https://mcp.mzizi.dev/mcp, and the agent skills. Each section gives the
vendor’s own config format, checked against that vendor’s documentation and linked, so you can
check it again. Each section says the date it was checked (20 October or 3 November 2026).
No client here was run. Every format is documented by the vendor, not tested in the tool;
the one thing tested from outside is the server, with curl, in Check the server
yourself. Formats move; where a vendor’s page and this one
disagree, the vendor’s page wins.
What you connect
What needs sign-in. The
fundi operation, and filing an issue with mzizi_feedback. They act as your Mzizi console user. Called without sign-in, fundi answers 200 with instructions to sign in, and a mzizi_feedback issue (without target: "docs") answers 401 with an OAuth challenge (WWW-Authenticate). A client that runs OAuth starts its sign-in flow from that 401. A client that does not can connect to /mcp/signin, which answers 401 until you sign in. The server supports dynamic client registration, so a client that asks how to register should register automatically. It does not publish a client ID metadata document.
How the skills arrive. One of three ways, depending on the client:
- Natively, from the
mziziplugin. Claude Code installs the skills and connects the MCP server in one step. Codex and ChatGPT will through the same plugin once mzizi-registry#520 reaches the registry’smain; until thencodex plugin marketplace addfinds nothing. - Natively, from npm.
npm install -D @nyuchi/mzizi-skills, thennpx skills experimental_sync, links the skills into.agents/skills/, which Codex, Cursor, VS Code, Gemini CLI, Antigravity, Zed, OpenCode and Devin Desktop all read. Cline reads.cline/skills/and.claude/skills/, and Kiro.kiro/skills/, not.agents/skills/. - As instruction files. For formats that cannot hold a skill, the
@nyuchi/mzizi-skillspackage’sadapters/folder holds files generated from the skills: Cursor.mdcrules, Antigravity and Devin Desktop rules, and one section to append toAGENTS.md,GEMINI.mdor.github/copilot-instructions.md. Each carries the skill’s own description and says where to load the full skill; none copies a skill body, because Antigravity cuts a rule file at 24,000 bytes and Codex stops readingAGENTS.mdat 32 KiB, and a skill is up to 59 KB. - Over MCP. Any client connected to the server can call
mzizi_executewithoperation: "get_skills": with no arguments it lists the skills and their categories,namereturns one in full, andcategory(dev,design,languageorbuild) returns every skill in it in full. Nothing to install. This is the route for Claude.ai, ChatGPT without the plugin and the APIs.
At a glance
Skills, instructions and plugins by client
Read from each vendor’s documentation on 3 November 2026; none run in the tool. “Not documented” means the page checked does not say, not that the tool lacks it.
Skill frontmatter limits. Every Mzizi skill uses only
name and description, the portable
core of the Agent Skills specification: name at most
64 characters, lowercase letters, digits and hyphens, matching the folder; description at most
1,024 characters, and these have no angle bracket. Claude Code truncates a description plus its
when_to_use at 1,536 characters; Cline, VS Code, OpenCode, Kiro and Zed state the 1,024 limit
(Zed loads a longer one with a warning); Codex, Cursor, Antigravity and Devin Desktop state no
limit. Claude.ai uploads and the Skills API reject any frontmatter key outside name,
description, license, compatibility, metadata and allowed-tools, and the skills
carried user-invocable: true until the next @nyuchi/mzizi-skills release removes it. Read
on 3 November 2026 from Create custom skills and
Extend Claude with skills.
Claude
Claude.ai
Add a custom connector: in Customize > Connectors, choose + Add, then Add custom connector, name itMzizi, and enter
https://mcp.mzizi.dev/mcp. Under Authentication, choose No sign in. On Team and
Enterprise plans an owner adds it once in
Organization settings > Connectors (Add,
then Custom, then Web), and each member connects to it.
For the sign-in operations, add a second connector at https://mcp.mzizi.dev/mcp/signin, choose a
sign-in option under Authentication, and Register automatically under OAuth client.
