Taiga MCP Server
MCP server for Taiga project management, enabling LLM clients to inspect and manage projects, work items, sprints, comments, attachments, and wiki pages via 6 op-dispatching tools.
README
Taiga MCP Server
Model Context Protocol (MCP) server for Taiga project management, written in TypeScript and built on the Model Context Protocol SDK over stdio transport. It connects LLM clients to Taiga instances to inspect and manage projects, work items (issues, user stories, tasks, epics), sprints, comments, attachments, and wiki pages.
The server consolidates all capabilities into 6 op-dispatching tools designed for minimal token overhead and dense, human- and LLM-readable text responses.
Requirements and Configuration
- Node.js >= 20.11
- A Taiga account on taiga.io or a self-hosted Taiga instance
- Three environment variables configured:
| Variable | Description | Default |
|---|---|---|
TAIGA_API_URL |
Base URL of the Taiga REST API (must include /api/v1) |
https://api.taiga.io/api/v1 |
TAIGA_USERNAME |
Taiga username or email | Required |
TAIGA_PASSWORD |
Taiga account password | Required |
MCP Client Configuration
Add the server to your MCP client configuration (such as claude_desktop_config.json or your harness configuration).
Using Local Checkout
When running from a local checkout, run npm run build first to compile the TypeScript source into dist/.
{
"mcpServers": {
"taiga": {
"command": "node",
"args": ["/absolute/path/to/dist/src/index.js"],
"env": {
"TAIGA_API_URL": "https://api.taiga.io/api/v1",
"TAIGA_USERNAME": "your_username",
"TAIGA_PASSWORD": "your_password"
}
}
}
}
Using npx
The package is not yet published to npm, so the npx configuration below is forward-looking:
{
"mcpServers": {
"taiga": {
"command": "npx",
"args": ["-y", "mcp-taiga"],
"env": {
"TAIGA_API_URL": "https://api.taiga.io/api/v1",
"TAIGA_USERNAME": "your_username",
"TAIGA_PASSWORD": "your_password"
}
}
}
}
Containers
The included two-stage Dockerfile uses Node 22 Alpine with a non-root node user. The build stage compiles the TypeScript source, and the runtime stage packages only the compiled dist/src output and production dependencies.
Build the container image:
docker build -t mcp-taiga .
Run the container attached to standard I/O:
docker run --rm -i --env-file .env mcp-taiga
Podman works by direct substitution: replace docker with podman in the commands above.
Point an MCP client configuration at the container runner:
{
"mcpServers": {
"taiga": {
"command": "docker",
"args": ["run", "--rm", "-i", "--env-file", "/absolute/path/to/.env", "mcp-taiga"]
}
}
}
There is deliberately no compose file. An MCP stdio server must be spawned attached directly to its client's stdin and stdout streams, and exits when that stdin stream closes. Process supervisors or compose setups that attempt to keep long-running background services alive cause infinite restart loops and container name conflicts.
Conventions
- Projects: Project arguments accept a numeric project ID (e.g.
19) or a slug (e.g."acme-web"). - Work Items: Work items accept a numeric database ID (e.g.
1888) or a reference prefixed with hash (e.g."#70"). A#referencerequires theprojectargument to resolve. - People: Member arguments accept a numeric user ID, username (e.g.
"jdoe"), full name (e.g."Jane Doe"), or the literal"me". - Taxonomies: Statuses, priorities, severities, issue types, and sprint names accept human-readable names and are resolved to numeric IDs server-side.
- Multi-Assignee User Stories: Taiga user stories support multiple assignees via
assigned_users. Filtering byassigneeonwork list type:storymatches co-assignees, and listings display all assignees rather than only the primary. - Dense Text Results: Results are plain text formatted with one dense line per record or clean key-value blocks for detail views. No consumer reads
structuredContent, so results are text only. An empty collection is reported as<items> in <project>: 0rather than an error. - Full Collections: List endpoints return complete collections because the client sends the
x-disable-pagination: truerequest header, eliminating multi-page roundtrips.
Why Six Tools
Single-purpose tool proliferation creates substantial context window overhead before any tool is invoked. Consolidating functionality into 6 op-dispatching domain tools keeps the tools/list payload to about 10,267 characters (~2,775 tokens).
