ArmorClaudeGetting Started

MCP Tools

ArmorClaude's MCP tools for intent plan registration, policy reading, and trust operations

MCP Tools

ArmorClaude exposes 5 MCP tools through the armorclaude-policy MCP server. These are available to Claude during a session - Claude uses them to register plans, read policy, and perform trust operations.

MCP tools are used by Claude (the LLM), not by the user directly. Users manage policy through /armorclaude:armor commands. There is intentionally no policy_update MCP tool - policy mutation is human-only.

policy_read

Read the current policy state.

Input:

ParameterTypeRequiredDescription
idstringNoSpecific statement ID to read

Without id: Returns the full policy state (version, policy JSON, history).

With id: Returns a single policy statement by its ID.

Example output:

{
  "version": 3,
  "policy": {
    "schemaVersion": "armor.policy.v1",
    "kind": "PolicyProfile",
    "defaults": { "decision": "deny" },
    "statements": [
      { "id": "stmt_1", "effect": "permit", "action": { "type": "tool", "eq": "Read" } }
    ]
  }
}

register_intent_plan

Declare the tools Claude intends to use for the current task. A restrictive active policy requires a plan for calls that change state. Reads such as Read, Grep, and Glob need no plan but still pass policy checks. Policies with default allow and only permit statements, such as all-allow, capture intent without blocking missing plans or drift. Empty default-deny policies do not qualify.

Input:

ParameterTypeRequiredDescription
goalstringYesOne-line summary of what the plan accomplishes
stepsarray or JSON stringYesOrdered list of tool calls
planobject or JSON stringNoAlternative: pass the whole plan as one object

Each step in the steps array:

FieldTypeRequiredDescription
actionstringYesTool name (e.g., Read, Edit, Bash, mcp__server__tool)
descriptionstringNoWhy this step is needed
argumentsobjectNoExpected inputs, advisory while strictParamCheck is false. Use policy conditions for parameter restrictions

Example input:

{
  "goal": "Read and summarize README.md",
  "steps": [
    {
      "action": "Read",
      "description": "Read the README file",
      "arguments": { "file_path": "/project/README.md" }
    }
  ]
}

What happens when called:

  1. The plan is validated against the intent plan schema
  2. If signed in, the plan is sent to the ArmorIQ backend for a signed intent token
  3. The result is stored as pending-plan.<sessionId>.json, or pending-plan.json if no session ID was resolved
  4. The next applicable enforcement-path PreToolUse hook consumes the pending plan. Read-only fast-path calls may return before this step

A token-request failure can still leave a locally recorded plan. An intent registration response alone does not prove a signed token was issued; inspect the token status in the response. Session routing uses a hook-maintained active-session pointer when the MCP process has no session ID, so concurrent windows do not have a guaranteed race-free routing boundary.

Example output:

Intent registered: 2 steps. Token valid 3600s.

trust_revoke

Revoke the currently active intent token. Use when the plan has gone wrong or needs to be stopped immediately.

Input:

ParameterTypeRequiredDescription
reasonstringYesWhy this revocation is happening
cascadebooleanNoIf true, also revoke child tokens (subtree delegations)

What happens when called:

  1. The active session's intent token is located
  2. A revocation delta is signed and broadcast to the ArmorIQ backend
  3. Enforcement points apply the revocation as it propagates; there is no fixed fleet-wide deadline guaranteed by the plugin
  4. The revocation is recorded in the runtime state's trust operation log

Example output:

Token revoked. trustId=tr_abc123 (2 descendants cascaded)

trust_reanchor

Re-anchor to an updated plan without revoking the current token. Use when the plan needs to change mid-session (e.g., additional tools needed).

Input:

ParameterTypeRequiredDescription
updatedPlanobjectYesThe revised plan (same schema as register_intent_plan)
reasonstringNoWhy the plan changed - surfaces in the audit timeline

What happens when called:

  1. The updated plan is validated
  2. A delta δ(h_P → h'_P) is signed, linking the old plan hash to the new one
  3. The working token is re-minted so subsequent tool calls succeed against the new plan
  4. The tamper-evident lineage is preserved in the IAP audit log

Example output:

ReAnchor recorded. trustId=tr_def456 delta(a1b2c3d4e5f6 → 7890abcd1234)

trust_delegate

Issue a subtree-bounded child token with a Merkle inclusion proof. Used for multi-agent scenarios where a sub-agent needs authority over part of the plan.

Input:

ParameterTypeRequiredDescription
delegatePublicKeystringYesPublic key of the sub-agent
subtreePathstringYesPlan subtree path (e.g., /steps/[1])
validitySecondsnumberNoOverride the default validity period
targetAgentstringNoHuman-readable label for the sub-agent

What happens when called:

  1. A child token is minted, scoped to the specified subtree path
  2. A Merkle inclusion proof is generated, proving the subtree is part of the parent plan
  3. The sub-agent receives authority confined to the named subtree
  4. The delegation is recorded in the trust operation log

Example output:

Delegation issued. trustId=tr_ghi789 subtreePath=/steps/[1] (delegationId=del_xyz)

Whitelisted Tools

These tools are never blocked by ArmorClaude's intent enforcement (they are needed for the system to function):

  • register_intent_plan - Claude needs this to declare plans
  • policy_read - Policy reading (read-only, no mutation)
  • trust_revoke / trust_reanchor / trust_delegate - Trust operations
  • ExitPlanMode - Plan mode approval
  • ToolSearch / TodoWrite / ListMcpResourcesTool / ReadMcpResourceTool - Claude Code internals with no side effects

Verifying MCP Server Connection

Check that the ArmorClaude MCP server is connected:

claude mcp list | grep armorclaude

Expected output:

Look for the installed variant's armorclaude-policy server and its connection status.

On this page