Menu
MCP Server · Threat-Modeled · Open Source

Linear MCP
Server

I built a read-only connection that lets an AI assistant summarize project information and help scope tasks in Linear.

6 MCP tools 93% eval pass rate Threat model documented TypeScript · Open source
At a glance 30-second read

What it is

I built six tools for sorting incoming work, scoping tasks, reviewing delivery data, and reporting status in Linear. Model Context Protocol, or MCP, connects the assistant to the tools. They read information and cannot change it.

What I explored

I explored how an assistant can help with planning while keeping project records unchanged.

Skills and tools

  • Production MCP server design (TypeScript SDK)
  • Evaluation harness (30 cases, 93% pass rate)
  • Documented threat model, prompt-injection defense
  • Per-tool cost analysis ($2.60/month at run-rate)
  • Read-only architecture — destructive ops never written
MCP SDK TypeScript Anthropic SDK Linear GraphQL Zod Vitest
What it is
Person You ask in plain words
→
AI Claude picks the right tool
→
This project linear-mcp six tools
→
Data Linear your team's work

An AI assistant can't reach into Linear on its own. linear-mcp is the connection, and it only reads.

6
focused tools
93%
live eval pass rate
$2.60
run cost / month
0
tools that can delete
01 — The Tools six jobs, six tools

One tool per job.

I organized the six tools around planning tasks: reviewing incoming work, defining scope, and reporting status.

Sort the inbox
triage_inbox
Ranks unread notifications by urgency and says what to do with each.
Fast model
Plan a task
scope_issue
Turns a rough task into steps, acceptance criteria, risks, and a size estimate.
Strong model
Find what's stuck
find_orphans
Surfaces stale, blocked, or unassigned work that has slipped through.
Fast model
Diagnose the team
audit_velocity
Reads recent sprints and explains whether the team is speeding up or slowing down.
Strong model
Write the update
compose_update
Drafts a Slack or email status update in the team's own voice.
Strong model
Recap the week
weekly_summary
A short, readable summary of what the team got done.
Strong model
02 — Proven test results

It works.

93%
Run live against a Linear workspace, an automated test of all six tools passed 28 of 30 checks. Most projects like this ship with no test at all.
weekly_summary · live run [ ACTUAL OUTPUT ]

“It was a productive week for the team, with five issues completed — all by Bruce Davis. …

… One thing worth flagging: every assigned issue this period belongs to Bruce Davis. Whether this reflects team size or a coverage gap, it’s worth keeping an eye on as the active work queue grows.”

weekly_summary produced that paragraph from sixteen tasks. The last line is something it noticed on its own.

03 — Security full threat model in the repo

Hard to misuse.

It acts on a company's data, so it was built to be hard to misuse. Three decisions do most of the work.

It can only read
No tool can delete or change anything. The code to delete a task was never written, so nobody can trick the AI into it.
The AI never owns the facts
Links and IDs come straight from Linear, added by my code. A made-up link is structurally impossible.
A bad actor finds nothing
Even if someone hides a sneaky instruction in a task, there is nothing harmful for the AI to reach.