Knolo

@knolo/mcp

@knolo/mcp is a read-only MCP server for pinned Knolo V5 Knowledge Images. It exposes the same four tools as @knolo/openclaw: knolo_search, knolo_get, knolo_verify, and knolo_status. Pins, policy checks, and query receipts come from @knolo/tools. When at least one mount includes a kar block, the server also exposes knolo_kar_evaluate, knolo_kar_verify, and knolo_cortex_verify.

The server speaks stdio only. It does not download images, open a runtime network connection, write or compile knowledge, or expose MCP resources or prompts. Install and pin each image before starting it. A hash, state-root, or policy-root mismatch exits non-zero before the MCP connection starts.

Caller identity is agentId in the config file, default mcp. That id must be listed in the mount's agents array. Tool arguments cannot set it. The MCP client's clientInfo.name is stored on the receipt as sessionId and is not used for authorization. One principal per mount is what policy evaluation sees.

Every search writes an openclaw-knolo-query-v1 receipt before it returns hits. An unwritable auditDir returns AUDIT_UNAVAILABLE and no hits. Restart the process to change pins. There is no hot reload.

@knolo/[email protected] is the in-tree server. This change does not publish it to npm or the MCP Registry. The registry name in server.json is io.github.HiveForensics-AI/knolo. Published @knolo/[email protected] remains the four-tool server.

Config

Paths are absolute. sha256 is 64 lowercase hex characters. stateRoot is sha256- plus 64 lowercase hex characters.

{  "agentId": "mcp",  "auditDir": "/var/lib/knolo/receipts",  "mounts": [    {      "id": "handbook",      "path": "/opt/knolo/handbook.knolo",      "sha256": "<64 hex>",      "stateRoot": "sha256-<64 hex>",      "policyPath": "/opt/knolo/handbook.policy.json",      "agents": ["mcp"],      "principal": "<policy principal>",      "kar": {        "graphPath": "/opt/knolo/handbook.kar.json",        "graphSha256": "<64 hex>"      }    }  ]}

kar is optional. graphSha256 is the raw file SHA-256, 64 lowercase hex characters, without a sha256- prefix. The server hashes that file and binds it to the mount before it accepts a connection. A hash or binding mismatch exits before MCP starts. A call for a mount that has no kar block fails closed when the three tools are enabled by another mount.

knolo_kar_evaluate takes mountId, proposition, and an inline plan object of at most 65536 UTF-8 bytes. SATISFIED means evidence requirements were met. It does not mean a claim is proven. UNSATISFIED_EVIDENCE_REQUIREMENTS is a completed evaluation. The plan is parsed by the KAR plan parser. Domain Pack template names are not accepted.

knolo_kar_verify takes the same mount, proposition, and plan, plus exactly one of result or dossier. knolo_cortex_verify takes mountId, an answer trace, and an optional question. It uses the pinned graph when the trace references KAR. These tools do not call a model and do not run askCortex.

Query receipts stay openclaw-knolo-query-v1. KAR and Cortex tools write knolo-mcp-evidence-receipt-v1 in auditDir before they return a result. An unwritable audit directory returns AUDIT_UNAVAILABLE and no result.

Start it with knolo-mcp --config /absolute/knolo-mcp.json or KNOLO_MCP_CONFIG. Logs go to stderr. Stdout is JSON-RPC only.

Clients

Claude Desktop, Claude Code, and Cursor:

{  "mcpServers": {    "knolo": {      "command": "npx",      "args": ["-y", "@knolo/mcp", "--config", "/absolute/knolo-mcp.json"]    }  }}

Claude Code CLI:

claude mcp add knolo -- npx -y @knolo/mcp --config /absolute/knolo-mcp.json

VS Code (.vscode/mcp.json):

{  "servers": {    "knolo": {      "type": "stdio",      "command": "npx",      "args": [        "-y",        "@knolo/mcp",        "--config",        "${workspaceFolder}/knolo-mcp.json"      ]    }  }}

The package requires Node 20 or newer, @knolo/core ^6.2.0, and @knolo/tools ^0.2.0. Coverage, composition, and authoring stay outside this server. Verification checks integrity. It does not decide that a claim or an answer is true.