Skip to content
HushLLM

Keep your data hushed. Use AI with confidence.

Sensitive data such as API keys and personal information (KVKK) is masked before it leaves the company. You see the real values; Anthropic doesn't. And Claude Code keeps working as if nothing happened.

We process your email address only for the early access list. Privacy notice

We'll write to this address when early access opens. HushLLM is a private POC today.

self-hosted · Anthropic API · early access

Animation: requests from Claude Code pass through the HushLLM detector; the AWS key inside each one turns into a locked placeholder before it reaches the Anthropic API. On responses coming back from Anthropic, the detector restores the placeholder so the developer sees the real value.

See what actually goes to Anthropic.

Pick a file Claude Code might read or a prompt with a pasted log, or type your own lines. On the way back, the values are restored for you.

4 secrets masked

editable

What Anthropic receivesafter HushLLM

DATABASE_HOST=db.internal
AWS_ACCESS_KEY_ID={{SECRET_AwsAccessKey_27f34c77fd8e}}
AWS_SECRET_ACCESS_KEY={{SECRET_HighEntropyToken_1e30b30c5272}}
STRIPE_KEY={{SECRET_StripeKey_c39aeb144c1d}}
GITHUB_TOKEN={{SECRET_GitHubToken_64d5ecf83238}}
LOG_LEVEL=info

Response from Anthropic

The access key in this file is defined as {{SECRET_AwsAccessKey_27f34c77fd8e}}. Deploy steps can keep reading this variable.

What you see (unmasked)

The access key in this file is defined as AKIAJDB4U6DRXHQ5UTBP. Deploy steps can keep reading this variable.

A simulation that runs in your browser; nothing you type is sent anywhere. It mimics, in simplified form, the patterns SecretMask recognises (AWS, GitHub, Slack, Stripe, Google, JWT, PEM, context-aware high entropy). Every key in the samples is fake.
Open in a new tab

Every request is on record. Prompt content stays off it.

Each request is recorded after it finishes: completed, blocked, failed or streamed. By default the record holds metadata only; raw secrets never reach it, and the caller's key is stored as a fingerprint.

The schema follows ECS and OpenTelemetry GenAI conventions. Records go to stdout, daily JSONL files or OTLP. An optional hash chain makes a deleted or altered line detectable.

audit.jsonl · schema v1 · examplelive example stream
  1. {"@timestamp":"2026-11-04T09:12:41.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:9c2e41d07ab35f18"}}}
  2. {"@timestamp":"2026-11-04T09:12:44.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":false}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:4b10e7a2c96d03f1"}}}
  3. {"@timestamp":"2026-11-04T09:12:47.000Z","event":{"action":"llm.request.blocked","outcome":"failure"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:e07d5a91b2c4f863"}}}
  4. {"@timestamp":"2026-11-04T09:12:50.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:9c2e41d07ab35f18"}}}
  5. {"@timestamp":"2026-11-04T09:12:53.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:4b10e7a2c96d03f1"}}}
  6. {"@timestamp":"2026-11-04T09:12:56.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":false}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:e07d5a91b2c4f863"}}}
  7. {"@timestamp":"2026-11-04T09:12:59.000Z","event":{"action":"llm.request.blocked","outcome":"failure"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:9c2e41d07ab35f18"}}}
  8. {"@timestamp":"2026-11-04T09:13:02.000Z","event":{"action":"llm.request.completed","outcome":"success"},"gen_ai":{"system":"anthropic","request":{"stream":true}},"hushllm":{"content_level":"metadata","caller":{"key_fingerprint":"sha256:4b10e7a2c96d03f1"}}}

Personal data gets masked before it leaves the company, too.

opt-in

The KVKK pack catches Turkish personal data in the files, logs and test data Claude Code reads, before any of it reaches Anthropic. It uses the same reversible masking as secrets: the developer sees the real value, Anthropic doesn't, and Claude Code's edits write the real value back to the file. It runs inside SecretMask and is turned on with one setting: HUSHLLM_KVKK_PACK=true.

musteri.json · example
What Claude Code read
"adSoyad": "Ayşe Örnek"
"tckn": "10000000146"
"iban": "TR20 0000 0000 0000 0000 0000 01"
What Anthropic receives
"adSoyad": "{{KVKK_PersonName_7c41e09a2b3d}}"
"tckn": "{{KVKK_Tckn_0a1b2c3d4e5f}}"
"iban": "{{KVKK_Iban_5e2f8a61c9d0}}"

The values are fake: test numbers with valid check digits. On the way back, the placeholders are restored to the real values.

What it masks

  • Validated by check digits

    Turkish ID number (TCKN), tax ID (VKN, with a context word nearby), TR IBAN (mod-97), payment cards (Luhn), IMEI (with a context word nearby).

