Keep keys and customer data out of what your coding agent sends
Locke masks keys, tokens and customer PII in what your coding agent sends, on your machine.
- For
- Software engineers, platform and security teams
- Agents
- Claude Code, GitHub Copilot, Codex CLI, gemini-cli, curl, the official SDKs
- Runs on
- x64 Windows and Linux. macOS is coming soon.
- Screens
- AI websites, through the free Locke Extension. Coding agents and SDK calls, through a local proxy.
- Sent to Sonomos
- Account, licence check, settings. Never prompts, files or detections.
- Price
- Free for one person. Teams is $12.99 per user a month. Pricing
Where secrets leave with the prompt
28.65 million new hardcoded secrets in public GitHub commits in 2025, up 34%; 1.28 million for AI services, up 81% (GitGuardian, The State of Secrets Sprawl 2026).
What goes into an engineering prompt
What you write
Fix the 401 here: key sk_live_51Kd7Vq2Wm4Rx8Tn, host 10.20.0.14, reported by [email protected].
What the AI gets
Fix the 401 here: key AKY_o2x0, host IP4_k6z5, reported by EML_7sja.
AKYAPI keyIP4IP addressEMLemail
Invented example.
| The work | What the prompt carries | What the AI receives |
|---|---|---|
Fix a failing test | Source, config keys, a database URL. | Code and error; secrets masked. |
Paste a stack trace or log | Tokens, JWTs, customer emails, IPs. | The trace; tokens and customers masked. |
Debug a data issue | Customer names, emails, phones. | The data's shape; personal fields masked. |
Write an infra or deploy script | Cloud keys, a PEM key, internal IPs. | The logic; keys and IPs masked. |
Set up an MCP server or tool config | Inline API keys and tokens. | The structure; credentials masked. |
Draft a reply to a support ticket | Customer name, email, account. | The problem; the customer masked. |
What Locke finds in code, logs and config
- Cloud and API keys
- AWS, GCP, Azure and others
- Tokens
- JWT, bearer, Discord
- Private keys and secrets
- PEM keys, secret-looking strings
- Connection strings
- Database and credential URLs
- Customer data in logs and fixtures
- Names, emails, phones, SSNs, cards
- Network and device identifiers
- IP, MAC, IMEI, serials
Code goes as written, and so do UUIDs, which agents use as tool-call IDs.
When it is more than a leaked key
| Rule | What it asks | Source |
|---|---|---|
| Trade secrets | Information is a trade secret only if its owner took reasonable measures to keep it secret (Defend Trade Secrets Act). | Source: 18 U.S.C. § 1839 |
| CCPA | Customer data in a log or ticket is personal information; sending it to AI can be a governed disclosure. | Source: CPPA: CCPA regulations |
| PCI-DSS | A card number in a log carries PCI DSS duties into a prompt. See PCI DSS and AI. | Source: PCI SSC document library |
Trade secret, NDA and vendor questions belong to your security and legal teams.
How Locke works on a developer's machine
Finds
67 types of sensitive data, as each request leaves your computer.
Masks
Each value becomes a one-way token, the same every time, so the AI can follow who is who. Requests Locke cannot check are refused.
Stays on your device
Your secrets never reach our servers. Detection runs on your computer, not our servers. Its local log keeps type, time and destination, never the value. Security review pack
What Locke does not cover
- Cursor and Devin Desktop (formerly Windsurf) use a binary format Locke cannot read yet.
- Linux: host HTTPS on port 443 only; not containers, VMs, VPNs or other ports.
- Windows: programs that skip the system proxy bypass Locke (install guide).
- Ollama and LM Studio need the client pointed at Locke's proxy.
- One-way masking: code the model writes never holds the real key.
- No repository or commit scanning; keep a secret scanner.
- Hardware and civil work: pasted specs, not CAD files or drawings.
Full list: known limitations.
For a developer or an engineering team
- Free
- One person, full protection, no card, no time limit.
- Teams
- $12.99 per user a month, up to 20 seats, 30-day free trial, US only. One admin policy for every machine; the extension deploys by browser policy.
- Programs
- Startups: six months free (startup program).
Questions engineers ask
Does Locke work with Claude Code?
Yes. Claude Code calls api.anthropic.com over HTTPS, and Locke's proxy screens those calls once capture is set up. On Windows, start Claude Code from a terminal opened after Locke started, because programs that were already running do not see the proxy setting.
Which AI coding tools does Locke screen?
Claude Code, GitHub Copilot in VS Code and Visual Studio, the Codex CLI, gemini-cli, curl and the official SDKs, through the proxy on your machine, plus AI websites through the Locke Extension. Cursor and Windsurf send prompts in a binary format and are not screened yet.
Will masking break my agent's tool calls?
UUIDs, which agents use as request and tool-call identifiers, are left unmasked by default for that reason. Everything else Locke finds is masked, and you can let a data type through in Controls if a workflow needs it. Try a prompt in the Sandbox first to see exactly what the model would receive.
Does Locke replace secret scanning?
No. Secret scanners find keys in repositories, commits and CI logs. Locke works earlier, on what leaves your machine for a model. Use both.
Does Locke slow down requests?
Each request is screened on your machine before it leaves, which adds a step. We have not published latency figures yet and will once our benchmarks are done.
What does Locke send to Sonomos?
Sign-in, a periodic licence check and your settings. Prompts, code, files and detections stay on your machine.
Someone will paste a live key into a prompt. Locke can mask it first.
Reading for engineers
Recent news for engineers
- Cursor traces breach to exposed Fireworks key: agent traces with live credentials, then a malicious extensionNov 16, 2026
- Fireworks AI breach widens: Notion AI chat transcripts, billing data and model weights takenNov 15, 2026
- NEAR AI says an SGLang flaw exposed account data and agent logs despite confidential computingNov 10, 2026