Enable it in a conversation from +, then Connectors. Skills arrive over MCP: ask Claude
to call mzizi_execute with the get_skills operation. Free plans are limited to one custom connector. Claude connects to
the server from Anthropic’s cloud, not from your device; mcp.mzizi.dev is public, so that works.
You can also upload a skill: zip one skill folder (the folder itself at the top of the ZIP, its
name matching name in SKILL.md) and add it under Customize > Skills (code execution
must be on). Upload accepts only the portable frontmatter keys, so use @nyuchi/mzizi-skills
from the release after 0.17.0, which removes user-invocable.
Checked on 3 November 2026 against Get started with custom connectors using remote
MCP,
Using skills in Claude
and Create custom skills. The connector article does
not name the transport; the server answers Streamable HTTP, as below.
Claude Desktop
The custom connector above works in Claude Desktop too; it is configured on your Claude account, not on the machine. To run the server on your machine instead, over stdio, open Settings from the Claude menu, then Developer, then Edit Config, and add it toclaude_desktop_config.json (~/Library/Application Support/Claude/ on macOS,
%APPDATA%\Claude\ on Windows):
fundi operation. Skills arrive over MCP.
Checked against the custom connectors article above and Connect to local MCP
servers.
Claude Code
The plugin carries the twelve skills and connects the server:claude plugin marketplace add mzizi-dev/mzizi-registry and
claude plugin install mzizi@mzizi. It lives in the plugin/ directory of
mzizi-dev/mzizi-registry; see
the skills bundle for which skills version it
carries.
Or the server alone, with skills over MCP:
--scope user to add it to the current project only. On 20 October 2026,
claude mcp list showed mzizi as connected and mzizi-signin as needing authentication,
and the plugin installed from a local copy of the registry listed seven skills and the mzizi
server, connected.
Checked against the Claude Code MCP,
plugin manifest and
marketplace references, and
on 3 November 2026 against its MCP,
skills and
marketplace pages: claude mcp add --transport http is the documented form for a remote server, and a JSON entry with a url and
no type is read as stdio and skipped. A project’s .mcp.json asks for approval before it
loads.
The Claude API
The Messages API’s MCP connector reaches a remote server for you. It needs themcp-client-2025-11-20 beta header, a server in mcp_servers and an mcp_toolset that names
it. Use the model you would otherwise call:
authorization_token is needed for /mcp. Only tool calls are supported, so skills arrive
through the get_skills operation. The fundi operation needs a token from the sign-in flow in
authorization_token, against /mcp/signin; the connector does not run that flow for you.
Checked against MCP connector.
Not run here: it needs an API key.
OpenAI
ChatGPT
Go to ChatGPT Plugins, select +, then Add custom MCP server. Name itMzizi, enter https://mcp.mzizi.dev/mcp under Connection, set no
authentication, review the warning, and select Create as a plugin. Install it, start a new
conversation, and type @ to pick it. Account and workspace policies decide whether you can
add a custom server. Skills arrive over MCP.
The mzizi plugin also packages the skills for ChatGPT, in the portable layout below. The
ChatGPT desktop app reads a repository’s .agents/plugins/marketplace.json, but installing it
from GitHub has not been tested here.
Checked against Connect and test your
plugin.
OpenAI Codex
The plugin. The registry’splugin/ carries a portable Agent Plugins manifest
(plugin/plugin.json and plugin/mcp.json) beside the Claude one, and a repo marketplace at
.agents/plugins/marketplace.json:
mzizi from the Plugins directory in the ChatGPT desktop app. OpenAI’s plugin
guide documents codex plugin marketplace add, list, upgrade and remove for the
marketplace source, and the desktop app for installing from it; it documents no CLI install
command.
Start a new session after installing. The plugin’s mcp.json connects /mcp over
streamable-http.
The Codex and ChatGPT manifests arrive in
mzizi-dev/mzizi-registry with its pull request
#520. Until that reaches the registry’s main branch, codex plugin marketplace add finds no
plugin there: use the server alone, below.~/.codex/config.toml (or a trusted project’s
.codex/config.toml), which the Codex CLI, the IDE extension and the ChatGPT desktop app share:
codex mcp add mzizi --url https://mcp.mzizi.dev/mcp writes the first table for you.
