The only client here that uses TOML. One line of config, or one CLI command.
One endpoint exposing 22 data capabilities across 57 pre-described tools — news, governance, DeFi, security and market data. It is a remote server, so there is no package to install, no process to keep alive and no key to obtain. That matters for Codex setup specifically: almost every MCP example online is command + args spawning a local binary, and none of it applies here.
The config below assumes all of this is already true. Most failed setups are one of these rather than a wrong field.
codex CLI, on a current build — older ones gated remote servers behind an experimental flag..codex/config.toml. Untrusted repos have theirs ignored outright.Codex has more than one route in, and they are genuinely different rather than alternatives to the same thing. Use the one that matches the surface you work in.
~/.codex/config.toml (user).codex/config.toml (trusted projects only)Confirm the connection before you go looking for a bug in your prompt. A server that is listed but not connected looks identical to a model that chose not to call it.
alphaday listed with its URL. The TOML table name becomes the server name, so [mcp_servers.alphaday] is what you should see here — if the name is wrong, the table header is why.
This is the first message of the MCP protocol, sent by hand. If it answers, the server is reachable from your machine and every remaining problem is on the Codex side.
Expected response
The version number moves; "name":"alphaday" is the part to look for. The response also carries an mcp-session-id header, which is what a client reuses for every later call.
Ask these in a `codex` session. Each one needs data the model cannot have — a live feed, an incident record, a yield table — so a good answer is proof the tools fired, not just that the server is listed.
--transport flag. Transport is inferred from whether you pass --url or a trailing stdio command. Guides showing --transport streamable-http are wrong..codex/config.toml is ignored entirely in untrusted repos — check trust first if nothing appears.experimental_use_rmcp_client. Current builds do not.[mcp_servers.alphaday] — underscore, plural. [mcpServers.alphaday] is JSON habit and will not be read.Symptom first, because that is what you have. The status codes below were produced against the live server rather than copied from elsewhere — if yours does not appear here, it is worth trusting that difference.
codex mcp add ... --transport streamable-http fails with an unknown-argument error.
There is no such flag. Several published guides show one anyway.
Drop it. --url is what selects the remote transport.
A project-level config has no effect.
Codex ignores .codex/config.toml in repos you have not marked trusted.
Trust the project, or move the server into ~/.codex/config.toml, where trust does not apply.
Remote servers are rejected outright on an older build.
They sat behind experimental_use_rmcp_client before the remote client shipped by default.
Update Codex. Enabling the flag on an old build is the fallback, not the fix.
Opening the endpoint in a browser returns {"detail": "Bad Request"}.
A browser sends a GET. MCP is JSON-RPC over POST, so a 400 here is the server working correctly — there is no page to serve at an MCP endpoint.
Nothing. A 400 on a plain GET is the expected response, and it confirms the server is reachable.
404 {"detail": "Not Found"}.
The URL has a trailing slash, or a path appended. /mcp/ and /mcp/sse are both distinct routes from /mcp, and neither exists.
Use the endpoint exactly as written, with no trailing slash and nothing after it.
The server shows as connected, but the model never calls it.
Not a connection problem. Most clients only call tools in an agent mode, and several ship new tools unticked or capped.
Switch to the client's agent mode, confirm the Alphaday tools are enabled in its tool picker, and ask something that cannot be answered without live data.
22 datasets, reached by 57 tools — most datasets have a list form, a detail form and a trending form, which is why the two numbers differ. These four are the ones a general crypto API does not have:
180+ on-chain exploits and hacks, each with a written incident record.
3,000+ GitHub activity snapshots — commits and contributors per coin.
AI-detected chart patterns with a bullish/bearish signal and confidence score.
32,000+ DeFi yield pools with base and reward APY, TVL and IL risk.
Real-time news from 49 crypto outlets — 440,000+ articles, tagged and deduped.
59 governance forums — 60,000+ posts — behind one endpoint.
133 project blogs — 19,000+ posts — in one feed.
118 crypto podcast feeds — 22,000+ episodes, each with its audio URL and duration.
121 YouTube channels — 34,000+ videos — timestamped.
Prices, metadata, history and trending movers for the top coins.
700+ token categories; about half carry a market-cap aggregate.
140+ centralized exchanges with trust scores and volume.
700+ DEXes with the chains and protocols behind them.
2,000+ protocol fee and revenue records across DeFi.
400+ stablecoins — peg type, mechanism and circulating supply.
Live Snapshot proposals across 51 DAOs.
6,000+ conferences, meetups, hackathons and side events.
A 17,000+ tag taxonomy — projects, chains and categories, with parents and children.
An AI briefing over the whole news corpus, focusable by project.
What crypto is talking about, ranked, right now.
Search project-tagged content across the whole superfeed.
The fear & greed index, per asset.
All 22 capabilities are listed on the MCP overview, with the tool count behind each.
Verified 2026-09-14 against vendor docs: