I’ve been using Linear as the project and issue tracking tool on a new project.
No wait, that’s not quite right. My AI coding agents have been using Linear as the project and issue tracking tool on a new project. I’ve opened Linear’s web interface...twice? And I’m pretty sure I’ve logged into the mobile client. But Claude and friends? They use Linear every day.
To date, I’ve been using the first-party Linear MCP and a third party one that I’d found before Anthropic started publishing an "official" Linear plugin in partnership with Linear.
It works great.
There’s just one problem. The official Linear MCP has 25 tools, using a total of 19,659 tokens of context on every single session. The third-party MCP is a little slimmer at 17k and change. But that…
I’ve been using Linear as the project and issue tracking tool on a new project.
No wait, that’s not quite right. My AI coding agents have been using Linear as the project and issue tracking tool on a new project. I’ve opened Linear’s web interface...twice? And I’m pretty sure I’ve logged into the mobile client. But Claude and friends? They use Linear every day.
To date, I’ve been using the first-party Linear MCP and a third party one that I’d found before Anthropic started publishing an "official" Linear plugin in partnership with Linear.
It works great.
There’s just one problem. The official Linear MCP has 25 tools, using a total of 19,659 tokens of context on every single session. The third-party MCP is a little slimmer at 17k and change. But that’s still nearly 10% of the full context window. For every context window.
This morning, after breakfast, I sat down and started chatting with Claude about what a better Linear tool might look like. We discussed just using a unix commandline tool. We discussed using a unix commandline tool + a skill. We discussed a Skill + a single-tool MCP client that was just a pure GraphQL client.
I asked Claude to read my blog post on MCP design.
We ended up with something nice and streamlined. It totals out at 975 tokens, including instructions for how to learn more about how to use the tool. I ended up talking Claude into making the MCP fully self-documenting by including a ‘help’ action.
We ended up compromising on tool design. Claude really thought that it would be fine always reading the instructions and just using raw GraphQL for everything. I overruled it and decided that the most common operations (working with tickets) merited first-class actions. Everything else is GraphQL backed up by the ‘help’ action.
It’s called Streamlinear. Ultimately, I’m responsible for the name. I didn’t say no. I asked Claude to come up with a list of punny names. Everything else it suggested was being used for a Linear client already.
I asked Claude to talk about the new tool and what it’s like:
This is the tool loadout for the ‘official’ MCP:
And this is what Streamlinear looks like:
Give it a spin and let me know how it goes.