Skills. Codex reads skills from .agents/skills/ in the repository and from
~/.agents/skills/, which is where npx skills experimental_sync links them; otherwise they
arrive over MCP. Codex also reads AGENTS.md before any work, so a repository can tell it to
load the skills:
plugin.json is the portable format and .codex-plugin/plugin.json a compatibility layout;
skills are discovered in skills/ and the MCP server in mcp.json, typed
streamable-http), Model Context Protocol
(url, bearer_token_env_var, http_headers; --url is in its codex mcp add example),
Build skills and Custom instructions with
AGENTS.md. The manifests validate
against the Agent Plugins schemas at agent-plugins.org/schemas/1.0.0. Not run here: the
Codex CLI was not installed.
The OpenAI API and Agents SDK
In the Responses API, the server is atype: "mcp" tool. Use the model you would otherwise
call:
/mcp, the live annotations (read from tools/list on 17 November 2026) are these. mzizi_search is readOnlyHint: true. mzizi_execute is readOnlyHint: false and destructiveHint: true. mzizi_feedback is readOnlyHint: false. Keep "require_approval": "never" only if you accept that. Narrowing is the safer choice: set allowed_tools: ["mzizi_search"], or require approval for mzizi_execute and mzizi_feedback. For the sign-in operations, point server_url at /mcp/signin and pass an access token in authorization.
In the Agents SDK for Python, the same object goes in a hosted tool, and the Responses API makes
the connection:
MCPServerStreamableHttp with
params={"url": "https://mcp.mzizi.dev/mcp"} and pass it in mcp_servers. Skills arrive over
MCP.
Checked against MCP servers
and the Agents SDK’s MCP page. Not run
here: it needs an API key.
Z.ai’s GLM Coding Plan
Z.ai’s coding plan runs inside coding tools rather than as its own client. Its documented Claude Code setup points Claude Code at Z.ai withANTHROPIC_BASE_URL set to
https://api.z.ai/api/anthropic and your Z.ai key in ANTHROPIC_AUTH_TOKEN, in
~/.claude/settings.json, or through Z.ai’s helper, npx @z_ai/coding-helper. Plugins and MCP
servers are Claude Code’s own, so connect Mzizi exactly as in Claude Code: the
mzizi plugin, or claude mcp add. Z.ai’s own MCP servers use the same commands
(claude mcp add -s user …).
Z.ai documents other tools too. In OpenCode, Cursor or another client, follow that client’s
section here.
Checked against Z.ai’s Claude Code and
Vision MCP server pages. Not run here: it
needs a Z.ai key.
Editors and other CLIs
OpenCode
Inopencode.json (the project’s, or your global config):
remote entry at /mcp/signin. OpenCode starts OAuth itself
when a server answers 401, or run opencode mcp auth <name>. Skills: OpenCode reads
.agents/skills/ and ~/.agents/skills/, so the npm route works; otherwise over MCP.
Checked against MCP servers and
Agent Skills.
Cursor
In.cursor/mcp.json in the project, or ~/.cursor/mcp.json for every project:
mzizi-signin entry at /mcp/signin for the sign-in operations; Cursor runs the OAuth flow.
Skills: Cursor reads .agents/skills/ (and .cursor/skills/, and Claude’s and Codex’s skill
folders), so the npm route works; otherwise over MCP.
Checked against Cursor’s MCP and
Skills pages.
VS Code and GitHub Copilot
Run MCP: Add Server from the Command Palette, choose HTTP, enterhttps://mcp.mzizi.dev/mcp and the ID mzizi. Or from a shell:
.mcp.json. Its portable format, which
GitHub Copilot CLI also reads from ~/.copilot/mcp-config.json, uses mcpServers:
.vscode/mcp.json uses servers instead of mcpServers, with the same entry.
Skills: VS Code reads .agents/skills/, .github/skills/ and .claude/skills/, so the npm
route works; otherwise over MCP.
Checked against Add and manage MCP
servers, the MCP
configuration reference
and Agent skills.
