Best MCP Servers for Jan (2026)
Jan runs a local model, so every tool you add competes for the same context — here are the six MCP servers that earn their slot, and the ones that don't.
The best MCP servers for Jan in 2026 are the ones that stay small, run locally, and pay for their tool slots: GitHub's official server, GitMCP, Anyquery, Octocode, Graphlit, and a package-registry lookup. Jan is a local AI app, so your model — not a frontier API — has to route every tool call, which makes discipline matter more here than in Cursor or Claude Desktop.
This is an opinionated shortlist, not a catalogue. Every server below is one I'd actually install, with the exact config Jan wants and a note on what to skip. If you're new to the protocol itself, start with what an MCP server is; if you want the cross-client picks, the best MCP servers list is broader.
How Jan handles MCP (and why the budget is tighter)
Jan adds MCP servers through Settings > MCP Servers, then the + in the top-right of the MCP box. The form is small: Transport (STDIO, HTTP, or SSE), Command / URL, Args, and Env for keys. You'll need Node.js and/or Python on the machine, and a model with tool-calling turned on — a non-tool model will simply ignore the servers.
Two Jan-specific settings change the calculus. There's an Allow All MCP Tool Permissions toggle that auto-approves every call (convenient, and exactly the setting to leave off while you're testing a write-capable server). And Smart Routing narrows which servers Jan queries for tools each turn, which is the closest thing you get to a safety valve on tool count.
That safety valve matters because the tool budget is real. Around 90% of MCP servers run locally over stdio, and most clients start degrading tool selection past roughly 40 active tools. The median server exposes about 10 tools, so five average servers already crowds the budget — and a local model is worse at picking from a long list than a frontier one. The tool-limit math was written for Cursor but the arithmetic is identical in Jan.
The six servers worth adding
Here's the shortlist, ranked by how often I'd reach for each one, with tool footprint in mind.
| Server | Transport | Why it earns the slot | Watch for |
|---|---|---|---|
| GitHub MCP Server | Local Docker or remote | Official server for repos, issues, PRs, Actions | Tool-heavy — enable a toolset |
| GitMCP | Remote (free) | Turns any repo into live docs, cuts hallucinations | Read-only by design |
| Anyquery | Local stdio | SQL over files, DBs, and 40+ apps | One SQL interface, small footprint |
| Octocode | Local stdio | Code research with GitHub + FS + LSP | Overlaps GitHub; pick one primary |
| Graphlit | Remote | RAG over Slack, Drive, Jira, Linear | Needs a Graphlit account |
| Package Registry | Local stdio | Up-to-date NPM/PyPI/crates/Go lookups | Tiny, keep it on |
GitHub, done officially
Start with the GitHub MCP Server if you touch code at all. It's GitHub's own first-party server — repos, issues, PRs, and Actions — and "official" buys you the boring wins: OAuth done right, kept current, security-reviewed. The catch is it's tool-heavy, so scope it to a single toolset rather than enabling everything, or it eats your budget alone.
GitMCP to stop the model guessing at APIs
Add GitMCP when your model keeps hallucinating a library's API. It's a free remote server that turns any GitHub repo into a live documentation source the model can read instead of inventing. It's read-only by design, which is exactly what you want here — retrieval with no write blast radius. As a remote server it costs you nothing locally.
Anyquery for everything with a SQL shape
Anyquery is the single most useful install on this list. It exposes SQL over local files, real databases, and 40+ apps — GitHub, Notion, Chrome, Todoist — behind one query interface. One server, a handful of tools, and you've replaced three narrow connectors. For a local-model setup where tool count is the constraint, that consolidation is the whole point.
Octocode vs. Octopus of connectors
Octocode is the pick if code research is your main job in Jan: it wires GitHub, the local filesystem, and LSP intelligence into one assistant that reads code like a staff engineer. It overlaps the GitHub server, so don't run both at full tilt — choose Octocode as your primary for deep code work, or the GitHub server for repo/PR management, not both maxed out.
Graphlit when your knowledge lives in five apps
Reach for Graphlit only if your context is scattered across Slack, Google Drive, Jira, Linear, and Discord. It ingests and RAGs over all of them through the Graphlit platform. It's a remote server and needs an account, so it's the one I'd add last — worth it for team knowledge, overkill for a solo coding setup.
Package Registry as the cheap, always-on utility
Package Registry is small enough to just leave on. It searches NPM, PyPI, crates.io, NuGet, and Go registries plus GitHub Security Advisories for current versions and CVEs. Local, tiny footprint, and it stops the model from suggesting a package version that shipped a year ago.
A minimal Jan config that works
Add Anyquery first — it's local, small, and immediately useful. In Settings > MCP Servers > +, fill the form like this:
- Transport: STDIO
- Command:
npx - Args:
-y anyquery mcp - Env: leave empty
Save, confirm the green active indicator, then add GitMCP as a HTTP transport with its endpoint URL — a remote server means nothing to install. Turn on Smart Routing once you're past two or three servers so Jan isn't handing your local model every tool on every turn.
If you'd rather generate a validated snippet than hand-type, the config generator produces the exact block, and the general how to add an MCP server walkthrough covers the gotchas.
What to skip
Skip anything that runs arbitrary code or SQL without a guardrail. A server with an unguarded execute_query or shell tool isn't capable — it's a bigger blast radius, and a local model with Allow All Permissions on is exactly the setup that fires a DELETE with no WHERE. Prefer read-only defaults; see what actually matters for MCP security.
Skip redundant connectors. Running the GitHub server, Octocode, and Anyquery's GitHub table all at once is three overlapping tool sets fighting for the same budget. Pick one primary per job.
And skip the mega-servers that ship 40-plus tools unless you'll use most of them — a single one of those can blow Jan's tool budget by itself. If you're building specifically for autonomous coding, the best servers for coding agents breakdown goes deeper on which tools survive contact with a real agent loop.