Check your MCP / Claude Code config before you start a session
Read the client config first, then check each enabled server against its listing in our directory: tested version, tool count, credentials and recorded tool-list changes. The steps below use our own recorded figures and official client docs.
Open the client's MCP config in an editor, list the servers it will load, and walk six checks before you start a session: the server list, pinned versus floating versions, tool count and context cost, credentials and env vars, tool lists that changed since our baseline, and deprecated, archived or paused listings. Our MCP config checker matches a pasted config to the directory. Drift Check, a GitHub Action we are building, will compare a server's current tools/list result to our baseline for that listing.
1. See which servers are configured
File locations below are from the vendor docs, which we checked on 2026-11-17 UTC. We have not re-run the clients ourselves for this page; if a path differs in your version, prefer the vendor page.
| Client | Project file | User / default file | Docs (checked 2026-11-17 UTC) |
|---|---|---|---|
| Claude Code | .mcp.json in the project root (project scope, meant to be committed) | ~/.claude.json: local scope (the default; stored under that project's path, private to you) or user scope (all projects) | Claude Code MCP docs |
| Cursor | .cursor/mcp.json | ~/.cursor/mcp.json | Cursor MCP docs |
| Codex | .codex/config.toml (trusted projects only, per the docs we read) | ~/.codex/config.toml (default) | Codex MCP docs |
The same Claude Code MCP docs also describe claude mcp list and claude mcp get <name> for reading the configured list, and they note that general local settings use .claude/settings.json, which is a different file from the MCP server entries in ~/.claude.json. For the rest of the .claude directory, see the Claude Code directory docs.
Paste the config into our MCP config checker to match each server against the directory. The checker runs in the browser tab; do not paste secret values. As of scored.json in this build we list 386 MCP servers (385 in the directory, 1 paused). Test dates on those listings run from 2026-10-27 UTC to 2026-11-14 UTC. Formats the checker accepts, and how it matches npm, PyPI, remotes and repos, are described on that page. A longer comparison of Claude Code, Codex and Cursor config shapes is in the cross-client guide.
2. Pinned versus floating versions
A floating install such as npx -y some-package (no @version) can resolve a different version than the one we tested. For npm and PyPI listings whose install completed in our tests, we recorded a pinned command in our pinned-install records: 300 listings, compiled from those test records. The file was first added on 2026-11-11 (UTC); this build's 300 entries are every npm or PyPI listing with a passing install in this build (the tested_date values in the file run from 2026-10-27 UTC to 2026-11-14 (UTC)). Remotes, git checkouts, GitHub-release binaries and local-http listings are out of scope for that file, and one PyPI listing is skipped because its install did not pass.
Example pin from that file for Calculator: uvx mcp-server-calculator==0.2.1 (version 0.2.1, tested 2026-11-10 UTC).
Compare the version in your config to the listing's tested version. The config checker reports a pin mismatch when the version in the config is not the version we tested.
3. Tool count and context cost
Each enabled MCP server adds its tool definitions to the model's context. We publish a token_cost estimate (tiktoken cl100k_base over compact JSON of each tool's name, description and input schema). Of the 385 directory listings, 287 have an estimate; How much context do your MCP servers cost? and the lowest-context ranking use 285 of them. The other 2 are not ranked: Atlassian (Jira/Confluence) listed 0 tools without credentials; SQLite (npx) failed its latest published run (the stored estimate is from an earlier capture). 98 listings are pending (no capture).
Among those 285 ranked estimates, the lowest non-deprecated figure is Calculator at 47 tokens (1 tool, measured 2026-11-12 UTC, version 0.2.1). The highest is Microsoft 365 MCP (Softeria) at 238,259 tokens (202 tools, measured 2026-11-12 UTC, version 0.178.0). Add up the servers you plan to enable on the context cost guide.
The same estimate is stored on each baseline as token_estimate (tokenizer cl100k_base, schema_version 1), together with the tool names, input schemas and a SHA-256 description_hash for each tool. 288 listings have a baseline; 98 of the 386 listings do not, because they have no tool snapshot yet.
