Mandu Security
Mandu Security builds Mandu, an open-source local security layer that lets people run coding agents beside real keys and sensitive files while showing the AI masked placeholders. Approved programs and destinations can still receive real values under user-controlled rules, while Mandu keeps the data local and requires no Mandu account, server, API key, or telemetry.
At a Glance
- Individual developers using AI coding agents
- Early teams whose agents run next to real credentials
- Software engineering and DevOps teams handling production keys
- Security-conscious organizations concerned about AI-provider data leakage
- +1 more
AI Tools by Mandu Security
(1)
Mandu
Secret Masking for Coding Agents
Discussions
No discussions yet
Be the first to start a discussion about Mandu Security
Latest News
Mandu 0.2.0-beta.7 released as a public beta
Mandu published Secret kept, task done, reporting a 450-session evaluation against multiple AI-agent protections
Mandu published Placeholders, not keys, explaining its architecture and updated evaluation results
Mandu published Why a sandbox isn't enough to keep secrets from your AI provider, comparing security approaches
Products & Services
Open-source local software distributed through npm as @mandusec/mandu and via an installer. It runs Claude Code, Codex, and other coding agents beside real credentials and sensitive files, presenting placeholders to the AI while selectively unwrapping values for approved programs, hosts, or user-approved actions.
Market Position
Mandu positions itself as an operating-system-level, placeholder-based control that protects the model view, program access, and network egress with one policy. Its engineering comparison contrasts it with agent sandboxes such as Claude Code sandboxing, Codex sandboxing, NVIDIA OpenShell, and nono; credential masks and autofill; AI gateways/DLP products such as LiteLLM, PrivAiTe, Bedrock Guardrails, Lakera/Check Point, Nightfall, and Prompt Security; and redesigned-agent approaches such as CaMeL and FIDES. The differentiator claimed is that programs can still use the real value while the AI sees only the wrapper, rather than simply blocking access or redacting model traffic.
Founding Story
Mandu Security says it started from systems-security research on keeping AI agents away from data they should not hold, building on the Prompt Flow Integrity and DualView work. Its initial vision was to turn that research into an operating-system-level tool that works next to existing agents rather than inside one agent framework; the name refers to a Korean dumpling, with the agent handling the wrapper and the user's secrets remaining the filling.
Business Model
Revenue Model
The publicly described product is open-source and free, runs locally on the user's computers, and does not require a Mandu account or Mandu server. Mandu Security is also working with a small group of early design-partner teams; no paid commercial model is stated on the reviewed sources.
Pricing Tiers
Open-source beta; everything runs on the user's own computers.
Target Markets
- Individual developers using AI coding agents
- Early teams whose agents run next to real credentials
- Software engineering and DevOps teams handling production keys
- Security-conscious organizations concerned about AI-provider data leakage
- Researchers working on AI-agent security and prompt-injection defenses
- Running Claude Code, Codex, OpenCode, OpenClaw, or Goose on software projects containing .env files and API credentials.
- Debugging applications while preventing database, payment, cloud, or GitHub credentials from entering model context.
- Deploying applications where tools need real credentials but the coding agent should see only placeholders.
- Using SSH keys and cloud logins through explicitly allowed programs.
- Handling confidential documents such as contracts, health records, board memos, and tax paperwork near an agent.
- Browser and web-agent workflows in which credentials should reach an intended site but not an AI provider or lookalike destination.