BLACKSAND

SELF-HOSTED · AGENTIC SECURITY RESEARCH

Let agents explore.
Keep the evidence.

A workspace for security research that connects the plan, the agents, and the proof. Your models. Your tools. Your infrastructure.

Explore the source Run it locally

GO + NEXT.JS / SQLITE OR POSTGRESQL / LOCAL OR E2B

01 / THE WORKSPACEBUILT TO BE INSPECTED

FROM OBJECTIVE TO OBSERVATION

Research is a process.
Not a chat transcript.

Blacksand turns an objective into a reviewable workflow. A planner creates work intents. Workers use your tools. Findings and evidence become part of the task’s persistent knowledge, ready for review and the next decision.

01

Set the scope

Define the goal and asset boundaries. Choose the model profiles, tools, and review rules for the task.

02

Follow the work

Inspect the planner’s intents, concurrent workers, conversations, approvals, and tool activity.

03

Review the proof

Keep findings, HTTP evidence bindings, reports, and retests together. Check the evidence before trusting a conclusion.

CONTROL

Your stack, connected.

Anthropic, OpenAI Chat Completions, and OpenAI Responses profiles. MCP servers, reusable skills, and an external agent adapter.

PERSISTENCE

A memory of the work.

Shared asset scope and task exploration graphs, backed by PostgreSQL or SQLite. Local execution or supported tool operations in E2B.

02 / GET STARTEDNO HOSTED ACCOUNT REQUIRED

YOUR MACHINE. YOUR WORKSPACE.

Start with
the source.

Use Go 1.26.10 and Node.js 22 or newer. A model provider is needed when you start agents.

Full setup guide
QUICK START / SQLITE
git clone https://github.com/arzumanabbasov/blacksand.git
cd blacksand
npm ci --prefix web

# Terminal 1: start the backend
BLACKSAND_DB_DRIVER=sqlite go run ./cmd/blacksand

# Terminal 2: start the interface
npm run dev --prefix web
Open localhost:3000 · Complete setup · Connect a model

THE FIELD JOURNAL

What worked.
What didn’t.
What comes next.

The track record is where real runs, findings, demos, experiments, and useful failures get written up—with the setup, outcome, and verification attached.

The journal starts empty. Results belong here when there is something real to show.

Open the track record
A TOOL. NOT A VERDICT.

Use controlled environments and targets you have permission to assess. Model output requires review. Local tools and remote sandboxes have different boundaries; inspect the workflow and the evidence.