Local MCP server and supporting skill(s)

[nblumhardt]

This ticket tracks TODOs for the seqcli skills install, seqcli mcp install, and seqcli mcp run commands, which will ship v1s in the 2026.1 release.

Here's how to get started today, compatible with many agents, substitute claude for whatever you're using:

dotnet tool install --global seqcli --prerelease
seqcli config -k connection.serverUrl -v https://seq.example.com
seqcli config -k connection.apiKey -v 1234567890abc
seqcli mcp install -a claude
seqcli skill install -a claude

The install commands have --global variants, too.

Remaining issues and feature gaps (in loose priority order; these may not all make it in time for the 2026.1 release):

• Signal support; currently expecting to add a seq_list_signals tool, a signal (expression) parameter on search and query tools, and a supporting entry in the search-and-query-seq skill
• Add --native as an output format option on seqcli search|query|tail, and nudge agents towards this for full property display
• Review the seq-search-and-query skill to ensure it covers CLI usage equally well
• Fix query syntax error reporting for seqcli query native and JSON output
• Metrics support
• Investigate and measure alternative schema discovery approaches to increase reliability, e.g. including a "new schema" block in search result responses
• Improve the examples in SKILL.md to cover non-OTel log/span property naming conventions, e.g. perhaps coalescing @Resource.service.name with Application, re-test
• Investigate supporting a Streamable HTTP transport via seqcli mcp listen (or similar), so that datalust/seqcli can run server-side in a container alongside Seq (or on a Windows server); would likely need API key forwarding and some careful state management
• Script some evaluations
• ...

[gertjvr-ghd]

love the attention on this one, as I am using seq, but using the playwright cli to extract logs for me etc,

Also been using lots of mcps in the past but a recent trend I have noticed is that everything is moving away from mcps to cli + skill, they are way more context friendly over mcp and secure by nature over mcps

just a few examples

• https://playwright.dev/agent-cli/installation
• https://github.com/upstash/context7
• https://developer.atlassian.com/cloud/acli/guides/introduction/

[nblumhardt]

Thanks for dropping by, @gertjvr-ghd!

Aiming for implementation/UX parity between direct seqcli CLI use and the MCP server is a worthwhile goal, thanks for the feedback.

Right now, this might look like:

• Being careful to keep the search-and-query-seq skill agnostic of whether the agent is using the CLI, MCP, or some form of direct UI manipulation/computer use
• Feeding what we've learned about data representation back through the search and query commands, and
• Back-filling feature gaps like schema discovery, where we feel the CLI command set is lacking.

The big difference right now between seqcli search/seqcli query and the MCP server is how results are represented; the MCP server keeps all representation strictly in agent-friendly Seq syntax terms. Both commands already internally support output format switching (e.g. we'd add --native), and we could introduce a global output.nativeTypes config option that the CLI or skills could somehow suggest agents use.

Right now, the MCP server is the CLI via STDIO, so there isn't an additional security boundary involved; if anything, the MCP server makes it easier to keep Seq credentials hidden from agents, so there might be pros/cons on the security front. Hosting the MCP server as an HTTP endpoint opens up the other can of worms 👍

We also tracked the CLI → MCP move earlier this year, but recently we've seen an uptick in requests for an MCP server, so right now it's hard for us to pick which will be the more popular. Which works best probably also comes down to your environment and preferences, and also the relative implementation quality of each option for whatever you're targeting. Happily, it looks like most of what we're doing here will cover both.

Cheers!

[KodrAus]

Improve the examples in SKILL.md

I suspect we'll need to experiment a fair bit with the structure of the skill in general. Right now it's very reference-like, and we may see better results if most of that was moved to a reference file and the skill itself was more role or task centric.

As an example, I played with it using my local Qwen 3.6 27B setup. Using the skill as-is, it didn't actually read or follow most of it and failed to interact with Seq properly. I played with moving most of the contents of SKILL.md into a references/seq-docs.md, and made the SKILL.md instruct the agent to read and understand it, leaving the examples and gotchas in the skill, and appear to have more success with that approach.

[nblumhardt]

Let's close this umbrella issue, we can raise standalone tickets/discussions for improvements going forward.

[userAXC001]

Wanted to post this here in case anyone else runs into it. I was running into issues with the mcp not starting up in claude. Turns out, there is something about how seqcli mcp run uses its working directory at startup: it makes the CWD the .NET host's content root and recursively sets up file watching over it.

My repo is bind-mounted into a devcontainer from a Windows host (WSL2, so 9p). From the repo root it registers thousands of inotify watches over that mount, taking 20s+ (past 30s when cold). Claude Code's default MCP timeout is 30s, so it kept giving up. It's not hung, just slow: run the same command by hand and it answers initialize fine once it gets going. From an empty directory it's instant.

Two fixes:

• "timeout": 300000 on the server config, so Claude waits it out
• DOTNET_contentRoot= — drops the watch count to essentially zero, since this is an empty directory and starts immediately

Seq itself was fine (I think). Just wanted to provide a heads up