Self-Evolving Deterministic Agent Control Protocol

A governance gateway for AI agents — making every action bounded, auditable, reversible, and explainable.

Works transparently with Cursor, Claude Code, Codex, and any MCP-compatible agent. Also supports shell command governance and a language-agnostic HTTP API.

🛡️ Governance in Action

1. Set Up Governance Rule Enable the governance rule in Cursor's settings to protect your workspace.
2. Block Secrets Exfiltration Agent attempts to read .env and write secrets — blocked instantly.	3. Block Credential Scanning Agent tries to search for credentials and secrets files — denied by policy.

https://github.com/user-attachments/assets/ec7a9524-1527-4e51-b837-7e05a24b189d

Table of Contents

• How It Works
• Core Principles
• Quick Start
• Agent Integrations
• Built-in Policies
• Integration Modes
• Policy Self-Evolution
• Architecture
• Policy DSL Reference
• Built-in Tool Adapters
• Custom Tool Adapters
• Development
• Contributing
• License

How It Works

Agents never execute tools directly. Every action flows through the control plane for evaluation, enforcement, and audit:

flowchart LR
    A["Agent"] -->|"action request"| CP["Control Protocol"]
    CP -->|"evaluate against policy"| D{"Decision"}
    D -->|"allow"| E["Agent Executes Action"]
    D -->|"deny"| F["Blocked + Reason Logged"]
    D -->|"gate"| G["Human Approval Required"]
    E -->|"record result"| L["Evidence Ledger"]
    G -->|"approved"| E

The protocol does not execute actions itself. It evaluates them against a policy, enforces session-level budgets, requires human approval for risky operations, and records everything in a tamper-evident audit ledger.

Core Principles

Quick Start

Install

npm i @det-acp/core

Set Up Governance (One Command)

npx det-acp init cursor        # Cursor
npx det-acp init codex         # Codex CLI
npx det-acp init claude-code   # Claude Code

This generates all required files (policy, MCP config, governance rules) with sensible defaults. Edit policy.yaml to customize — everything else is handled automatically.

# Use your own policy instead of the default
npx det-acp init cursor --policy ./my-policy.yaml

After running init, restart your agent to pick up the MCP server.

Define a Policy

Create agent.policy.yaml:

version: "1.0"
name: "my-agent"

capabilities:
  - tool: "file:read"
    scope:
      paths: ["./src/**"]
  - tool: "file:write"
    scope:
      paths: ["./src/**"]
  - tool: "command:run"
    scope:
      binaries: ["npm", "node", "tsc"]

limits:
  max_runtime_ms: 1800000
  max_files_changed: 50

gates:
  - action: "file:delete"
    approval: "human"
    risk_level: "high"

evidence:
  require: ["checksums", "diffs"]
  format: "jsonl"

forbidden:
  - pattern: "**/.env"
  - pattern: "rm -rf"

session:
  max_actions: 100
  max_denials: 10
  rate_limit:
    max_per_minute: 30
  escalation:
    - after_actions: 50
      require: human_checkin
    - after_minutes: 15
      require: human_checkin

Use as a Library

import { AgentGateway } from '@det-acp/core';

const gateway = await AgentGateway.create({
  ledgerDir: './ledgers',
  onStateChange: (sessionId, from, to) => console.log(`${from} -> ${to}`),
});

// Create a session
const session = await gateway.createSession('./agent.policy.yaml', {
  agent: 'my-coding-agent',
});

// Evaluate an action (does NOT execute it)
const verdict = await gateway.evaluate(session.id, {
  tool: 'file:read',
  input: { path: './src/index.ts' },
});

if (verdict.decision === 'allow') {
  // Execute the action yourself
  const content = fs.readFileSync('./src/index.ts', 'utf-8');

  // Record the result
  await gateway.recordResult(session.id, verdict.actionId, {
    success: true,
    output: content,
    durationMs: 5,
  });
}

// Terminate and get report
const report = await gateway.terminateSession(session.id, 'task complete');
console.log(`Allowed: ${report.allowed}, Denied: ${report.denied}`);

Agent Integrations

Ready-to-use guides for popular AI agents. Each integration includes policy, config templates, governance rules, test sandbox, and step-by-step instructions.

For any MCP-compatible agent not listed above, see MCP Proxy (General).

Built-in Policies

Production-ready policies in examples/ — usable out of the box:

Validate any policy with: npx det-acp validate ./policy.yaml

Integration Modes

MCP Proxy (General)

Works with any MCP-compatible client.