  • By shape, with false-positive controls

    Mobile and landline numbers (real operator and area codes), e-mail addresses (example and noreply addresses exempt), licence plates, passport numbers, public IPv4 addresses (private and documentation ranges exempt).

  • Labeled fields

    Names, addresses and birth dates when they appear as the value of a label such as adSoyad: … or musteri_adresi: ….

What it blocks

  • Special-category data (KVKK art. 6)

    Health, religion and belief, race and ethnicity, political opinion, union membership, criminal record, biometric data and sexual life. When it appears as a label and value, such as kanGrubu: …, the request is never sent.

  • Bulk data

    If the newest message holds more than 20 personal-data values, for example a pasted customer table or a dump, the request isn't sent.

What Claude Code shows · HTTP 451

API Error: 451 HushLLM engelledi / blocked: Kişisel veri politikası — dışarı gönderilmesine izin verilmeyen veri bulundu / personal data policy — data that may not leave the network was found: KVKK.Health. İstek gönderilmedi; bu veriyi içeren mesajı geri alın (/rewind) veya yeni bir konuşma başlatın. / Nothing was sent; remove the message with this data (/rewind) or start a new conversation.

Mentioning a topic is not blocked: “write an appointment API for diabetes patients” is not health data. Claude Code shows a blocked request once and doesn't retry it, and the data itself is never shown anywhere.

Proof

For every request, the audit record keeps how many values of each data type were masked, blocked or flagged, and never the values themselves. With the optional Elasticsearch + Kibana profile, the “HushLLM — KVKK raporu” dashboard installs itself: masked personal data by type, blocked requests and their reasons, bulk-data events, a per-developer view and flagged rules.

Kibana · HushLLM — KVKK raporu

What the dashboard shows

  1. Masked personal data
  2. Blocked requests
  3. Bulk-data events
  4. Breakdown by type
  5. Block reasons
  6. Per-developer view
  7. Flagged rules

Records and the dashboard hold only data types and counts; values are never recorded.

Limits

  • Names and addresses in unlabeled free text aren't detected yet.
  • Image and PDF content isn't read.
  • claude.ai in the browser doesn't go through HushLLM, so it isn't seen.

On its own, the KVKK pack doesn't make your company KVKK-compliant. It is a technical measure that greatly reduces the risk of personal data leaving the country, and proves it.

The path of a request

Plugins are added and ordered through a registry in config; core code doesn't change. The contract is language-agnostic gRPC.

example request · POST /v1/messages
  1. Developerfix the deploy

    request: Types a task into Claude Code.

    response: Reads the answer with the real values.

  2. Claude Code

    request: Reads files for the task. The key in .env ends up in the request.

    response: Edits files with the real values, so the edits still match.

  3. HushLLM coreYARP, Polly

    request: Hands the request to each plugin in registry order. Credential headers never reach a plugin.

    response: Streams the reply back unbuffered.

  4. HijackGuardexperimental

    request: Looks for prompt-injection and jailbreak patterns.

  5. SecretMaskmask / unmask

    request: Swaps the key for a {{SECRET_…}} placeholder.

    response: Restores the original value, streaming included.

  6. Anthropic API

    request: Receives the placeholder. The key stays in the company.

    response: Streams the reply; the placeholder sits where the key was.

AuditLoggerobserver · after the request completesRecords each request once it finishes. Adds no latency and can't change traffic.llm.request.completed · content_level: metadata
What Claude Code read (.env)
AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE
What Anthropic receives
AWS_ACCESS_KEY_ID={{SECRET_AwsAccessKey_3f9a1c0b7d2e}}
What the developer gets back
AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE
HijackGuard and SecretMask run with FailureMode: Closed. AuditLogger only observes, so it never stops traffic.

Roadmap

None of this exists in HushLLM today. Items marked R&D are what we're working on now; the rest are planned. We're not giving dates.

  • Turkish name and address detection (KVKK)

    R&D · in progress

    A model that catches names and addresses in unlabeled free text. Today the KVKK pack catches them only as the value of a label such as “adSoyad: …”.

  • Turkish prompt-injection detection (HijackGuard)

    R&D · in progress

    Teaching HijackGuard to recognise attacks written in Turkish. Today's experimental HijackGuard doesn't catch Turkish attacks.

  • More tools and providers

    not built yet

    OpenAI, Gemini, GitHub Copilot, Cursor and other mainstream providers and tools. Today the focus is Claude Code and the Anthropic API; HushLLM sees Anthropic API traffic only.

  • LDAP/AD, Entra ID, Okta integration

    not built yet

    Tying gateway access to groups in the company directory, closing access automatically when someone leaves, and showing name and department in reports.

Frequently asked questions

How is it different from LiteLLM?

LiteLLM is a router that puts many providers behind one API, and it offers guardrail integrations. HushLLM is designed around security policy. Each plugin runs in its own process, traffic stops when a security plugin is unreachable, and masking is reversible, so Claude Code's file edits keep working. HushLLM supports only the Anthropic API today.

