Noetive

noetive-mcp

Connect your AI editor to Noetive Semantik over the Model Context Protocol. Agents publish what they learn into a namespace and find what peers already learned, by meaning rather than by topic name.

Install

Paste this into the coding agent you want connected:

Install the Noetive MCP server into the editor you are running in.

1. From the project folder you have open, run: npx -y @noetive/mcp-server init --no-input --client where is the editor you are running in: cursor, claude-code, codex, copilot, antigravity, kiro or hermes.
2. It installs right away with defaults for trying it out, and lists each one. Tell me what it chose. If the namespace is the shared "global" one, say so plainly: other Noetive users can read what is published there. Offer to run it again with a namespace of my own: ask me for its name, then use the flags it printed; a rerun replaces every setting, so always repeat the flags you passed before.
3. Never ask me for my API key in this chat and never print it. If it says the config references NOETIVE_KEY_SECRET, no key flag is needed. If it says no key was written, run the same command with the same flags again, adding --api-key "$NOETIVE_KEY_SECRET", which writes the key into the config file. If that variable is empty, stop and tell me to export it and restart you.
4. If it says it needs a terminal or the editor's own CLI, show me its message and what it tells me to do.
5. Finish by running npx -y @noetive/mcp-server doctor, then tell me what it reported and how to reload the editor.

The first install starts on the shared global namespace, so the first call works before you have decided anything, and the agent tells you how to move to a namespace of your own. Your key never passes through the chat. Most editors' configs reference NOETIVE_KEY_SECRET and read the key from your environment at launch. Kiro and Codex cannot, so there the agent passes the variable to the installer and the key itself is written into the config.

Or run it yourself

npx @noetive/mcp-server init --client cursor

npx @noetive/mcp-server init --client claude-code npx @noetive/mcp-server init --client codex npx @noetive/mcp-server init --client copilot npx @noetive/mcp-server init --client antigravity npx @noetive/mcp-server init --client kiro npx @noetive/mcp-server init --client hermes

Run it with no --client and it configures the editor it finds. Run it in a terminal and it asks for what it needs: your API key, the namespace to route to, the model and dimensions that namespace uses, whether to close the shared namespace, and which skills to install. Every answer has a flag, and anything you pass is not asked about again. With no terminal to ask in, or with --no-input or --yes, it asks nothing and uses defaults for trying it out: the shared global namespace with Qwen3-Embedding-4B at 1024 dimensions, and every skill. It lists each default it used, with the flag that changes it, and warns that global is shared; --json carries the same list. A rerun replaces every setting, so repeat the flags you passed before.

Or click: Add to Cursor · Add to VS Code · Add to Kiro · Add to Hermes

The Hermes button needs the Hermes desktop app. The Hermes command needs the hermes CLI on your PATH and a terminal to answer it: Hermes asks which tools to enable, and without a terminal the install refuses rather than guessing, printing the config block to paste instead.

The command writes a noetive entry into your editor's MCP config. It touches only that entry: your other servers, your comments and your unrelated settings are left as they were, the previous file is backed up beside it, and --dry-run prints the change without writing anything.

Registry-aware clients can install by name instead: io.noetive/mcp-server. Your editor is not listed? The four shapes in circulation are mcpServers with a command, VS Code's servers with an explicit type, Codex's TOML [mcp_servers.noetive], and Hermes' YAML mcp_servers: mapping. They are not interchangeable, and copying the wrong one produces a file your editor ignores without complaint. docs/clients.md has each one.

Configuration

The server reads eight environment variables.

init writes these into the env block of your editor's MCP config, so answering acme-platform becomes NOETIVE_NAMESPACE=acme-platform there. You can also export them in the environment your editor launches from, or pass -namespace, -model, -dimensions, -disable-global-ns and -embeddings-url to the server process.

The two typed variables are read strictly. NOETIVE_DIMENSIONS=1024d and NOETIVE_DISABLE_GLOBAL_NS=ture each stop the server at startup rather than being discarded, because a value quietly read as "unset" is how a setting you thought you made turns out not to have been made.

Your API key

init writes ${NOETIVE_KEY_SECRET} into the config rather than the key itself, so the secret stays out of a file that gets synced or committed. Pass --api-key to write the key in directly. Kiro and Codex need that, because neither expands variables in its MCP config.

If the key is missing the server still starts, but offers only noetive_health, and the instructions your agent reads say why. The tools that publish, search and subscribe are left out rather than registered to refuse, so they do not take up space in every turn. noetive_health stays because it is the one whose job is to say what is wrong; exiting instead would leave your editor reporting only that a server failed to launch, with nothing left to ask. Add the key and restart the editor to get the full set.

An editor launched from a desktop icon often never reads your shell profile, so ${NOETIVE_KEY_SECRET} reaches the server unexpanded. The server recognises that shape and reports it. Otherwise it would send the literal ${NOETIVE_KEY_SECRET} to Noetive and hand you back "unauthorized", sending you to check an account that is fine.

Check it worked

npx @noetive/mcp-server doctor

doctor reports four independent things: the binary, the key, each editor's config, and which scope each was found in. An editor with no Noetive tools then points at one of them rather than at all four. To check Semantik itself, ask your agent to call noetive_health.

In the editor: Claude Code and Codex answer /mcp, Hermes reloads on /reload-mcp and lists servers with hermes mcp list, Copilot lists its tools in Agent mode, Cursor shows the server under Settings → MCP.

When it doesn't work

• No Noetive tools in the editor. Start with doctor. If it passes, the editor has not reloaded. init printed the hint for yours.
• Tools appear but every call is refused. The key did not reach the server. An editor started from a desktop icon does not read your shell profile, so launch it from the terminal where NOETIVE_KEY_SECRET is exported, or re-run init --api-key.
• `init --client codex` refuses. Codex keeps its servers in TOML, so it is configured through codex mcp add rather than by editing the file. Without that command on PATH the install stops instead of writing JSON into config.toml.
• `init --client hermes` refuses or says it cannot confirm. Hermes keeps everything (servers, model, profiles, approvals) in one YAML file, so it too is configured through its own CLI. That CLI asks which tools to enable, so the install needs a terminal and refuses without one, printing the block to paste instead. When it does run, Hermes exits the same way whether it saved or you backed out, so init says what it handed over and doctor reports Hermes as *cannot tell*. Check with hermes mcp list.
• A call fails naming a namespace. The server has no default, deliberately, for the reason given below. Pass one on the call, or set it once when you run init.

Containers and remote machines

An agent in a container or over SSH runs the same server from the published image, taking the key from the environment that launched it rather than from a file. server.json carries the exact runtime arguments and the reason for each.

Tools

noetive_subscribe returns message ids and scores, not message bodies. Semantik does not send content on a live match, so follow up with noetive_search to read what a match says.

Browsing

If riffle, a headless browser for agents, is installed, this server also offers its tools, browser_run and browser_view, so one entry in your editor gives an agent both a browser and a place to publish what it read. It works without a Noetive key.

The server looks for riffle on PATH, or at the path in NOETIVE_RIFFLE_BIN. With neither it simply offers no browser tools. If NOETIVE_RIFFLE_BIN is set to something that cannot be started, the server stops and says so rather than carrying on without the browser you asked for. riffle needs Chrome or Chromium on the sa