Simplified mode — point at a policy file, auto-configures filesystem backend:

npx det-acp proxy --policy ./policy.yaml
npx det-acp proxy --policy ./policy.yaml --dir /path/to/project

cat > mcp-proxy.config.yaml << 'EOF'
policy: ./agent.policy.yaml
ledger_dir: ./.det-acp/ledgers
transport: stdio
backends:
  - name: filesystem
    transport: stdio
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "./src"]
EOF

npx det-acp proxy ./mcp-proxy.config.yaml

Shell Proxy

Execute commands through the policy gateway:

npx det-acp exec ./agent.policy.yaml echo "hello"    # Allowed
npx det-acp exec ./agent.policy.yaml rm -rf /tmp      # Denied (forbidden)

HTTP Session Server

npx det-acp serve --port 3100

# Create a session
curl -X POST http://localhost:3100/sessions \
  -H "Content-Type: application/json" \
  -d '{"policy": "version: \"1.0\"\nname: test\ncapabilities:\n  - tool: file:read\n    scope:\n      paths: [\"./src/**\"]"}'

# Evaluate an action
curl -X POST http://localhost:3100/sessions/<session-id>/evaluate \
  -H "Content-Type: application/json" \
  -d '{"action": {"tool": "file:read", "input": {"path": "./src/index.ts"}}}'

# Record result
curl -X POST http://localhost:3100/sessions/<session-id>/record \
  -H "Content-Type: application/json" \
  -d '{"actionId": "<action-id>", "result": {"success": true, "output": "..."}}'

# Terminate session
curl -X POST http://localhost:3100/sessions/<session-id>/terminate

CLI Reference

npx det-acp init <agent>                  # Set up governance (cursor, codex, claude-code)
npx det-acp init <agent> --policy <file>  # Use custom policy
npx det-acp validate <policy-file>        # Validate a policy
npx det-acp proxy --policy <policy-file>  # Start MCP proxy (simplified)
npx det-acp proxy --policy <policy-file> --evolve  # With policy self-evolution
npx det-acp proxy <config-file>           # Start MCP proxy (full config)
npx det-acp exec <policy-file> <command>  # Execute via shell proxy
npx det-acp report <ledger-file>          # View audit report
npx det-acp serve [--port <port>]         # Start HTTP session server

Policy Self-Evolution

When an action is denied by the policy, the self-evolution feature can suggest a minimal policy change that would allow it, prompt you for a decision, and optionally update the policy (in memory and/or on disk). This keeps governance strict by default while letting you relax policy incrementally when you approve.

What it does

• Analyses denials — Pattern-matches denial reasons (missing capability, path/binary/domain outside scope, forbidden pattern) and produces a single, minimal policy edit.
• Prompts you — Presents the suggestion (e.g. “Add file:read capability for path ./config/**?”) via the agent’s own chat UI.
• Three choices:
  • Add to policy — Apply the change to the session’s policy and persist it to the policy YAML file.
  • Allow once — Apply the change in memory only for the current session (no disk write).
  • Deny — Keep the block; no change.
• Retry — After approval, the agent retries the original tool call against the updated policy.

Budget and session-limit denials (e.g. “Budget exceeded”, “Rate limit exceeded”) are not suggestible; only permission/scope/forbidden denials can trigger evolution.

How it works (MCP-native)

When running as an MCP proxy (the default for Cursor, Claude Code, and Codex), evolution uses a two-step asynchronous protocol built entirely on standard MCP tool calls — no terminal stdin/stdout required:

1. On deny: The proxy returns the denial to the agent with a structured suggestion and a unique suggestion_id.
2. On approve: The agent presents the suggestion to the user in chat, collects their decision, and calls policy_evolution_approve with the suggestion_id and decision.
3. On retry: The proxy applies the policy change. The agent retries the original tool call, which now succeeds.

sequenceDiagram
    participant User
    participant Agent as Agent (Cursor / Claude Code)
    participant Proxy as MCP Proxy
    participant GW as Gateway

    Agent->>Proxy: callTool("read_text_file", path)
    Proxy->>GW: evaluate()
    GW-->>Proxy: deny
    Proxy->>Proxy: suggestPolicyChange()
    Proxy-->>Agent: Denied + suggestion + suggestionId

    Agent->>User: "read_text_file denied. Add capability for path X?"
    User-->>Agent: "Yes, add to policy"

    Agent->>Proxy: callTool("policy_evolution_approve", suggestionId, "add-to-policy")
    Proxy->>Proxy: applyPolicyChange + writePolicyToFile
    Proxy-->>Agent: "Policy updated"

    Agent->>Proxy: callTool("read_text_file", path)
    Proxy->>GW: evaluate()
    GW-->>Proxy: allow
    Proxy-->>Agent: file contents

This works in every MCP client because it uses only the standard MCP tool-call protocol — no readline, no stdin conflicts.

