MCP Tools

These tools are available to any agent connected to the kcp-harness MCP server.

kcp_plan

Get a deterministic load plan without loading content. Use this to inspect what units are available before loading.

Parameters:

Parameter	Type	Required	Description
task	string	Yes	Description of what you're trying to do
manifest	string	Yes	Path to knowledge manifest
env	string	No	Environment (affects temporal gates)
audience	string	No	Requester audience tag

Returns: Plan object with eligible units, their scores, and gate results.

kcp_load

Plan + load eligible unit content. This is the primary tool for accessing governed knowledge.

Parameters:

Same as kcp_plan.

Returns: Plan object + loaded unit content for all eligible units.

kcp_trace

Get a full 14-gate decision trace. Shows exactly which gates each unit passed or failed, and why.

Parameters:

Same as kcp_plan.

Returns: Decision trace with per-unit, per-gate results.

kcp_validate

Validate a knowledge manifest (knowledge.yaml).

Parameters:

Parameter	Type	Required	Description
path	string	Yes	Path to the manifest or directory containing it

Returns: Validation report with errors and warnings.

harness_status

Get the current governance state.

Returns:

{
  "session": "abc123",
  "governed_domains": 1,
  "approved_plans": 2,
  "known_units": 5,
  "budget_spent": { "USDC": 0.25 },
  "policy": { "fail_closed": true, "max_units": 5 }
}

harness_session

Get session information including known units and approved plans.

Returns: Session state with plan cache and known set.

harness_budget

Get the current budget ledger — itemized spend tracking.

Returns:

{
  "ceiling": { "amount": 1.00, "currency": "USDC" },
  "spent": { "USDC": 0.25 },
  "remaining": { "USDC": 0.75 },
  "entries": [
    {
      "seq": 1,
      "timestamp": "2025-07-07T10:30:00.000Z",
      "source": "plan",
      "description": "kcp_load: 3 units",
      "cost": { "amount": 0.25, "currency": "USDC" },
      "runningTotal": { "USDC": 0.25 }
    }
  ]
}

harness_temporal_check

Check all watched plans for temporal drift. Re-evaluates plans against the current time and reports units that have changed status.

Returns:

{
  "drifted": false,
  "plans_checked": 2,
  "results": []
}

Or if drift is detected:

{
  "drifted": true,
  "plans_checked": 2,
  "results": [
    {
      "manifest": "./knowledge.yaml",
      "summary": "1 unit expired: 'api-v1-docs' (temporal gate failed)",
      "diff": { ... }
    }
  ]
}

harness_approvals

List human-approval tickets: governed calls held for a named reviewer. A call matching a governance.approvals rule is denied with its ticket id — re-try it after a human resolves the ticket (see kcp-harness approvals).

Parameters:

Parameter	Type	Required	Description
state	string	No	Filter: pending_review, approved, dismissed, expired

Returns:

{
  "approvals": [
    {
      "id": "2d62d5a2-…",
      "state": "pending_review",
      "toolName": "Write",
      "target": "records/customer-7.md",
      "requiredRole": "account-owner",
      "requestedAt": "2026-07-20T17:01:04.696Z",
      "expiresAt": "2026-07-23T17:01:04.696Z",
      "policyRef": "POL-7.2"
    }
  ]
}

Resolved tickets additionally carry reviewer, reviewedAt, and (for dismissals) note.

harness_assess

Confidence-gate a synthesized answer before acting on it. Runs kcp-agent's post-synthesis assess(): the answer's self-reported confidence is adjudicated deterministically against the org threshold (the planner gates loading, grounding gates asserting, this gates acting).

Parameters:

Parameter	Type	Required	Description
task	string	Yes	The task the answer concludes
answer	string	Yes	The synthesized answer to gate
threshold	number	No	Tightens the configured threshold — the strictest wins, a caller can never loosen org policy
severity	string	No	Severity label override (e.g. critical)

Returns:

{
  "allowed": false,
  "verdict": {
    "gate": "confidence",
    "passed": false,
    "threshold": 0.7,
    "score": 0.4,
    "signals": [{ "source": "self", "score": 0.4, "reasoning": "self-reported: \"Confidence: 0.4\"" }],
    "detail": "confidence 0.4 < threshold 0.7 on critical task — self: …",
    "severity": "critical",
    "asOf": "2026-07-20"
  },
  "ticket": { "id": "…", "state": "pending_review", "requiredRole": "account-owner" }
}

Behavior on a failed verdict when governance.confidence.route_to_role is set:

• No ticket yet → one is opened with the verdict embedded as evidence (ticket in the result)
• Pending → the same ticket is returned; no duplicate is opened
• Approved by a named human → allowed: true with the override (reviewer, policyRef) attached — the verdict still records that the gate itself failed
• Dismissed → terminal; dismissed (reviewer, note) attached, no new ticket

Fail-closed: an answer with no confidence signal at all fails the gate with a specific reason.