Repository navigation
Local MCP server and supporting skill(s) #459
Description
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 claudeThe 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_signalstool, asignal(expression) parameter on search and query tools, and a supporting entry in thesearch-and-query-seqskill - Add
--nativeas an output format option onseqcli search|query|tail, and nudge agents towards this for full property display - Review the
seq-search-and-queryskill to ensure it covers CLI usage equally well - Fix query syntax error reporting for
seqcli querynative 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.namewithApplication, re-test - Investigate supporting a Streamable HTTP transport via
seqcli mcp listen(or similar), so thatdatalust/seqclican 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
- ...
Activity
gertjvr-ghd commentedon Jun 1, 2026
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
nblumhardt commentedon Jun 1, 2026
Thanks for dropping by, @gertjvr-ghd!
Aiming for implementation/UX parity between direct
seqcliCLI use and the MCP server is a worthwhile goal, thanks for the feedback.Right now, this might look like:
search-and-query-seqskill agnostic of whether the agent is using the CLI, MCP, or some form of direct UI manipulation/computer usesearchandquerycommands, andThe big difference right now between
seqcli search/seqcli queryand 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 globaloutput.nativeTypesconfig 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 commentedon Jun 1, 2026
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.mdinto areferences/seq-docs.md, and made theSKILL.mdinstruct 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 commentedon Jun 11, 2026
Let's close this umbrella issue, we can raise standalone tickets/discussions for improvements going forward.
userAXC001 commentedon Aug 11, 2026
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 runuses 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:
Seq itself was fine (I think). Just wanted to provide a heads up