Hushwire

zero-retention inference for coding agents

A private backend for Claude Code

Point Claude Code, Cline, OpenCode or Aider at one endpoint and get the best open coding models: Qwen3 Coder, Kimi K2.5, GLM and DeepSeek. Your prompts and code are never stored, and nothing goes to OpenAI or Anthropic.

~/.zshrc · that is the whole setup
export ANTHROPIC_BASE_URL="https://hushwire.vibiz.app"
export ANTHROPIC_AUTH_TOKEN="hw_your_key_here"
export ANTHROPIC_MODEL="hush-coder-large"
claude
nothing you type is stored

Ask a coding question. Try:

20 free messages a dayopen full screen →

The only row we wrote

Send a message. This panel then shows the exact metadata row saved for it: counts, model and timing. There is no column for your text.

three privacy proofs

Privacy you can check, not a promise in a footer

01

Zero retention, no toggle

Prompts, code and completions stream through memory and are gone. There is no column for them in our database and no log line that holds them.

request_meta: counts · model · ms · status

02

See exactly what we keep

Your dashboard lists every metadata row we hold about you, live, for the last 30 days. If it is not on that page, we do not have it.

GET /app/privacy

03

Never a big lab, by design

The proxy can only call open-weight models in a fixed alias table. Claude model names are mapped to open models, never forwarded to Anthropic or OpenAI.

claude-sonnet-* → hush-coder-large

for contractors and agencies

The answer to “which AI touches our code?”

Client NDAs and security questionnaires now ask which AI tools see their source. Hushwire gives you a one-page security summary and data-flow diagram to attach: what is stored (metadata only), what is not (your code), and which models run. Team plans put your company name on it.

Your agentClaude Code · Cline · AiderHushwire proxymemory only · no body logsOpen model hostQwen · Kimi · GLM · DeepSeekTLSTLSmetadata rowtokens · model · ms · 30 days✕ prompts✕ code✕ completionsnever written

models

Stable aliases for the best open coders

Point your agent at an alias once. When a better open model ships, the alias table is updated and announced; your config stays the same.

AliasModelContextQuota weight
hush-coderQwen3 Coder 30B A3B · Qwen
Fast, cheap default for edits, refactors and tool loops.
262k1×
hush-coder-largeQwen3 Coder Next · Qwen
The largest Qwen coder. Best for multi-file agent work.
262k3×
hush-kimiKimi K2.5 · Moonshot AI
Strong agentic coder with long tool chains.
262k4×
hush-glmGLM 4.7 · Z.ai
Reliable tool calling, good at planning and front-end work.
200k4×
hush-deepseekDeepSeek V4 Flash · DeepSeek
Very cheap general coder, good for quick questions.
128k1×

pricing

Flat monthly plans. Cheaper than Claude Max.

Predictable cost with a fair-use token allowance. No per-token bill surprises.

Free

$0forever

  • Browser playground, 20 messages a day
  • Small coding model
  • No API key, no card

Solo

$19/month

  • 1 API key
  • Every coding model
  • ~30M tokens a month (fair use)
  • Anthropic and OpenAI endpoints
  • Security summary PDF

Pro

most teams

$49/month

  • ~100M tokens a month (fair use)
  • Priority models: Qwen3 Coder Next, Kimi K2.5, GLM 4.7
  • 5 API keys with per-key spend caps
  • Higher rate limits (60 rpm, 4 streams)

Team

$149/month

  • 5 seats, 400M pooled tokens
  • Org keys and usage by member
  • Security summary with your company name
  • 120 rpm, 8 concurrent streams

7-day money-back guarantee on every paid plan · Top-ups: $5 per 10M tokens · Larger models count 3–4× against the allowance (see model weights).

Questions

Do you store my prompts or code?+

No. Request and response bodies are never written to our database or logs. Per request we keep one metadata row: key id, model alias, token counts, latency, status and time. It is deleted after 30 days, and your account’s “What we store” page shows every row.

Does anything go to OpenAI or Anthropic?+

No. Hushwire only calls the open-weight models in its alias table (Qwen, Kimi, GLM, DeepSeek). When Claude Code asks for a Claude model name, we map it to an open model; nothing is forwarded to Anthropic. The full upstream policy is on the security page.

Does Claude Code really work with it?+

Yes. Hushwire speaks the Anthropic Messages API at /v1/messages, with streaming and tool use, so you set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN and run claude. Open models are not Claude: expect strong but different results.

Why is there no free API trial?+

Free API keys attract key resellers and stolen cards, which would put everyone’s service at risk. Instead the playground is free with no sign-up, and every paid plan has a 7-day money-back guarantee.

What happens when I hit my monthly tokens?+

Requests return a clear error that says the quota is used and when it resets. Buy a 10M-token top-up for $5 or upgrade, and requests work again at once. We email you at 80% and 100%.

Can I put this in a client security questionnaire?+

Yes: that is what the security summary is for. Download it from your dashboard (Team plans get it with your company name) and attach it. It says what we store, where data flows and which models run. Hushwire is not SOC 2 or HIPAA certified, and the summary says so.