Linear MCP (multi-workspace, PAT auth)

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:

1. The OAuth flow is brittle across MCP client session boundaries — auth state expires when a session resumes mid-flow.
2. The localhost callback fails in some setups, forcing fragile paste-back flows.
3. 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-mcpcd ~/.claude/linear-mcppip install -e .

Claude Desktop one-click

Download the latest .mcpb from Releases and double-click.

Configure

1. 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.

2. Create ~/.claude/linear-mcp/admin.env (chmod 600):

LINEAR_WORKSPACES=personal,workLINEAR_PRIMARY_WORKSPACE=personal
LINEAR_PAT_PERSONAL=lin_api_xxxxxxxxxxxxxxxxxxxxxxxxxxxxLINEAR_PAT_WORK=lin_api_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
LINEAR_LABEL_PERSONAL=PersonalLINEAR_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

3. 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

4. Restart Claude Code. healthcheck should return ok: true per 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.

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 p1linear-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

Core entities

Webhooks (v0.2)

Notifications / inbox (v0.2)

Attachments (v0.2)

Issue relations (v0.2)

Agent sessions (v0.2)

Search (v0.2 — replaces v0.1's broken search_documentation)

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.