Devin Desktop (formerly Windsurf)
Windsurf’s documentation now redirects to Devin Desktop, whose agent, Devin Local, shares its harness with the Devin CLI. MCP servers live in.devin/mcp_config.json (the project, shared
through version control), .devin/mcp_config.local.json (yours, ignored by git) or
~/.config/devin/mcp_config.json (every project). The page supports stdio, Streamable HTTP
and SSE, with OAuth for each. The Devin CLI’s configuration page names a remote server’s key
url, with transport set to "http" (the default) or "sse":
serverUrl in mcp_config.json, is not on the current page; if
your install still reads it, use that. Skills: Devin Desktop reads .devin/skills/,
.agents/skills/ and the legacy .windsurf/skills/, so the npm route works. Rules:
AGENTS.md and .devin/rules/*.md (with a trigger: model_decision and a description), and
the legacy .windsurf/rules/ is imported.
Checked on 3 November 2026 against Devin Desktop MCP,
skills and rules and the Devin CLI MCP
configuration. The two pages
do not show the same example, so the JSON above is the CLI page’s, applied to the Desktop paths
its own page names.
Antigravity CLI and IDE
Google moved free individual access and Google AI Pro and Ultra from Gemini CLI to Antigravity CLI (agy) on 18 June 2026. Antigravity has no command that adds a server: edit
~/.gemini/config/mcp_config.json (every project) or .agents/mcp_config.json (the
workspace). A remote server is Streamable HTTP only, under serverUrl, and the page says the
legacy HTTP+SSE transport and httpUrl are not supported:
/mcp opens a manager that lists servers and reloads the config. OAuth runs
through dynamic client registration, so a /mcp/signin entry needs no extra config.
Skills: Antigravity reads .agents/skills/ in the workspace and each parent up to the repo
root, and ~/.gemini/config/skills/ globally, so the npm route works. Instructions:
AGENTS.md or GEMINI.md, and .agents/rules/*.md with a trigger (model_decision needs a
description); a file with a missing or invalid trigger is silently discarded, and each file
is cut at 24,000 bytes.
Checked on 3 November 2026 against MCP,
Agent skills and
Rules.
Gemini CLI
Gemini CLI is now maintained for enterprise Gemini Code Assist workflows; individual and Google AI Pro and Ultra users are on Antigravity CLI.~/.gemini/settings.json; without --scope user it writes the project’s
.gemini/settings.json. By hand, a Streamable HTTP server takes httpUrl (url is the SSE
endpoint):
.agents/skills/ and .gemini/skills/ in the workspace, and the same
under ~, so the npm route works; otherwise over MCP. GEMINI.md is the instruction file, and
"context": {"fileName": ["AGENTS.md", "GEMINI.md"]} in settings.json makes it read
AGENTS.md too.
Checked on 3 November 2026 against MCP servers,
Agent skills, GEMINI.md
and the project’s repository notice.
Zed
In Zed’ssettings.json, or Settings > AI > MCP Servers > Add Server > Add Remote Server:
Authorization header starts the standard MCP OAuth flow, so a /mcp/signin
entry prompts you to sign in. Skills: Zed reads .agents/skills/ in a trusted worktree and
~/.agents/skills/ globally, so the npm route works. Its skills page does not mention
AGENTS.md or rules files.
Checked on 3 November 2026 against Model Context Protocol and
Skills.
Cline
Incline_mcp_settings.json (~/.cline/data/settings/ for the CLI; in the IDE extension, MCP
Servers, Configure, Configure MCP Servers). Set the type: without it Cline falls back
to the legacy SSE transport.
Authorization header and the CLI’s
Authorize OAuth action, so for the sign-in operations add /mcp/signin and use that action.
Skills: Cline reads .cline/skills/, .clinerules/skills/ and .claude/skills/, and
~/.cline/skills/ globally; its page does not list .agents/skills/, so link the skills into
.cline/skills/. Rules: .clinerules/ and AGENTS.md.
Checked on 3 November 2026 against Adding and configuring
servers,
Rules and
Skills.
