linear-mcp
Multi-workspace Linear MCP server using Personal API Keys instead of OAuth. Provides tools for managing issues, projects, teams, and more across multiple workspaces.
README
linear-mcp
<!-- mcp-name: io.github.adelaidasofia/linear-mcp -->
Multi-workspace Linear MCP server with Personal API Key auth. Drop-in replacement for the official OAuth-only Linear MCP at mcp.linear.app/mcp.
Why this exists. The official Linear MCP is OAuth-only and single-workspace per instance. PAT auth plus persistent token storage plus multi-workspace routing kills three failure modes:
- The OAuth flow is brittle across MCP client session boundaries — auth state expires when a session resumes mid-flow.
- The localhost callback fails in some setups, forcing fragile paste-back flows.
- One MCP instance per workspace doubles config + OAuth dances.
This server replaces both with one entry. One install, N workspaces, never an OAuth dance again.
Install
From PyPI
pipx install adelaidasofia-linear-mcp
From source
git clone https://github.com/adelaidasofia/linear-mcp ~/.claude/linear-mcp
cd ~/.claude/linear-mcp
pip install -e .
Claude Desktop one-click
Download the latest .mcpb from Releases and double-click.
Configure
-
Generate one Personal API Key per workspace at https://linear.app/settings/account/security. You must be logged into each workspace separately while generating the key for that workspace.
-
Create
~/.claude/linear-mcp/admin.env(chmod 600):
LINEAR_WORKSPACES=personal,work
LINEAR_PRIMARY_WORKSPACE=personal
LINEAR_PAT_PERSONAL=lin_api_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
LINEAR_PAT_WORK=lin_api_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
LINEAR_LABEL_PERSONAL=Personal
LINEAR_LABEL_WORK=Work
Aliases are arbitrary — name them whatever helps you route (acme,beta, team1,team2, home,client).
chmod 600 ~/.claude/linear-mcp/admin.env
- Register in Claude Code (user scope, so it loads in every project):
claude mcp add -s user linear-mcp python3 -m linear_mcp.server
Or, with pipx install:
claude mcp add -s user linear-mcp linear-mcp
- Restart Claude Code.
healthcheckshould returnok: trueper workspace.
Substrate-layer enforcement (v0.3)
Three server-side checks that protect issue quality without depending on any markdown rule file or client-side memory. Apply only to WRITE paths; reads stay unconstrained.
| Layer | What it does | Bypass env var |
|---|---|---|
[source:] first-line check |
save_issue / save_project reject CREATE calls whose description (or project content) does not start with [source: <canonical-key>]. UPDATE calls (id passed) skip the check so legacy backfills remain unblocked. |
LINEAR_MCP_SKIP_SOURCE_CHECK=1 |
| Idempotency check | Before any CREATE, the server runs searchIssues / searchProjects for [source: <key>] and refuses to create a duplicate. The error names the existing identifier + UUID so the caller can update in place. |
LINEAR_MCP_SKIP_IDEMPOTENCY=1 |
auth_phrase on bulk_save_issues |
bulk_save_issues now requires auth_phrase ∈ {"go", "yes do it", "confirmed", "execute", "go cancel", "go update"} (case-insensitive). Mass-modification surface stays explicit. |
(no bypass — surface the phrase to the operator) |
Canonical-key examples:
[source: 🍄 Mycelium AI/📝 Meeting Notes/2026-05-22 - sync.md][source: ⚙️ Meta/Decisions/2026-05-23-merger-public-comms.md][source: linear-kickoff:sweep-myc-p1][source: ~/.claude/linear-mcp/BUILD_PROMPT_V03.md]
Fast issue execution
linear-exec is the short path for starting real work from a Linear issue.
It resolves the issue, checks blocking relations, scans same-project siblings
for obvious scope overlap, infers the repo when possible, and can create the
standard claude-dev-worktree branch.
Dry-run a specific issue:
linear-exec execute MYC-150 --workspace mycelium
Start the work after reviewing the preflight:
linear-exec execute MYC-150 --workspace mycelium --repo memory-runtime-pro --go
Find the next unblocked P1 issue in a workspace:
linear-exec sweep mycelium p1
linear-exec sweep mycelium p1 --go --repo memory-runtime-pro
The command refuses to proceed when an upstream blocks relation is still
incomplete unless --force is passed. --no-state-update and --no-worktree
let agents use only the parts of the preflight they need.
Tool surface (v0.3 — 57 tools + 3 prompts)
Every tool takes an optional workspace parameter (the alias from LINEAR_WORKSPACES). Omit it to use LINEAR_PRIMARY_WORKSPACE.
