Codex CLI and MCP Server Discovery: Navigating Registries, Smithery, Glama, and the Official MCP Registry

The MCP ecosystem has exploded. As of late May 2026, over 27,000 MCP servers are indexed across public registries ¹, spanning databases, cloud providers, SaaS integrations, developer tools, and bespoke domain-specific services. For Codex CLI users, the challenge is no longer whether a tool exists — it is finding the right one, verifying its trustworthiness, and wiring it into config.toml without friction.

This article maps the MCP server discovery landscape, compares the major registries, and provides concrete workflows for integrating discovered servers into Codex CLI projects.

The Discovery Problem

Codex CLI’s MCP integration is straightforward once you know what to install. You add a [mcp_servers.<name>] table to ~/.codex/config.toml (or a project-scoped .codex/config.toml for trusted projects) ², and the CLI handles the rest. The bottleneck is upstream: which server, which version, and is it maintained?

With four major registries, several meta-indexes, and an emerging protocol-level discovery standard, the ecosystem resembles early npm — fast-growing, duplicative, and lacking a single canonical source.

The Major Registries

The Official MCP Registry

The official registry at registry.modelcontextprotocol.io is maintained by Anthropic with community contributions ³. It serves as the canonical, protocol-level source and exposes a REST API at /v0/servers:

# Fetch ten servers from the official registry
curl "https://registry.modelcontextprotocol.io/v0/servers?limit=10"

Each entry uses a standardised server.json format containing the server’s unique name (e.g. io.github.user/server-name), location metadata (npm package, remote URL), execution instructions (command-line args, env vars), and capability declarations ⁴. The registry also supports an updated_since parameter for incremental synchronisation — useful for tooling that needs to track new additions.

Best for: Programmatic access, client implementations, canonical metadata.

Glama

Glama (glama.ai/mcp) is the largest directory by volume, indexing over 21,000 servers with visual previews, daily updates, and broad coverage of niche and newly published servers ¹. Its interface is closer to a search engine than a curated marketplace — you get everything, including experimental and low-quality entries.

Best for: Maximum coverage when exploring what exists.

mcp.so

The mcp.so community directory lists approximately 19,700 servers, with strong coverage of third-party and unofficial tools ¹. It leans towards community-submitted entries and newer or experimental servers.

Best for: Finding community-built servers, especially newer or experimental ones.

Smithery

Smithery (smithery.ai) positions itself as the Docker Hub of MCP — a curated registry with approximately 7,000 listings, clean search, one-click install commands, and both local and hosted deployment options ⁵. It offers a CLI (smithery) and an MCP server of its own (smithery-cli-mcp) that lets AI agents browse and install servers programmatically ⁶.

Best for: Developer experience, install convenience, hosted remote servers.

Comparing the Registries

graph LR
    subgraph Registries
        OR["Official Registry<br/>Canonical metadata<br/>API-first"]
        GL["Glama<br/>21,000+ servers<br/>Broadest coverage"]
        MS["mcp.so<br/>19,700+ servers<br/>Community-driven"]
        SM["Smithery<br/>7,000+ servers<br/>Install UX"]
    end

    subgraph Meta-Indexes
        MF["MCPfinder<br/>27,000+ aggregated<br/>Agent-readable"]
    end

    OR --> MF
    GL --> MF
    SM --> MF

    subgraph Codex CLI
        CT["config.toml"]
    end

    MF --> CT
    SM --> CT
    OR --> CT

Registry	Servers	API	CLI tool	Hosted servers	Trust signals
Official Registry	Growing	REST `/v0`	No	No	Canonical
Glama	21,000+	Yes	No	No	Community ratings
mcp.so	19,700+	Limited	No	No	Stars, recency
Smithery	7,000+	Yes	`smithery` CLI	Yes	Verified badge

MCPfinder: Agent-Native Discovery

MCPfinder (mcpfinder.dev) is not another directory site — it is an MCP server that sits on top of every public registry, providing agent-readable search with trust scoring and install-ready configuration ⁷. Install it as an MCP server, and Codex CLI gains the ability to discover tools programmatically during a session:

[mcp_servers.mcpfinder]
command = "npx"
args = ["-y", "@anthropic/mcpfinder"]

With MCPfinder configured, you can ask Codex to find and install servers conversationally:

“Find an MCP server for Elasticsearch that supports index management and query debugging.”

MCPfinder searches across the Official Registry, Glama, and Smithery, inspects environment variables and trust signals, and returns install-ready JSON — which the agent can then translate into config.toml entries.

Wiring Discovered Servers into Codex CLI

STDIO Servers (Local Processes)

Most discovered servers run as local processes started by npx, uvx, or a binary. The standard config.toml pattern:

[mcp_servers.context7]
command = "npx"
args = ["-y", "@upstash/context7-mcp"]

[mcp_servers.context7.env]
LOCAL_TOKEN = "your-token-here"

Key configuration options ²:

startup_timeout_sec — Server initialisation deadline (default: 10 seconds)
tool_timeout_sec — Per-tool execution deadline (default: 60 seconds)
enabled — Toggle without removing the config block
enabled_tools / disabled_tools — Allowlist or blocklist specific tools
default_tools_approval_mode — auto, prompt, or approve

Streamable HTTP Servers (Remote)

Remote servers — including Smithery-hosted instances — use the url parameter:

[mcp_servers.figma]
url = "https://mcp.figma.com/mcp"
bearer_token_env_var = "FIGMA_OAUTH_TOKEN"

For OAuth-enabled servers, run codex mcp login <server-name> to complete the flow. Codex stores tokens securely and refreshes them automatically ².