Kiro CLI (formerly Amazon Q Developer CLI)
Amazon Q Developer CLI is now Kiro CLI. In.kiro/settings/mcp.json (the workspace) or
~/.kiro/settings/mcp.json (every workspace; the two merge, and the workspace wins):
/mcp auth signs in again. Skills: Kiro reads .kiro/skills/ in the workspace (IDE, CLI
and web) and ~/.kiro/skills/ globally (IDE only), so link the skills into .kiro/skills/.
Kiro’s page states the same 64 and 1,024 character limits. It does not mention AGENTS.md.
Checked on 3 November 2026 against MCP configuration
and Skills.
JetBrains AI Assistant and Junie
AI Assistant. Settings > Tools > AI Assistant > Model Context Protocol (MCP) takes amcpServers JSON with a url, and lists Streamable HTTP as a supported transport. The page
does not show headers, and it also describes bundled skills under Settings > Tools > AI
Assistant > Skills.
.junie/mcp/mcp.json (the project) and ~/.junie/mcp/mcp.json (yours),
the same mcpServers shape, with url and headers for a remote server. Its page does not
say whether that transport is Streamable HTTP, so if Junie cannot connect, run the stdio
server instead: npx -y @nyuchi/mzizi-mcp. Junie reads AGENTS.md in the project root.
Checked on 3 November 2026 against AI Assistant: MCP
and Junie’s MCP configuration.
Junie’s older help page said it supported stdio only; the current one does not.
Any other MCP client
Most clients take one of the shapes above: a Streamable HTTP URL,https://mcp.mzizi.dev/mcp,
or the stdio command npx -y @nyuchi/mzizi-mcp. The server is listed in the
MCP Registry
as io.github.mzizi-dev/mzizi-mcp, so a client that installs from the registry finds both.
Check the server yourself
Any client that sends theAccept header the MCP spec requires, with no Origin header, gets
an answer. On 3 November 2026, with curl:
200 with application/json, protocolVersion 2025-11-25 and no
Mcp-Session-Id: the server is stateless. tools/list returned 14 tools when this was probed, before the server moved to three (see the MCP server); a notification got
202; a GET got 405; an Accept of application/json alone got 406. It also answers the
2026-07-28 revision’s per-request form: server/discover and tools/list with _meta and
MCP-Protocol-Version, and tools/call without Mcp-Name gets 400 with -32020. The main
deviation from the spec’s Streamable HTTP page: it accepts any https Origin, where the spec
says an invalid one must get 403 (a plain-http origin other than localhost, and null, do get
403). That does not affect a CLI or desktop client, which sends no Origin. The full table of probes is in the MCP server’s own README.
A client that only speaks stdio, or cannot reach a remote server, runs
npx -y @nyuchi/mzizi-mcp (every operation except fundi, offline except docs_*).
mcp-remote, a third-party stdio-to-HTTP bridge, is another route this project has not tested.
What is not verified here
- Every client above, run. No client was installed or run for this page. Config paths, keys and skill folders are what each vendor’s page says, read on the date given in its section.
- JetBrains and Kiro on the details. Junie’s Streamable HTTP support, and
.junie/skills/, rest on a vendor page that does not say, and a third-party source. Kiro’s page does not name an instruction-file format. - Gemini CLI’s extension format, and Antigravity’s and Cursor’s plugin manifests: read as existing, not read in full, and this project publishes none of them.
- The Codex plugin in a live Codex CLI or ChatGPT desktop app. The manifests follow
OpenAI’s packaging guide and validate against the Agent Plugins schemas, but no Codex install
ran, and they are not yet on the registry’s
main(mzizi-registry#520). - Per-server sign-in in the portable
mcp.json. OpenAI’s guide puts OAuth settings under a server’sextensions["com.openai"].auth, but the Agent Plugins MCP schema allows noextensionskey on a server. The plugin’smcp.jsonconnects the anonymous/mcpand declares noauth; connect/mcp/signinyourself for the sign-in operations. - The Claude API, the OpenAI API and Z.ai snippets come from the vendors’ pages, last read on 20 October 2026; none was run, because each needs an account key.