Enabling self-evolution

MCP proxy (simplified mode):

npx det-acp proxy --policy ./policy.yaml --evolve

The init command includes --evolve by default in generated MCP configurations.

Programmatic (library): pass policyEvolution in GatewayConfig for non-MCP setups (e.g. CLI scripts, custom agents):

import { AgentGateway, createCliEvolutionHandler } from '@det-acp/core';

const gateway = await AgentGateway.create({
  ledgerDir: './ledgers',
  policyEvolution: {
    policyPath: './agent.policy.yaml',
    handler: createCliEvolutionHandler(),
    timeoutMs: 30_000,
  },
});

You can plug a custom EvolutionHandler (e.g. GUI dialog, webhook) instead of createCliEvolutionHandler().

Evolution architecture (high level)

flowchart LR
    subgraph Denial["On Deny"]
        A["Action Denied"] --> B["Suggestion Engine"]
        B --> C{"Suggestible?"}
        C -->|No| D["Keep Deny"]
        C -->|Yes| E["PolicySuggestion + ID"]
    end
    E --> F["Agent presents to user"]
    F --> G{"User Decision"}
    G -->|Add to policy| H["policy_evolution_approve → Apply + Write YAML"]
    G -->|Allow once| I["policy_evolution_approve → Apply in-memory"]
    G -->|Deny| D
    H --> J["Retry original tool call"]
    I --> J
    J --> K{"Verdict"}
    K -->|allow| L["Proceed"]
    K -->|deny| D

The Suggestion Engine maps denial reasons to one of: add capability, widen scope (paths/binaries/domains/methods/repos), or remove a forbidden pattern. In MCP proxy mode, the MCP Evolution Handler returns the suggestion to the agent as a structured denial with a suggestion_id, and the agent calls policy_evolution_approve after collecting the user’s decision. In library mode, the Policy Evolution Manager uses a pluggable handler (CLI, GUI, webhook) with a configurable timeout and re-evaluates the action inline.

Architecture

Component Architecture

graph TB
    subgraph External["External Systems"]
        BackendMCP["Backend MCP Servers"]
        Approvers["Human / Webhook Approvers"]
    end

    subgraph Integration["Integration Layer"]
        subgraph MCPProxyGroup["MCP Proxy Server"]
            MCPProxy["MCP Proxy"]
            subgraph McpEvo["MCP Evolution (optional)"]
                McpEvoHandler["MCP Evolution Handler"]
                EvoTool["policy_evolution_approve tool"]
            end
        end
        ShellProxy["Shell Proxy"]
        HTTPServer["HTTP Server"]
        LibrarySDK["Library SDK"]
    end

    subgraph Core["Core Engine"]
        Gateway["Agent Gateway"]
        SessionMgr["Session Manager"]
        PolicyEval["Policy Evaluator"]
        GateMgr["Gate Manager"]
        ActionReg["Action Registry"]
        subgraph Evolution["Policy Self-Evolution (optional)"]
            EvolutionMgr["Policy Evolution Manager"]
            SuggestionEngine["Suggestion Engine"]
        end
    end

    subgraph Infra["Infrastructure"]
        Ledger["Evidence Ledger<br/>(JSONL + SHA-256)"]
        Rollback["Rollback Manager"]
    end

    subgraph Tools["Tool Adapters"]
        subgraph FileTools["File Operations"]
            FR["file:read"]
            FW["file:write"]
            FD["file:delete"]
            FM["file:move"]
            FC["file:copy"]
        end
        subgraph DirTools["Directory Operations"]
            DL["directory:list"]
            DC["directory:create"]
        end
        subgraph GitTools["Git Operations"]
            GD["git:diff"]
            GA["git:apply"]
            GC["git:commit"]
            GS["git:status"]
        end
        subgraph NetTools["Network & System"]
            CR["command:run"]
            HR["http:request"]
            ND["network:dns"]
            ER["env:read"]
            AE["archive:extract"]
        end
    end

    MCPProxy --> Gateway
    ShellProxy --> Gateway
    HTTPServer --> Gateway
    LibrarySDK --> Gateway

    Gateway --> SessionMgr
    Gateway --> ActionReg
    Gateway --> GateMgr

    SessionMgr --> PolicyEval
    SessionMgr --> GateMgr
    SessionMgr --> Ledger
    SessionMgr -->|on deny| EvolutionMgr
    EvolutionMgr --> SuggestionEngine

    MCPProxy -->|on deny| McpEvoHandler
    McpEvoHandler --> SuggestionEngine
    EvoTool -->|apply change| SessionMgr

    ActionReg --> FileTools
    ActionReg --> DirTools
    ActionReg --> GitTools
    ActionReg --> NetTools