4. Auth and env vars that are needed
66 of the 385 directory listings needed credentials in our tests (an API key, or an OAuth sign-in). Those rows are listed on Needs an API key or sign-in. Test dates on the directory run from 2026-10-27 UTC to 2026-11-14 UTC.
We recorded the env var names 32 servers asked for in our test output (stdout or stderr). Those names are the ones we recorded in test output. We do not invent names for servers whose prompts we never saw; the config generator also makes no key-name claim for those. Check the listing page and the vendor's own docs for anything we did not record. The Claude Code MCP docs describe ${VAR} expansion in .mcp.json and ~/.claude.json, including the warning that appears in claude mcp list when a referenced variable is unset and has no default (Claude Code MCP docs, checked 2026-11-17 UTC).
5. Tool lists that changed since the baseline
When we re-run a server with the same arguments and the tool list differs, we record a tools_change on the listing. In this build that field is present on 21 listings; 6 of those are marked breaking. Every one of the 21 records is marked confirmed_by_rerun. The at dates on those records run from 2026-10-24 UTC to 2026-11-17 UTC. The public list is on Tool history.
Drift Check is a GitHub Action we are building for MCP server repositories; it is not released yet. The version we reviewed (commit 5fab254, committed 13 Nov 2026 at 05:59 UTC) starts the server, sends initialize and tools/list, and compares the result to the baseline. It reports a breaking difference when a tool was removed, a required field was added, a field type changed, or the server did not complete initialize and tools/list. It reports a non-breaking difference when a tool was added, an optional field was added, or a description changed. It does not call tools.
On 2026-11-17 UTC we computed 288 baseline documents from the tool snapshots in this build, one per listing with a snapshot (the listings with a tools_snapshot_at date in scored.json). They are not published yet. Their tool snapshot dates (tools_snapshot_at, UTC days) run from 2026-10-22 UTC to 2026-11-17 UTC (204 of them on 2026-11-14 UTC); 252 use transport stdio and 36 use streamable-http, and all 288 include a tested_version. 98 of the 386 listings have no baseline yet.
6. Deprecated, archived or paused servers
Of the 386 listings, 26 are deprecated or archived (19 with a registry deprecated flag, 21 with an archived GitHub repository; overlap makes 26 unique). 1 listing is paused and omitted from the directory. The paused listing is MailVeri MCP (mailveri-mcp). The public pause note on that page records that its source GitHub repository returned 404 (checked 2026-11-14 UTC). See all deprecated listings and deprecated servers that may still install. The config checker flags packages we mark deprecated or archived.
10 of the 288 baselines belong to deprecated or archived listings. A completed tools/list on an old package is not a maintenance promise; prefer a replacement when the registry message or our deprecation data names one.
Data sources
- Directory figures: scored.json in this build, 386 listings, test dates 2026-10-27 UTC to 2026-11-14 UTC.
- Pinned install commands: our pinned-install records, 300 listings in this build, covering every npm or PyPI listing with a passing install (listing tested dates in that file run from 2026-10-27 UTC to 2026-11-14 (UTC)).
- Env var names: names we recorded in test output for 32 servers; we do not invent names for the others.
- Tool-list changes: the
tools_changefield on scored.json (21 listings, 6 breaking,at2026-10-24 UTC to 2026-11-17 UTC; all 21 confirmed by a re-run) and Tool history. - Baselines: computed 2026-11-17 UTC from this build's tool snapshots, 288 documents (not published yet), tool snapshot dates 2026-10-22 UTC to 2026-11-17 UTC.
- Client file paths and
claude mcp list: Claude Code MCP docs, Cursor MCP docs, Codex MCP docs, checked 2026-11-17 UTC.
Questions about this guide: support@testedmcp.com.
Vendor docs checked 2026-11-17 UTC. Baselines computed 2026-11-17 UTC. Listing test dates 2026-10-27 UTC to 2026-11-14 UTC.