Using Smithery CLI for Installation

Smithery’s CLI can install servers to supported clients directly ⁶:

# Install globally
npm install -g smithery@latest

# Add an MCP server
smithery mcp add @anthropic/brave-search --client codex \
  --config '{"apiKey": "your-key"}'

# List installed servers
smithery mcp list --client codex

⚠️ As of May 2026, Smithery CLI has native support for Claude Desktop and Cursor clients, but Codex CLI support requires manual config.toml editing after discovery. The --client codex flag is community-contributed and may not cover all configuration options.

Using `codex mcp` Commands

Codex CLI provides its own MCP management subcommand ⁸:

# Add an STDIO server interactively
codex mcp add github-mcp \
  --env GITHUB_TOKEN=ghp_xxx \
  -- npx -y @github/mcp-server

# List configured servers
codex mcp list

# Remove a server
codex mcp remove github-mcp

The .well-known/mcp.json Discovery Standard

The MCP v2.1 specification introduces Server Cards via /.well-known/mcp.json — a standardised endpoint that exposes server metadata without requiring a full connection ⁹. This enables:

IDE extensions auto-configuring when pointed at a domain
Registries crawling domains to discover servers automatically
Clients displaying capabilities before initialisation

sequenceDiagram
    participant C as Codex CLI
    participant D as example.com
    participant R as MCP Registry

    C->>D: GET /.well-known/mcp.json
    D-->>C: Server metadata (transports, auth, capabilities)
    C->>C: Generate config.toml entry
    C->>D: Connect via advertised transport

    R->>D: Crawl /.well-known/mcp.json
    R-->>R: Index server capabilities

An IETF draft (draft-serra-mcp-discovery-uri) also defines DNS TXT record discovery for environments where HTTP well-known URIs are impractical ¹⁰. This is still early-stage but signals the ecosystem’s trajectory towards zero-configuration discovery.

A Practical Discovery Workflow

For senior developers evaluating MCP servers for a Codex CLI project, here is a pragmatic workflow:

1. Start with the Official Registry

Query the API for servers matching your domain:

curl -s "https://registry.modelcontextprotocol.io/v0/servers?q=elasticsearch" \
  | jq '.servers[] | {name, description, updated_at}'

2. Cross-Reference on Smithery

Check Smithery for install convenience, hosted options, and verified badges:

smithery mcp search elasticsearch

3. Verify Trust Signals

Before adding any server to config.toml, check:

GitHub stars and recent commits — Is it maintained?
Environment variables required — What credentials does it need?
Tool count and scope — Does it expose more tools than you need? (MCP schema bloat is real ¹¹)
Licence — Is it compatible with your project?

4. Scope to Project

For team projects, use project-scoped configuration in .codex/config.toml rather than the global file. This keeps MCP dependencies explicit and version-controlled:

# .codex/config.toml (project-scoped, trusted projects only)
[mcp_servers.project-db]
command = "npx"
args = ["-y", "@askdba/mysql-mcp-server"]

[mcp_servers.project-db.env]
DSN = "${MYSQL_DSN}"

5. Lock Tool Exposure

Use enabled_tools to restrict the agent to only the tools it needs, reducing context window consumption and attack surface:

[mcp_servers.github]
command = "npx"
args = ["-y", "@github/mcp-server"]
enabled_tools = ["get_issue", "list_pull_requests", "create_comment"]
default_tools_approval_mode = "prompt"

What is Coming Next

The discovery landscape is consolidating rapidly. Three trends to watch:

Protocol-level discovery — .well-known/mcp.json and DNS-based discovery will reduce reliance on centralised registries for well-known services ⁹.
Agent-driven installation — Tools like MCPfinder and Smithery’s CLI MCP server are making it possible for agents to discover and install their own tools mid-session, closing the loop between “I need a capability” and “I have a capability” ⁷.
Enterprise governance — As organisations adopt MCP at scale, expect private registries with approval workflows, SBOM-style manifests for MCP dependencies, and policy enforcement for which servers can be configured in project-scoped configs.

For Codex CLI users, the practical advice is straightforward: use the official registry as your canonical source, Smithery for install convenience, MCPfinder for agent-native discovery, and always scope and restrict tools to what your project actually needs.

Codex CLI and MCP Server Discovery: Navigating Registries, Smithery, Glama, and the Official MCP Registry

The Discovery Problem

The Major Registries

The Official MCP Registry

Glama

mcp.so

Smithery

Comparing the Registries

MCPfinder: Agent-Native Discovery

Wiring Discovered Servers into Codex CLI

STDIO Servers (Local Processes)

Streamable HTTP Servers (Remote)

Using Smithery CLI for Installation

Using codex mcp Commands

The .well-known/mcp.json Discovery Standard

A Practical Discovery Workflow

1. Start with the Official Registry

2. Cross-Reference on Smithery

3. Verify Trust Signals

4. Scope to Project

5. Lock Tool Exposure

What is Coming Next

Citations

Using `codex mcp` Commands