Is it open source?

Not today. HushLLM is a private POC and the license hasn't been decided. There's no public repo or documentation site yet.

Where does my data go?

When you self-host, HushLLM runs on your servers and by default connects out only to the Anthropic API. By default the audit log stores no prompt content and stays with you: stdout, a JSONL file or your own OTLP endpoint. If you'd rather not keep logs yourself, we can store them for you, priced by package. Either way, secrets are masked before they reach Anthropic and the audit log never sees raw secrets. This page loads no third-party scripts, fonts or analytics either.

Will it break Claude Code?

Our end-to-end suite runs on real Claude Code traffic and passes 30/30. Streaming passes through unbuffered. Masking touches only conversation content and is reversed on the way back, so tool use, prompt caching and extended thinking keep working.

Which tools does it work with?

Today, Claude Code and apps that call the Anthropic API directly; that's where our focus is right now. We plan to add integrations for mainstream APIs and tools such as OpenAI, Gemini and GitHub Copilot, but none of them exist yet and we're not giving dates. claude.ai in the browser doesn't go through HushLLM.

Does it run on a company server or on a developer's own machine?

Both. In a company setup HushLLM runs on a central server and the team's Claude Code traffic goes through it. In Inject mode the company's API key stays on the gateway, each developer gets a personal gateway key and usage is attributable per person. Plugins and the audit log are managed in one place. For individual use, the same package runs on a developer's own machine with Docker Compose and by default is reachable only from 127.0.0.1. In Passthrough mode your own Claude login (Pro, Max, Team, Enterprise) or API key passes through unchanged, and secrets are still masked before they reach Anthropic. With a personal card, Passthrough usage is attributable per person too.

Can we see who sent what while developers use their own Claude subscriptions?

Yes, with a personal card. Each developer keeps their own Claude login (Pro, Max, Team, Enterprise); on top of it, Claude Code sends a personal gateway card as an extra header (ANTHROPIC_CUSTOM_HEADERS), and the developer shows up by name in the audit log. The gateway checks the card and removes it, so it never reaches Anthropic or the plugins. The default is Optional: a request without a card is recorded without a name. Set it to Required and a request without a valid card is refused with 401, which turns the card into access control. Revoked cards are always refused. To make the gateway mandatory company-wide, use Claude Code's managed settings: ANTHROPIC_BASE_URL and allowedProviders: ["customEndpoint"] are set centrally, so developers can't route around the gateway.

Does it make us KVKK-compliant?

Not on its own. The KVKK pack is a technical measure that greatly reduces the risk of personal data leaving the country, and proves it; it doesn't replace a legal assessment. It doesn't detect names and addresses in unlabeled free text yet, doesn't read image or PDF content, and doesn't see claude.ai used in the browser.

How is it deployed?

Two ways. Self-hosted (on-premise): you install the core and the plugins you need on your own server or machine with Docker Compose. Traffic and logs stay with you; containers ship hardened and by default are reachable only from 127.0.0.1. Managed (all-in-one): we handle setup, operations and log storage, priced by the package you choose. In this model your traffic and logs pass through our infrastructure. We'll share setup steps and package details with early access.

Plans

Plans and prices aren't final yet; they'll be settled during early access. Nothing can be purchased today, and every button leads to the early access list.

FeatureTrialFree/ 1 dayTry the full Individual plan on one machine for a day.Join the list for a trialIndividual[PRICE]For a single developer on their own machine, with their own Claude subscription.Join early accessEnterprise[PRICE]A ready-made set for companies rolling Claude Code out to the team.Join early accessCustomQuoteFor needs the ready-made set doesn't cover. Price depends on scope.Join early access
SecretMaskReversible secret maskingIncludedIncludedIncludedFeatures, deployment model and number of users are agreed together based on your needs.
KVKK packMasking and blocking Turkish personal dataNot includedNot includedIncluded
AuditLoggerMetadata-only audit logIncludedIncludedIncluded
Audit log hash chainDeletion and tampering are detectableIncludedIncludedIncluded
Fail-closed, process-isolated pluginsIncludedIncludedIncluded
HijackGuardPrompt-injection detection, experimentalIncludedIncludedIncluded
Passthrough modeYour own Claude login or API key; per-person records with a personal cardIncludedIncludedIncluded
Inject modeCompany key on the gateway, personal gateway keysNot includedNot includedIncluded
Users11Team
Where it runsYour machineYour machineYour server or managed
Managed log storageLogs on our infrastructureNot includedNot includedIncluded

Put a gateway in place before Claude Code reaches the whole team.

Join the early access list. We'll email you when early access opens.

We process your email address only for the early access list. Privacy notice

We'll write to this address when early access opens. HushLLM is a private POC today.