Achromat Machine Operational Memory

Achromat MCP server

An MCP server that lets any MCP host (Claude Desktop, an IDE, an agent framework) use Achromat Machine Operational Memory (MOM) as native tools. MOM keeps verified memory of tool calls that policy blocked and the corrected calls that were then allowed. With this server your agent can ask for the compliant version of a call before it acts, and check a call against live policy, without you writing any integration code.

The server runs on your machine as a child process of the host and talks to it over stdio. It signs every request to MOM with an Ed25519 key that stays on your machine: only signatures leave it, never the key.

Add it to Claude Desktop

Open the configuration file:

Add the server under mcpServers, then restart Claude Desktop:

{  "mcpServers": {    "achromat": {      "command": "npx",      "args": ["-y", "@achromat-ai/mcp-server"]    }  }}

That is the whole setup. Requires Node.js 20 or newer; -y lets npx install the package without asking. Other MCP hosts take the same command and arguments in their own configuration.

First use: your agent sets up the identity

Every request to MOM is signed, so the server needs a key that MOM knows. You do not create or register one yourself. The first time you use the server it has no key, so it offers a single tool, achromat_register_identity, and your agent will ask you whether to set up an identity. If you say yes, the server:

1. generates an Ed25519 key pair on your machine,
2. does a short proof of work (about a second; it is how MOM limits automated sign-ups),
3. registers the public key with MOM, and
4. saves the private key to ~/.achromat/achromat_agent_key.pem, readable only by you.

Then achromat_recall and achromat_calibrate appear in place of the registration tool. If your host does not refresh its tool list by itself, restart it once.

What leaves your machine: the public key, the proof of work, and a signature proving you hold the private key. The private key never does. MOM keeps a record of the registration and limits it to 10 a day per network.

The key file is the identity. Keep it private, out of git and out of screenshots, and back it up: it cannot be recovered if you lose it. The server never overwrites an existing key file, and it only saves the file once MOM has accepted the key, so a failed attempt leaves nothing behind.

Choosing where the key lives

By default the key is at ~/.achromat/achromat_agent_key.pem. To use another path, set ACHROMAT_AGENT_KEY_FILE in the host configuration. This works for a new key (the server creates it there) and for one you already have:

{  "mcpServers": {    "achromat": {      "command": "npx",      "args": ["-y", "@achromat-ai/mcp-server"],      "env": {        "ACHROMAT_AGENT_KEY_FILE": "/Users/you/keys/achromat_agent_key.pem"      }    }  }}

On Windows use the full path with escaped backslashes, for example "C:\\Users\\you\\keys\\achromat_agent_key.pem". A leading ~ is expanded by the server. A desktop app may start the server with a different home directory than your terminal has, so spelling the path out removes any doubt.

An existing key must be an unencrypted PKCS#8 Ed25519 PEM file that MOM already knows.

Attributing an install to a link (optional)

If you share a link or a setup snippet and want to know which one brought an install, add an ACHROMAT_REF tag to the host configuration. It is a short label of 1 to 32 characters from a-z, 0-9, _ and -, for example launch-post:

{  "mcpServers": {    "achromat": {      "command": "npx",      "args": ["-y", "@achromat-ai/mcp-server"],      "env": { "ACHROMAT_REF": "launch-post" }    }  }}

The tag is sent only while the server registers its identity, as ?ref=launch-post on the two registration calls, and nowhere else. It is the label itself and nothing about you or your machine. A tag that does not fit the rule is ignored; it never stops the registration.

Run from source

git clone https://github.com/achromat-ai/achromat-mcp-server.gitcd achromat-mcp-servernpm installnpm run build

Point the host at the build instead of npx:

{  "mcpServers": {    "achromat": {      "command": "node",      "args": ["/absolute/path/to/achromat-mcp-server/dist/index.js"]    }  }}

To try the tools in a browser without a host, use the MCP Inspector:

npm run inspect

Tools

achromat_register_identity

Offered only while there is no key file. Creates and registers an identity as described above. It requires user_confirmed: true, which the agent may set only after you have agreed; the tool description tells it to ask first.

achromat_recall

Call this before running a tool that policy might block (payments, credit decisions, robot or plant commands, shell, SQL, web requests). It asks whether a verified compliant version of the call already exists. On a match you get compliant_arguments, a steering_directive and the rule that applied. Use the whole compliant_arguments object as your template, keep any [VERIFIED_STATE] placeholders exactly as written, and change only your own operational values.

When the lesson is an approved one, the answer ends with a citation line: Tool arguments follow a verified Achromat MOM lesson: https://achromat.ai/?ref=audit#lesson-<id>. The agent is told to add it to text a person reads (its final answer, a pull request description, a commit message), never to the tool's arguments, and to skip it if you ask for no attribution. The link opens that lesson on the Achromat site. A lesson that is only your own, not yet shared, gets no citation.

achromat_calibrate

Call this when recall found nothing, or when the stakes are high. It evaluates the call against live policy and returns the verdict, the steering directive and the receipt id. A BLOCK comes with a trap_id: fix the arguments as the directive says and call the tool again with the same task_description and the trap_id. An ALLOW means the call complies. A block followed by an allow becomes a lesson for other agents.

Never put secrets or credentials in arguments or task_description; the service rejects them.

The tools use the name tool_name; the service's own field is tool, and the server maps between them.

Lesson text (rules, directives, arguments) comes from the community. The server labels it as data in every answer, and your agent should use it only to choose parameter values.

Configuration

Whether the key file exists decides which tools are offered, and it is checked at startup and right after a registration. If you delete, move or replace the key file while the host is running, restart the host. The key itself is read on the first tool call, so a wrong key never makes the server disappear from your host: the tool call returns a message saying what to fix.

When something goes wrong

Every failure comes back to the agent as a readable tool result, with the next step.

Server logs go to stderr, which hosts write to their own MCP log files (Claude Desktop keeps them in its logs folder; on macOS that is under ~/Library/Logs/Claude/). The server never logs keys, signatures or request contents.

If the host cannot start npx, it may not see the PATH your terminal does (common with nvm). Put the absolute path of npx or node in command.

Development

npm run typecheck   # source and testsnpm test            # vitestnpm run build       # compile to dist/ and mark dist/index.js executablenpm start           # run the built server on stdio (waits for a host)

stdout is the protocol channel, so the server writes nothing else to it.

The signing and registration code follows the byte-level contract in the agent protocol (sections 2 and 3), and the tests reproduce its published self-check vectors exactly: the request signature, and the proof-of-work nonce. Ed25519 is deterministic, so a change that alters a single byte of what is signed fails them.

The proof of work searches on several worker threads (up to 8, never all cores) so that it takes about a second and the server stays responsive while it runs.

Licence

MIT