Meta
| Tool | Purpose |
|---|---|
list_workspaces |
Show configured workspaces and primary |
healthcheck |
Verify each PAT + surface remaining rate-limit budget |
Core entities
| Tool | Purpose |
|---|---|
list_teams / get_team |
Teams (with inline workflow states) |
list_users / get_user |
Users (me resolves to PAT owner) |
list_projects / get_project / save_project |
Projects (v0.3: save_project enforces [source:] on content + idempotency on CREATE) |
list_initiatives / get_initiative / save_initiative |
Initiatives |
list_issues / get_issue / save_issue / bulk_save_issues |
Issues (id or ONDE-123; bulk uses issueBatchUpdate; v0.3: save_issue enforces [source:] + idempotency on CREATE; bulk_save_issues requires auth_phrase) |
list_cycles |
Cycles |
list_milestones / get_milestone / save_milestone |
Project milestones |
list_issue_statuses / get_issue_status |
Workflow states |
list_issue_labels / create_issue_label |
Labels |
list_comments / save_comment |
Comments |
list_documents / get_document / save_document |
Documents |
save_status_update |
Post a project status update |
Webhooks (v0.2)
| Tool | Purpose |
|---|---|
list_webhooks / get_webhook |
Inspect subscriptions |
create_webhook / update_webhook |
Manage subscriptions |
delete_webhook |
Destructive — draft+confirm |
Notifications / inbox (v0.2)
| Tool | Purpose |
|---|---|
list_notifications / get_notification |
Inbox read |
notifications_unread_count |
Top-of-mind counter |
mark_notification_read / mark_all_notifications_read |
Triage |
archive_notification |
Sweep |
Attachments (v0.2)
| Tool | Purpose |
|---|---|
list_attachments / get_attachment |
Per-issue reads |
attachments_for_url |
Reverse lookup: which issues link to this URL? |
link_url_to_issue |
Attach any URL to an issue |
delete_attachment |
Destructive — draft+confirm |
Issue relations (v0.2)
| Tool | Purpose |
|---|---|
list_issue_relations |
The blocks/duplicate/related graph |
create_issue_relation / delete_issue_relation |
Manage the graph |
Agent sessions (v0.2)
| Tool | Purpose |
|---|---|
list_agent_sessions / get_agent_session |
Linear's first-class agent surface |
create_agent_session_on_issue / create_agent_session_on_comment |
Spawn |
Search (v0.2 — replaces v0.1's broken search_documentation)
| Tool | Purpose |
|---|---|
search_issues / search_documents / search_projects |
Full-text per entity type |
semantic_search |
Workspace-wide semantic across all entities |
MCP prompts (v0.2)
Available as slash commands in MCP clients that surface prompts:
/triage-issue— full triage pass: classify, label, prioritize, assign, link duplicates/project-status— draft a weekly status update from current Linear state/inbox-sweep— sweep today's notifications, propose actions, archive what's handled
Multi-workspace usage
Switch workspaces inline:
list_teams(workspace="work")
save_issue(workspace="personal", title="Ship", team_id="...")
Without workspace, the primary is used.
Auth
Linear PATs use header Authorization: <key> (no Bearer prefix). Each PAT is scoped to one workspace and grants access only to data the owning user can see. There is no shared org token.
Rate limit: 2500 requests/hour per token (verified against live API 2026-05-23). The server passes Linear's Retry-After header through on 429 and surfaces remaining budget via healthcheck.
Safety
Read tools and routine writes (create/update issues, comments, labels, status updates) pass through. Destructive ops (delete_webhook, delete_attachment) use the draft+confirm pattern: the first call stages the change and returns a draft_id + preview of what will happen; the second call (with confirm_draft_id) commits. Drafts expire after 1 hour (override with LINEAR_MCP_DRAFT_TTL_SECONDS).
Every tool call appends one JSONL line to ~/.claude/linear-mcp/audit.log (override with LINEAR_MCP_AUDIT_LOG_PATH, disable with LINEAR_MCP_AUDIT_LOG=false). Tokens are stripped from audit records.
healthcheck surfaces each PAT's remaining rate-limit budget (X-RateLimit-Requests-Remaining + X-Complexity-Remaining) per workspace, so agents can self-throttle without making a separate observability call.
Related MCPs
- adelaidasofia/slack-mcp — multi-workspace Slack with draft+confirm
- adelaidasofia/whatsapp-mcp — WhatsApp via whatsmeow + vault export
- adelaidasofia/imessage-mcp — iMessage chat.db + vault export
- adelaidasofia/github-mcp — GitHub PR/issue/release
License
MIT.
Built by Adelaida Diaz-Roa. Full install or team version at diazroa.com.
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。