Guardana Control MCPサーバー

Guardana Control

Supervise your AI agents. Decide what runs.

An open-source AI agent supervision framework that checks recorded runs against procedures and uses policy to decide connected actions.

Website · Try the demo · Docs · Status · Roadmap · Contributing · Security

Try the alpha

Run the local demo: allow a call, block another and approve one exact action, without Go or a model key. Connect your MCP setup next.

The gateway is experimental. Caller identity comes from configuration. The evidence trail is not tamper-evident. Do not deploy this as a security boundary. Status is the inventory of what exists.

What you can do today

• Hold writes for an expiring approval of one exact action: starter policies.
• Let your authorization service veto configured calls through the AuthZEN API.
• Let a tool of your own ask before it acts, through the Enforcement API.
• Import runtime traces and egress proxy logs, inspect coverage, check a recorded run against its procedure, and stop its later calls where a signed route allows: supervision.
• Under approvals.authority: signed, take approvals, pauses, runs and stop routes only from operators named in a signed roster: who may approve.
• Block a tool whose definition no longer matches its reviewed fingerprint, except under OBSERVE: MCP gateway.

Supervision and reaction run when you invoke them. A stop never interrupts a running operation. Paths around the gateway stay outside its control.

The 1.0 goal: observe, supervise, react

Guardana Control is a control and supervision layer for AI agents. Its 1.0 goal is planned (roadmap):

• Observe actions and outcomes across connected tools, APIs, runtimes, network and process sources. State each path's coverage, source trust and gaps.
• Supervise runs against procedures and permissions. Report deviations, denied actions tried another way, skipped work and uncertain outcomes.
• React within scoped authority: inform, alert, request review, hold an exact action for approval, pause calls or stop further actions.

Full observation means coverage of connected paths with unknowns visible, not every action visible or controllable. Sensors report; enforcement points decide before effects. Read the product model.

How a call is decided

The MCP gateway consults signed policy before a tool call. An external PDP can veto but never widen local authority (experimental). Its silence blocks the calls it governs in an enforcing mode. It receives action metadata, not arguments. Evidence names the policy version and records the outcome.

flowchart LR
    accTitle: How a tool call is decided today
    accDescr: The enforcement point decides an agent's proposed call from the built-in policy engine, an external decision point that can only veto, and the approval authority when a person has to approve. An allowed call goes to the tool or API. Decisions are appended to the evidence, which is exported over OpenTelemetry; a call whose record cannot be appended is blocked, unless the operator let a read run unrecorded.
    AG[Agent] --> PEP[Enforcement point]
    POL[Built-in policy engine] --> PEP
    PDP[External PDP] -.->|can veto| PEP
    APR[Approval authority] --> PEP
    PEP -->|allowed call| TOOL[Tool or API]
    PEP --> EV[(Append-only evidence)]
    EV --> OTEL[OpenTelemetry export]
    classDef accent fill:#E6F4F2,stroke:#0B8F80,color:#0F1115
    class PEP accent

Decisions do not wait for evidence export. A full spool blocks calls unless the operator explicitly let a read run unrecorded.

Verdict	Under ENFORCE
ALLOW	Proceed.
DENY	Block.
REQUIRE_APPROVAL	Hold this exact action for one approval.
ALLOW_WITH_OBLIGATIONS	Proceed under enforced conditions, unless advisory.
INDETERMINATE	No decision; never an allow.

OBSERVE records decisions; pauses, stops and halts still block calls. No model grants authority. Content capture is off by default; hidden reasoning is never stored.

Connected channels towards 1.0

The planned target connects enforcement and observation through one vocabulary (ADR-0039). An Enforcement API (experimental) is distinct from AuthZEN vetoes: custom tools ask before acting and report completion or abort. A cooperating client enforces its answer; the receiver does not.

flowchart TB
    accTitle: The connected-channel target for 1.0
    accDescr: The planned 1.0 target connects tools through the MCP gateway and cooperating clients over an enforcement API, and observations from runtimes, proxies and processes. A supervisor checks procedures and raises findings. Notifiers report them; scoped reactions stop later calls only at connected enforcement points. Coverage and missing observations remain explicit.
    AG[Agents]
    subgraph EP[Enforcement points]
        MCP["MCP gateway, today"]
        HK[Custom tools and framework hooks]
    end
    subgraph SN[Sensors]
        TR[Runtime traces and logs]
        PX[Proxy access logs]
        PR[Process events]
    end
    TL[Tools and APIs]
    EV[(Evidence)]
    OB[(Observations)]
    SV["Supervisor: procedures, detectors, coverage"]
    subgraph OUT[Outputs]
        AL["Alerts: webhooks, chat"]
        EX["OpenTelemetry, SIEM"]
        GR[Run graph]
    end
    RE["Reaction: scoped stops where allowed"]
    AG --> MCP
    AG --> HK
    MCP -->|allowed calls| TL
    HK --> TL
    AG -.-> TR
    AG -.-> PX
    AG -.-> PR
    MCP --> EV
    HK --> EV
    TR --> OB
    PX --> OB
    PR --> OB
    EV --> SV
    OB --> SV
    SV --> AL
    SV --> EX
    SV --> GR
    SV -.-> RE
    classDef accent fill:#E6F4F2,stroke:#0B8F80,color:#0F1115
    class MCP,SV accent

Install

Each release has archives of guardana-gateway and guardana-control for Linux and macOS on amd64 and arm64. Download one from releases, check it against the signed checksums.txt (RELEASING.md) and put the binaries on your PATH.

With Go 1.27.4 or later, build from source (version dev):

go install github.com/guardana/control/cmd/guardana-gateway@latest
go install github.com/guardana/control/cmd/guardana-control@latest

Related project: Guardana

Guardana verifies AI systems before and after deployment. Control observes and controls agents while they run. The projects are independent.

flowchart LR
    accTitle: Where Guardana and Guardana Control sit in a system's life
    accDescr: Guardana checks a build before its release. While the released agents run, Guardana Control observes their runs and decides connected calls. After the release, Guardana compares what changed.
    B[Build] --> G1["Guardana<br/>checks before release"]
    G1 --> R[Release]
    R --> C["Guardana Control<br/>observes and controls runs"]
    C --> G2["Guardana<br/>compares after release"]

Governance · Decision records · Code of conduct · Trademarks