DeleGuard

Authorize actions. Request approval. Record the outcome.
Deterministic authorization to integrate before your agent executes a sensitive action.

Quick start · Demo · How it works · Screenshots · Integrations · Documentation

_{Actual console with demo data. Click the screenshots for a closer look.}

What it does

Your agent proposes an action; DeleGuard evaluates policies and returns ALLOW, DENY, or REQUIRE_APPROVAL. The AI model stays outside the authorization decision.

Capability	What you can do
Business policies	Define YAML rules for the agent, resource, and context; test them and publish versions.
Human approvals	Review a sensitive operation and grant time-limited approval or deny it.
Permits and evidence	Consume a permit once and associate the reported outcome with its execution.
Console and audit trail	Inspect decisions, approvals, and outcomes within each organization.

Includes an HTTP API, remote Guard MCP, .NET/Python/TypeScript clients, and an optional AI Gateway. You do not need an AI provider key to try the platform.

Quick start

The verified setup runs PostgreSQL and Redis in Docker, with the API and worker running locally using .NET.

Requirement	Version or condition
.NET SDK	10
Docker + Compose	Linux containers and Compose v2
Git	To clone the repository
Terminal	Examples use PowerShell 7; Bash environment variables are shown below

1. Clone and start the data services

git clone https://github.com/juliansprt/DeleGuard.git
cd DeleGuard
docker compose -f deploy/docker-compose.yml up -d --wait postgres redis

PostgreSQL uses localhost:5432 and Redis uses localhost:6379. PostgreSQL data persists in a volume; Redis is optional for decision correctness.

2. Start the API and console

From the repository root, in your first terminal:

$env:ASPNETCORE_ENVIRONMENT = "Development"
$env:DOTNET_ENVIRONMENT = "Development"
dotnet run --project src/Deleguard.Api --no-launch-profile --urls http://127.0.0.1:5080

In Development, the API applies migrations and seeds the demo organization, agent, and policies. Wait for it to announce that it is listening on port 5080.

3. Start the worker

From the same repository root, in another terminal:

$env:ASPNETCORE_ENVIRONMENT = "Development"
$env:DOTNET_ENVIRONMENT = "Development"
dotnet run --project src/Deleguard.Workers --no-launch-profile

Keep both processes running. The worker processes the outbox, webhook deliveries, and approval expiration.

Using Bash on Linux or macOS?

Replace the environment assignments in each terminal with these two lines, then run the same dotnet run command from the corresponding step:

export ASPNETCORE_ENVIRONMENT=Development
export DOTNET_ENVIRONMENT=Development

--no-launch-profile does not load variables from launchSettings.json; set Development explicitly.

4. Open the console

In a third terminal, check the connection:

Invoke-RestMethod http://127.0.0.1:5080/health

The response should be Healthy. Open http://127.0.0.1:5080/login:

Email	Password
admin@demo.local	DeleguardDemo!23

Note

These public credentials are only for the local demo. The quick start uses local login and does not require Keycloak. The full Compose stack needs adjustments before startup: see the development guide.

Demo in action

The first integration protects invoice.cancel. With the API running and Python 3 installed, run this from the repository root:

$env:DELEGUARD_URL = "http://127.0.0.1:5080"
python examples/python/authorize_demo.py

No external packages are required. In Bash, use export DELEGUARD_URL=http://127.0.0.1:5080 and python3 examples/python/authorize_demo.py.

Scenario	Decision
Draft invoice for COP 5,000,000, not reported to the tax authority	🟢 ALLOW
Issued invoice for COP 35,000,000, not reported to the tax authority	🟡 REQUIRE_APPROVAL
Invoice reported to the tax authority	🔴 DENY

The example evaluates the invoice policy and creates the pending approval; it does not cancel invoices. Open Approvals to inspect and resolve it with the demo administrator. The policy requests billing-manager; the organization administrator can also resolve it.

How it works

flowchart LR
    Agent[Agent] --> PDP["POST /v1/authorize"]
    PDP -->|ALLOW| Permit[Consume permit]
    PDP -->|REQUIRE_APPROVAL| Human[Human approval]
    PDP -->|DENY| Stop[Stop]
    Human -->|Approved and decision retrieved| Permit
    Human -->|Rejected or expired| Stop
    Permit -->|"acquired: true"| Execute[Execute action]
    Execute --> Evidence[Report evidence]
    classDef allow fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef approval fill:#fef3c7,stroke:#b45309,color:#78350f
    classDef deny fill:#fee2e2,stroke:#b91c1c,color:#7f1d1d
    class Permit,Execute,Evidence allow
    class Human approval
    class Stop deny

Loading

For a real action, preserve operation.id, the SHA-256 of the exact arguments in operation.hash, and the same Idempotency-Key across retries. After approval, retrieve the decision; before the external side effect, refresh the resource facts and consume the permit. Execute only if you receive acquired: true, then report the outcome with its execution_id.

Evidence records what your integration reports; it does not independently verify an external side effect. The execution contract details the flow and how to handle uncertain outcomes.

Inside the console

Approvals with context

Review the resource, reason, and validity period before resolving a sensitive operation.

From permit to outcome: view a decision record

Inspect the decision, issued permit, and evidence reported by the integration.

View the visual policy editor and regression cases

The editor lets you adjust rules, simulate outcomes, and save cases before publishing a version.

Integrations

Connect	Entry point	Guide
Your agent or backend	HTTP API and beta .NET, Python, and TypeScript clients	SDKs · OpenAPI
Remote MCP tools	/mcp/guard, with a dgk_ agent credential	Guard MCP
Administration through MCP	/mcp, with a separate dgm_ token	Local access
Model providers	Optional AI Gateway, compatible with OpenAI	AI Gateway
IDE and desktop client	Per-project connection	IDE integration

Provider keys stay in the gateway. Only metadata is sent for authorization. The legacy local MCP adapter remains available for compatibility; new connections use the remote Guard.

Documentation and support

To…	Start here
Configure variables, credentials, and ports	Development guide
Run the web demo, tests, or optional services	Components and validation
Open the documentation in English and Spanish	Documentation site
Understand the modules and contracts	Repository map · START_HERE
Stop services or diagnose a problem	Troubleshooting
Report a bug or suggest an improvement	Open an issue

Initial validation results and known issues

Initial validation covered the Release build, API, worker, /health, Python example, and documentation site. 534 unit tests, 36 security tests, 98 integration tests, and 15 demo agent tests passed.

At the time of that validation, unresolved issues included NuGet NU1902/NU1903 warnings for cryptographic dependencies and a width check in the desktop suite in headless mode. The acceptance script treats NU1903/NU1904 as errors. Review the results and commands before evaluating the project for production.

Contributing

Contributions go through a pull request to main, maintainer review, and the configured CI checks. Read the contribution guide for the fork and branch workflow, validation commands, and authorization guarantees. Use issues to propose changes or report reproducible bugs.

The project's own code does not yet include a license. Licenses for redistributed components are listed in the third-party notices.