    Rollback --> ActionReg
    Rollback --> Ledger

    MCPProxy --> BackendMCP
    GateMgr --> Approvers

Action Evaluation Flow

sequenceDiagram
    participant Agent
    participant Integration
    participant Gateway as Agent Gateway
    participant Session as Session Manager
    participant Policy as Policy Evaluator
    participant Gate as Gate Manager
    participant Ledger as Evidence Ledger

    Agent ->> Integration: Action request (tool, input)
    Integration ->> Gateway: evaluate(sessionId, action)
    Gateway ->> Session: evaluate(sessionId, action)

    Session ->> Policy: evaluateSessionAction(action)

    Note right of Policy: 1. Session state check<br/>2. Rate limits & escalation<br/>3. Forbidden patterns<br/>4. Capability & scope match<br/>5. Budget limits<br/>6. Gate lookup

    Policy -->> Session: allow / deny / gate

    Session ->> Ledger: append(action:evaluate)

    alt Gate required
        Session ->> Gate: requestApproval(action, gate)
        alt Approved
            Gate -->> Session: approved
        else Pending
            Gate -->> Session: pending
            Note over Agent, Gate: Session paused until resolved
        end
    end

    Session -->> Gateway: EvaluateResponse
    Gateway -->> Integration: decision + reasons
    Integration -->> Agent: allow / deny / gate

    alt Allowed
        Note over Agent: Executes action externally
        Agent ->> Integration: recordResult(actionId, result)
        Integration ->> Gateway: recordResult(sessionId, actionId, result)
        Gateway ->> Session: recordResult(result)
        Session ->> Session: Update budget tracking
        Session ->> Ledger: append(action:result)
    end

    opt Session complete
        Agent ->> Integration: terminateSession()
        Integration ->> Gateway: terminate(sessionId)
        Gateway ->> Session: terminate(sessionId)
        Session ->> Ledger: append(session:terminate)
        Session -->> Agent: Session Report
    end

Session Lifecycle

stateDiagram-v2
    [*] --> Created: createSession()
    Created --> Evaluating: evaluate(action)
    Evaluating --> Allowed: policy allows
    Evaluating --> Denied: policy denies
    Evaluating --> Gated: gate required
    Gated --> Allowed: approved
    Gated --> Denied: rejected
    Allowed --> Recording: recordResult()
    Recording --> Evaluating: next action
    Denied --> Evaluating: next action
    Recording --> Terminated: terminateSession()
    Evaluating --> Terminated: terminateSession()
    Terminated --> [*]

Evidence Ledger

Every action produces an immutable audit record in JSONL format with SHA-256 hash chaining:

{"seq":1,"ts":"...","hash":"sha256:abc...","prev":"sha256:000...","type":"session:start","data":{...}}
{"seq":2,"ts":"...","hash":"sha256:def...","prev":"sha256:abc...","type":"action:evaluate","data":{...}}
{"seq":3,"ts":"...","hash":"sha256:ghi...","prev":"sha256:def...","type":"action:result","data":{...}}

If any entry is tampered with, the hash chain breaks and integrity verification fails.

Policy DSL Reference

See examples/ for complete policy files.

Built-in Tool Adapters

File Operations

Directory Operations

Command Execution

Git Operations

Network & HTTP

Environment & System

Custom Tool Adapters

Extend the ToolAdapter base class to add your own tools:

import { ToolAdapter } from '@det-acp/core';
import { z } from 'zod';

class MyCustomTool extends ToolAdapter {
  readonly name = 'custom:mytool';
  readonly description = 'My custom tool';
  readonly inputSchema = z.object({
    target: z.string().min(1),
    options: z.record(z.string()).optional(),
  });

  validate(input, policy) { /* ... */ }
  async dryRun(input, ctx) { /* ... */ }
  async execute(input, ctx) { /* ... */ }
  async rollback(input, ctx) { /* ... */ }
}

// Register it
gateway.getRegistry().register(new MyCustomTool());

Every tool adapter follows the execution lifecycle: validate -> dryRun -> (gate check) -> execute -> verify. Each step is recorded in the evidence ledger.

Development

npm install          # Install dependencies
npm run lint         # Type check (TypeScript strict)
npm test             # Run tests (Vitest)
npm run build        # Build

Contributing

Contributions are welcome! Please follow these guidelines:

1. Fork the repository and create a feature branch
2. Follow the coding standards (TypeScript strict, ESM, Zod validation)
3. Write tests mirroring src/ structure under tests/
4. Run npm test && npm run lint before submitting
5. Use Conventional Commits for commit messages
6. Open a PR with a clear description of changes

License

MIT

Star History