Output schemas are deliberately absent from tool registrations: MCP client bridges concatenate text content and ignore outputSchema and structuredContent, so omitting output schemas eliminates unnecessary token overhead on session startup.
Tool Reference
The server exposes 6 tools covering 28 operation pairs.
1. projects
List or inspect Taiga projects and verify credentials.
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List projects where authenticated user is a member | (none) | (none) |
get |
Inspect project metadata, owner, member count, active modules | project |
(none) |
whoami |
Verify credentials and show current user info | (none) | (none) |
2. work
Manage issues, user stories, tasks, and epics (type: issue, story, task, epic).
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List work items with server-side filters | type, project |
assignee, watcher, sprint, status, tags, closed, q, orderBy, limit, parent (tasks) |
get |
Get complete details and description for a work item | type, item |
project (required if item is #ref) |
create |
Create a single work item or batch items | type, project, subject (or items for batch; tasks require parent) |
description, status, assignee, sprint, tags, priority (issue), severity (issue), issueType (issue), points (story), parent (epic for story / default for batch), items (max 20) |
update |
Update fields on an existing work item | type, item |
project (required if item is #ref), subject, description, status, assignee, sprint, tags, priority, severity, issueType, points |
link |
Link a user story to an epic | type (story), item (story), parent (epic) |
project (required if item or parent is #ref) |
unlink |
Remove a user story from an epic | type (story), item (story), parent (epic) |
project (required if item or parent is #ref) |
delete |
Permanently delete a single work item | type, item |
project (required if item is #ref) |
3. sprints
Manage sprints (milestones) and inspect progress statistics.
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List sprints in a project | project |
(none) |
get |
Get sprint details and assigned user stories | sprint |
project (required if sprint is a name) |
create |
Create a new sprint milestone | project, name |
start (YYYY-MM-DD), finish (YYYY-MM-DD) |
stats |
Get sprint progress statistics and completion metrics | sprint |
project (required if sprint is a name) |
Sprint deletion is intentionally not exposed: removing a milestone detaches every story and task on it, making it a board-wide edit that belongs in the Taiga UI.
4. comments
List, add, edit, or delete comments on work items and wiki pages (type: issue, story, task, epic, wiki).
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List comments oldest first | type, item |
project (required for #ref or wiki slug), includeDeleted |
add |
Add a comment to an item | type, item, text |
project (required for #ref or wiki slug) |
edit |
Edit an existing comment by UUID | type, item, commentId, text |
project (required for #ref or wiki slug) |
delete |
Soft-delete a comment by UUID | type, item, commentId |
project (required for #ref or wiki slug) |
5. attachments
Manage file attachments on work items and wiki pages (type: issue, story, task, epic, wiki).
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List attachments on an item | type, item |
project (required for #ref or wiki slug) |
upload |
Upload a file from local path or base64 | type, item, filePath or fileContent |
project, fileName, mimeType, description |
download |
Fetch attachment metadata; optionally writes file to disk | type, attachmentId |
savePath (path to save downloaded file) |
delete |
Permanently delete an attachment | type, attachmentId |
(none) |
6. wiki
Manage wiki pages and page subscriptions within a project.
| Op | What it does | Required args | Optional args |
|---|---|---|---|
list |
List all wiki pages in a project | project |
(none) |
get |
Inspect wiki page metadata and Markdown content | page (ID or slug) |
project (required if page is a slug) |
create |
Create a new wiki page | project, page (slug) |
content |
update |
Update wiki page content | page (ID or slug), content |
project (required if page is a slug) |
delete |
Permanently delete a wiki page | page (ID or slug) |
project (required if page is a slug) |
watch |
Watch or unwatch a wiki page | page (ID or slug) |
project (required if page is a slug), watch (boolean, default true) |
Reliability and Safety
- Rate Limiting (429): The server retries HTTP 429 responses at most twice, honoring the server
Retry-Afterheader. If the required wait exceeds the 5-second ceiling (MAX_THROTTLE_WAIT_MS), it throws immediately with a retry message instead of sleeping. - 5xx Errors Never Retried: 5xx responses are never retried automatically because mutating requests (such as POST) may have already been applied on the server; repeating them risks creating duplicate records.
