Product 02 · MCP Server

Confidence Gate

Scores proposed agent actions and returns allow, review, or block decisions — before tools touch files, APIs, customers, or production. The primary control point in the CertaWorks suite.

Score Route to review Block Decision trace MCP Server

Install Local → Open Dashboard Join Hosted Beta

Code change review

Proposed
Scored
Routed
Logged

Agent proposes

Rewrite authentication middleware after failing tests.

The agent has test output, touched files, and rollback notes, but the change affects login behavior.

Confidence 78%

Decision Review required

Decision trace

08:42 · Code change · Proposed · 78% · policy: high_risk_review

Touches authentication flow.
Test evidence is present but incomplete.
Rollback path should be confirmed.

← Product 01: Identity Engine SDK Next: Product 03 Multi-Voice Deliberation Framework →

The Problem

Agents act. Then you find out.

AI agents call tools, write to files, send customer emails, and push production changes without any pre-action control point. By the time a team reviews the log, the action has already happened — a destructive database query has run, an irreversible API call has been made, or a commit has altered code that touches critical paths.

Most agent frameworks offer no systematic way to evaluate a proposed action before it reaches a tool. Operators see the result of what the agent did, not a checkpoint where the action could have been evaluated, held, or blocked before execution.

The window for review closes before it opens. Confidence Gate adds that window back.

How It Works

Score → Route → Trace

Agent proposes an action

Your agent calls the check_gate tool with the proposed action text and any available context — touched files, test evidence, source, project, or risk metadata.

Confidence Gate scores it

The gate runs a 0–100 confidence score, checks the action against configured policy rules, and produces a factor analysis covering linguistic certainty, completeness, contextual alignment, and factual grounding.

Returns a decision with trace

The gate returns allow, review, or block along with the reasoning trace and a durable local log entry. The agent proceeds, pauses for human review, or halts based on the result.

Confidence Gate is a decision-support layer. It adds a pre-action control point between agent intent and tool execution. It does not replace human judgment or production controls.

Policy Defaults

Default confidence thresholds

Thresholds are configurable. These are the defaults shipped with Confidence Gate.

Domain / Scope	Default threshold	Outcome below threshold
Default	0.75	Review or block based on policy
General	0.70	Review or block based on policy
Code execution	0.85	Review required
Financial	0.90	Review required
Legal	0.90	Review required
Medical	0.92	Review required
Risk level: low	0.60	Review or block based on policy
Risk level: medium	0.75	Review or block based on policy
Risk level: high	0.90	Review required
Risk level: critical	0.95	Block unless explicitly overridden

Thresholds are set per policy rule. The values above reflect factory defaults from config.ts. All thresholds are overridable via the set_threshold MCP tool or environment variables.

Core Decisions

Three outcomes. One clear interface.

Allow

Confidence is above the policy threshold and supporting evidence is present. The action proceeds with a trace record logged to local durable storage. No human intervention required.

Example: low-risk file read, documentation update, non-destructive query

Review

Confidence is below the configured threshold or a high-risk policy rule was matched. The action is held in the review queue until a human approves, rejects, or resolves it before the agent proceeds.

Example: authentication change, customer-facing message, production deploy

Block

A hard policy rule was matched — the action is rejected immediately. The reason is logged with a full trace record. The agent does not proceed without a policy change or explicit override.

Example: destructive data deletion without migration evidence, policy-prohibited action type

Local Install

Run it locally in under 5 minutes

Confidence Gate ships as a local MCP server. No hosted endpoint required to get started.

Once the server is running, any MCP-compatible agent runtime can call check_gate to score and route proposed actions.

View on npm →

npm package

npm install -g @certaworks/confidence-gate-mcp-server

Claude Desktop config

{
  "mcpServers": {
    "confidence-gate": {
      "command": "npx",
      "args": ["-y", "@certaworks/confidence-gate-mcp-server"]
    }
  }
}

MCP Interface

Available MCP tools

Confidence Gate exposes the following tools to any MCP-compatible agent runtime.

Scoring

score_confidence

Score agent output text from 0 to 1 with a full factor breakdown.

check_gate

Gate an action and return allow, review, or block with confidence trace.

explain_confidence

Retrieve a saved confidence trace by gate_id.

Policy & Thresholds

set_threshold

Update the default, domain, or risk-level threshold at runtime.

get_config

Inspect the active gate configuration including all current thresholds.

list_policies

List named project policies stored in the local policy store.

upsert_policy

Create or version a named project policy with full audit metadata.

list_policy_audit

List policy version and threshold-change audit events.

Trace & Audit

list_traces

List stored gate traces from the local trace store.

query_traces

Filter traces by decision, approval status, project, source, policy, domain, risk, or time range.

resolve_approval

Approve, reject, or resolve a review-queue item with approver identity and note.

record_enforcement_receipt

Record what a client actually did after receiving a gate decision.

Dashboard

get_dashboard_summary

Return dashboard counts, recent checks, review queue, config, policies, and policy audit events.

Dashboard

Local dashboard — prototype status

Confidence Gate includes a local HTML dashboard for reviewing scored decisions, exploring policy settings, and inspecting trace history. The dashboard runs locally alongside the MCP server and does not require any cloud connection.

Roadmap — not currently built

Hosted dashboard, team seats, persistent cloud trace storage, and managed alerting are roadmap items. They are not currently built or available. What is available today is the local MCP server and the local prototype dashboard.

View local dashboard prototype →

Scope

What Confidence Gate does and does not do

Does

Score proposed agent actions with configurable policy rules
Return allow / review / block with confidence factors and trace context
Log decisions to local durable storage with hash-chain integrity
Provide a local MCP server interface for any MCP-compatible agent runtime

Does Not

Guarantee safety or prevent all harmful agent actions
Replace human review, legal review, or production controls
Provide hosted storage, team accounts, or managed alerting (roadmap)
Cover actions the agent takes without calling the gate
Recover hidden prompts or provide full agent observability

Hosted Beta

Get early access to hosted Confidence Gate

The local MCP server is available now for private testing. Hosted beta with cloud trace and team features is in development. Leave your details and we will reach out when hosted access opens.