:

Claude Code:

claude mcp add --scope user tracker \
  -e TRACKER_TOKEN="$TRACKER_TOKEN" \
  -e TRACKER_CLOUD_ORG_ID="$TRACKER_CLOUD_ORG_ID" \
  -- uvx yandex-tracker-mcp-lite

Codex (~/.codex/config.toml):

[mcp_servers.tracker]
command = "uvx"
args = ["yandex-tracker-mcp-lite"]
env = { TRACKER_TOKEN = "y0_...", TRACKER_CLOUD_ORG_ID = "bpf..." }

Cursor, Claude Desktop, anything else (mcpServers JSON):

{
  "mcpServers": {
    "tracker": {
      "command": "uvx",
      "args": ["yandex-tracker-mcp-lite"],
      "env": { "TRACKER_TOKEN": "y0_...", "TRACKER_CLOUD_ORG_ID": "bpf..." }
    }
  }
}

Restart the client and ask something like «что у меня в работе в трекере?».

Troubleshooting

you see	it means	fix
tracker:read OAuth scope is required	the OAuth app has no Tracker access	add both Tracker accesses to the app (step 1), then issue a new token (step 2)
401 … expired_token / INVALID	wrong or expired token	copy it again, or issue a new one
403 Organization is not available, not ready or not found	wrong org id, wrong header, or the token belongs to another account	re-copy the id; try the other variable; sign in to the right account and repeat step 2
OAuth token … issued after '2026-06-01', is not supported for IAM token exchange	you went the yc CLI route	yc tokens no longer exchange for IAM tokens; use the OAuth app above
everything is read-only in Tracker	the organisation's free period ended	add a billing account in Tracker; reads keep working

Keep the token secret. It acts as you in Tracker. If it leaks (pasted into a chat, committed), revoke it in Yandex ID → Безопасность and issue a new one.

Variables

variable
TRACKER_TOKEN	OAuth token. Or TRACKER_IAM_TOKEN (12 h, Yandex Cloud orgs, e.g. from a service account).
TRACKER_CLOUD_ORG_ID / TRACKER_ORG_ID	organisation id, see step 3
TRACKER_READ_ONLY=1	optional: only whoami, search_issues, get_issue and GET-only call_api (≈ 600 tokens)

Saving the extra process. uvx keeps a small uv process (~26 MB) alongside the server. To avoid it, install once with uv tool install yandex-tracker-mcp-lite and use yandex-tracker-mcp-lite as the command instead of uvx yandex-tracker-mcp-lite. Update with uv tool upgrade yandex-tracker-mcp-lite.

The names match aikts/yandex-tracker-mcp, so switching is a change of command.

Tools

tool	what it does
whoami	login, name, email; a quick check that the token works
search_issues	Tracker query language or a preset (mine, created, following). One line per issue: key | status | assignee | updated | summary, with a next-page hint
get_issue	all fields in one call; comments=N adds the last N comments, attachments=true downloads images to local files the agent can open, changelog=true adds field history
add_comment	markdown comment, optionally summoning people
transition_issue	without status: lists the transitions available now; with it: matches an id, a name or the target status
update_issue	any fields: {"assignee": "me", "tags": {"add": ["x"]}, "deadline": "2026-12-01"}
create_issue	new issue, any extra fields
link_issues	relates, depends on, subtask, duplicates, epic
call_api	any Tracker API v3 request: boards, sprints, worklogs, checklists, projects, queues

How it stays small

Following Anthropic's Writing effective tools for agents and the Claude Code MCP docs:

• A few tools for whole workflows, one escape hatch for the rest. get_issue answers "what's going on with DEV-123" in one call instead of four. Rare operations go through call_api(method, path, body), which costs ~200 tokens and covers the whole API.
• Server instructions that say when to look here. With tool search only tool names and server instructions load up front, so the instructions lead with the words people use: Yandex Tracker, Яндекс Трекер, задачи, тикеты, issue keys like DEV-123.
• Help arrives with the error. The query-language cheat sheet is not in any schema. When Tracker rejects a query, the error comes back with it. 401/403/404 errors say what to check next, and a search that has more pages says how to get them.
• Answers are shrunk, not dumped. Reference objects become their display text ("assignee": "Анна Петрова", "parent": "DEV-0 Auth epic"), bookkeeping fields (self, version, uids, counters) and empty values are dropped, timestamps lose their seconds, JSON has no indentation. call_api keeps ids and keys, because follow-up calls need them.
• People by name. assignee, followers and summon take a login, an email, a display name or me. If a name matches several people the call fails with the candidates instead of guessing.
• Annotations that clients act on. Read tools are readOnlyHint (Claude Code runs them in parallel), update_issue is idempotentHint (safe to retry after a dropped connection), creating tools are non-destructive.
• Schemas without pydantic decoration. No auto-generated titles, no anyOf: [X, null].

Limitations

• File upload is not supported yet (call_api sends JSON).
• Name resolution loads the organisation's user list once per session (up to 1000 users).
• Not affiliated with Yandex.

Development

uv run pytest                                      # offline tests
TRACKER_TEST_QUEUE=TEST uv run pytest -m live -s   # creates, edits, links and closes issues in TEST
uv run python bench/measure.py --issue TEST-1 --queue TEST [--aikts ~/src/yandex-tracker-mcp]

MIT licence.