Does Cloak see my prompts?

Only if you choose the hosted plan, and only what the dashboard stores. Self-hosted, nothing leaves your machine and no telemetry is sent anywhere. You can also run the CLI in mask-only mode and never connect it to a dashboard at all.

Which agents and providers work with it?

Anything that speaks the OpenAI or Anthropic wire format: Claude Code, Cursor, Codex, Aider, the OpenAI SDKs and more. Providers include OpenAI, Anthropic, Ollama and any OpenAI-compatible endpoint such as Groq, Together or a local server.

Does masking break my prompts?

No. A value is swapped for a stable, readable token like [EMAIL_1], and the same value always maps to the same token within a field, so the model can still reason about it. You choose per pattern whether to redact, flag or block.

How is this different from shim-cli?

Cloak keeps the same local-first idea but adds a real dashboard: stats, cost and a request-by-request replay of any session, plus a native MCP server so agents audit themselves. It is a proxy you point any SDK at, not a set of per-client hooks.

Is it a data-loss-prevention boundary?

Treat it as a guard, not a boundary. Detection is pattern and checksum based and can miss a value, especially names and addresses with no key around them. It narrows exposure and gives you an audit trail — it does not replace a security review.

Yes. The whole thing is one Next.js app and a Node CLI. A Dockerfile and a docker-compose file ship in the repo, and Coolify deploys it in one click with your own database and domain.

Apache-2.0Open source · no telemetry · self-hosted

See what your agents send. Keep the secrets home.

Cloak is an open-source proxy for coding agents. It logs every model request, masks secrets and personal data before they leave your machine, and audits each session from one dashboard.

Get started Explore the live demo

Works withOpenAIAnthropicOllamaAny OpenAI-compatible

cloak proxy --mask127.0.0.1:8787

POST /v1/messages200$0.092
claude-sonnet-4.6 · 24,880 tok[OPENAI_KEY_1][ANTHROPIC_KEY_1][DATABASE_URI_1]
POST /v1/chat/completions200$0.078
gpt-5.2 · 15,600 tok[CARD_1][IBAN_1]
POST /v1/chat/completions200local
llama3.3:70b · 40,840 tok[EMAIL_1][PHONE_1]
POST /v1/chat/completions429—
gpt-5.2 · 400 tokpassed clean

How it works

Set up in a minute, without changing your workflow

01
Install the CLI
One command installs the proxy. No config file to learn, no daemon to babysit — it only runs while you ask it to.
02
Point your agent at it
Set one base URL. Cloak forwards every byte unchanged, or rewrites the sensitive ones when masking is on.
03
Watch and audit
Sessions stream into the dashboard. Open one to replay every request, the tokens it cost and the secrets it carried.

terminal

# 1. run the proxy
npx cloak proxy --mask --port 8787

# 2. point your agent at it
export OPENAI_BASE_URL=http://127.0.0.1:8787/v1
export ANTHROPIC_BASE_URL=http://127.0.0.1:8787

# 3. that's it — sessions stream to your dashboard
npx cloak proxy --mask --log $CLOAK_URL --key $CLOAK_KEY

What you get

Everything you need to see, mask and justify what your agents send

One proxy, every provider

OpenAI, Anthropic, Ollama and any OpenAI-compatible endpoint. Point your SDK at Cloak and change nothing else — same requests, same streaming.

OpenAIAnthropicOllamaAny OpenAI-compatible

Masked before they leave

API keys, tokens, emails, cards, IBANs, IPs — plus your own regex. Sensitive values become stable tokens like [EMAIL_1] before the model ever sees them.

A dashboard, not a scrollback

Every request logged with tokens, latency, cost and detections. Replay any session, request by request, and see exactly what left your machine.

Agents that audit themselves

A built-in MCP server exposes get_logs, mask_text and audit_session, so your coding agent can check what it just sent without leaving the terminal.

mask_text()get_logs()audit_session()

Cost and drift, visible

Token spend per session, per model, per day. Catch a runaway agent loop before it shows up on the invoice at the end of the month.

Open source and self-hosted

Apache-2.0, no account required, no telemetry. Run it locally or deploy the dashboard to your own server with one Docker command on Coolify.

The masking engine

Secrets become [TOKENS] the model can still reason about

Built-in patterns for keys, tokens, cards, IBANs, emails and IPs, plus your own regex. The same value always maps to the same token, so nothing breaks. Edit the text on the left — it runs locally.

Input — what the agent would send

Runs locally in your browser. Nothing here is sent anywhere.

Output — what the model receives6 masked

User report — please summarise before I paste it to the model.

Account: [EMAIL_1] (preferred contact)
Card on file: [CARD_1]
IBAN: [IBAN_1] 00
OPENAI_API_KEY=[OPENAI_KEY_1]
Authorization: Bearer [JWT_1]
Postgres: [DATABASE_URI_1]

CARDDATABASE_URIEMAILIBANJWTOPENAI_KEY

Why it matters

A request carries more than your prompt

The system prompt, the tools array and every file the agent read along the way. When it opens a .env, those keys go to the model too — often with nothing in the session to tell you.

In a normal session	With Cloak
Guesswork	Full request log
Sent to the model	Masked or blocked
End-of-month invoice	Live, per request
Nothing	Exportable audit trail
Trust the client	Your machine or your server

Rows describe the difference in visibility, not a guarantee.

Model Context Protocol

Agents that can audit themselves

Cloak ships an MCP server. Point your agent at it and it can check what it just sent, mask a value before a request, or summarise a session — without you leaving the terminal.

mask_textReturn the masked version of any text, with detections.
get_logsList recent proxied requests, filtered by provider or session.
audit_sessionSummarise one session: tokens, cost and a risk score.

.mcp.json

{
  "mcpServers": {
    "cloak": {
      "url": "https://your-cloak.app/mcp"
    }
  }
}

Pricing

Open source forever. Hosted when you want it handled.

The tool is free and self-hostable. The hosted plan is for teams who want the dashboard managed for them, billed per seat, with no usage caps per event.

Open source

The full CLI and dashboard, on your machine.

€0forever

Multi-provider proxy (OpenAI, Anthropic, Ollama)
Masking engine: built-in patterns + your own regex
Dashboard with logs, stats and session replay
MCP server for coding agents
Self-host with Docker or Coolify
Apache-2.0, no telemetry, no account

Start for free

Cloud

14-day trial

The hosted dashboard, managed for your team.

€5per seat / month

Everything in Open source, hosted and updated for you
Unlimited usage — no request or seat caps per event
Team workspaces and shared audit trail
90-day log retention with CSV / JSON export
Priority support from a human
One-click deploy, TLS and backups included

Start 14-day trial

Paying yearly? See every plan detail.

FAQ

Straight answers, including the limits

Cloak is a guard, not a data-loss-prevention boundary. If a question isn't here, the source is on GitHub — read it.

See what your agents actually send

Run the proxy locally in a minute. No account needed to try it, and the whole tool is open source.