- Metadata Cache: Project metadata (slug lookups, user memberships, and taxonomy lists for statuses, priorities, severities, and issue types) is cached for 60 seconds (
METADATA_TTL_MS) viagetMetadata. Work items, comments, and attachments are never cached. - Timeouts: HTTP requests enforce a 30-second timeout (
REQUEST_TIMEOUT_MS). - HTTPS Enforcement: The server emits a warning to stderr if
TAIGA_API_URLuses unencrypted HTTP to a non-loopback host. - Restricted Attachment Downloads: Attachment downloads are restricted strictly to the configured Taiga hostname with no redirects allowed (
maxRedirects: 0), bounded to a maximum file size of 10 MB (MAX_ATTACHMENT_BYTES). The download request does not send the Taiga bearer token to media hosts. - File Overwrite Protection: Attachment download with
savePathrefuses to overwrite an existing local file. - Single-Target Deletions: Deletion operations accept exactly one target at a time. Batch operations are create-only (up to 20 items), preventing accidental board-wide deletions.
Development
File Layout
src/index.ts # Server entrypoint, stdio transport, resource registration, tool registry
src/api.ts # Authenticated axios transport, generic HTTP helpers (get, post, patch, del), token management, retry policy, metadata cache
src/taiga.ts # Domain helpers: resolution (projects, items, members, taxonomies, sprints) and optimistic concurrency patch
src/types.ts # Taiga payload interfaces, tool definitions, and type contracts
src/format.ts # Dense pipe-separated single-line renderers and detail views
src/utils.ts # MCP response builders (createSuccessResponse, createErrorResponse, guard) and formatting helpers
src/constants.ts # Endpoints, limits (batch size, attachment size), status labels, error messages
src/tools/index.ts # Tool registry aggregating all tools and registering with McpServer
src/tools/projects.ts # projects tool (list, get, whoami)
src/tools/work.ts # work tool (list, get, create, update, link, unlink, delete across issues, stories, tasks, epics)
src/tools/sprints.ts # sprints tool (list, get, create, stats)
src/tools/comments.ts # comments tool (list, add, edit, delete)
src/tools/attachments.ts # attachments tool (list, upload, download, delete)
src/tools/wiki.ts # wiki tool (list, get, create, update, delete, watch)
test/unitTest.ts # Offline unit tests for pure helpers, formatting functions, response builders, and tool invariants
test/protocolTest.ts # Protocol tests verifying MCP stdio handshake, server capabilities, tool count, and tools/list budget
test/apiContractTest.ts # Contract tests driving every tool op against an in-process mock Taiga HTTP server, asserting outgoing HTTP requests
test/integration.ts # Live integration smoke test against a real Taiga instance (read-only, skips without credentials)
NPM Scripts
npm run build: Compiles TypeScript fromsrc/andtest/intodist/viatsc.npm run check: Type-checks TypeScript code without emitting output (tsc --noEmit).npm run lint: Runs oxlint acrosssrc/andtest/.npm start: Runs the compiled server (node dist/src/index.js).npm test: Compiles and runs unit, protocol, and contract test suites in sequence.npm run test:unit: Compiles and runs offline unit tests.npm run test:protocol: Compiles and runs MCP protocol tests over stdio.npm run test:contract: Compiles and runs API contract tests against the mock Taiga server.npm run test:integration: Compiles and runs live integration tests against a live instance.npm run prepublishOnly: Runs type check, linting, and full test suite before publishing.
Test Suites
- Unit Tests (
test/unitTest.ts): Offline unit tests asserting pure formatting functions, response builders, identifier resolution helpers, and tool-definition invariants without network calls or credentials. - Protocol Tests (
test/protocolTest.ts): Protocol tests asserting the real MCP stdio handshake, server version and capabilities, tool schemas, and thetools/listcharacter budget against a spawned server process. - Contract Tests (
test/apiContractTest.ts): Contract tests asserting that every tool and op sends the expected outgoing HTTP requests (method, endpoint, query parameters, headers, and payload) and processes responses against an in-process mock Taiga HTTP server. - Integration Tests (
test/integration.ts): Live integration smoke tests asserting read-only tool operations against a real Taiga instance over stdio (skipped when credentials are not configured).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。