OpenAI Compatible Private API: Drop-In Privacy for Existing Code

Why compatibility matters

OpenAI-compatible private API endpoints let you swap base_url without rewriting SDK calls. Migration cost drops from weeks to hours when privacy requirements change. https://secrypt.space/v1 — familiar shape for scripts, agents, and IDE plugins.

Compatibility is ergonomics; privacy is policy. You need both. Open-source frameworks hardcode OpenAI environment variable names; fighting that shape wastes contributor goodwill when privacy policy forces a vendor change.

Contributor pull requests that only swap base_url merge faster — lower friction keeps privacy migration from stalling in git limbo while legal reviews the new vendor. Security reviewers should treat base_url swaps as infrastructure changes worthy of changelog entries, not silent Friday deploys. When legal approves Secrypt, developers can ship the same week because client shapes stay familiar across Python, Node, and IDE agents.

Minimal integration diff

Set base URL to https://secrypt.space/v1. Model: cipher (alias secrypt-cloud). Authorization: Bearer sk_sec_…. Point OpenAI SDKs at the new base and keep temperature and message arrays identical for A/B tests.

Run three production-shaped prompts before cutting over billing. Keep existing OpenAI env var names and swap values only — reduces developer confusion during cutover. Feature-flag old and new base_url during a canary so rollback stays one flag flip. Staging should mirror production secrets patterns.

Parity prevents “works on my machine” cutover failures when the first real user hits rate limits or credit exhaustion. Document the exact environment variables your platform team must set in staging versus production. Include a rollback snippet that restores the legacy base_url in one commit so on-call engineers are not improvising during the first outage.

• Create key in Account → API keys
• Never expose keys in frontend bundles
• Use environment variables or secret managers
• Handle 401 and 429 explicitly

Privacy beyond compatibility

Cipher inherits Secrypt’s stance: conversations are not used to train models; no ad profiling from messages. Read live policy for API-specific retention. Compatibility does not encrypt prompts from the operator — HTTPS protects transit, not your trust model. Secrypt does not claim HIPAA or SOC2.

Add a privacy review gate in CI for new LLM environment variables, same as you would for new database URLs. Architecture review slides should include one honest bullet: plaintext is processed by Secrypt during inference. When legal asks about subprocessors, compatibility does not replace paperwork. List Secrypt in your risk register with clear scope boundaries instead of assuming OpenAI-shaped APIs imply OpenAI-shaped compliance.

Add Cipher to your data processing inventory with plain language about plaintext inference. Compatibility does not shrink subprocessors or retention — it only shrinks refactor time.

Billing model

Chat subscriptions and API usage bill separately on Secrypt. 01 from prepaid credits. Buy credits from Account or earn via Invite & Earn.

Predictable per-request pricing beats surprise token math for many internal tools. Finance teams understand requests times one cent faster than token calculus — use that clarity in budget meetings. Prepaid credits suit bursty agent prototypes.

Finance often approves petty-cash-style credit purchases faster than uncapped token invoices tied to an unknown model release cycle. Finance spreadsheets should model worst-case agent loops explicitly: requests times one cent adds up when cron jobs fan out. Prepaid credits make burn visible weekly instead of surprising token invoices at month end.

Chat subscription vs API

Use the chat UI for interactive private conversations. Use the API when cron jobs, backends, or Cursor need programmatic access. Same assistant stack; different metering surfaces.

Monitor both dashboards during pilots. Do not assume chat tier limits apply identically to API without reading current docs. Engineering managers should review both meters in weekly ops review during the first migration month — surprises come from forgetting API exists.

A Slack bot posting weekly Cipher spend creates visibility and prevents month-end shock when an agent loop runs hotter than expected. Product managers often forget API meters while celebrating chat adoption. During migration month, post both numbers in the team channel every Friday until the new baseline feels normal.

Cursor and Continue patterns

In Cursor, set the OpenAI-compatible base URL to https://secrypt.space/v1, and select model cipher. Keep keys out of repositories — use environment injection supported by your IDE. Test on a branch before enabling agent mode on proprietary monorepos.

Pin a working IDE configuration in your internal wiki quarterly — UI moves buttons between releases and onboarding copies stale screenshots. New hire checklists should include an IDE screenshot step. That small habit reduces repeated Slack pings asking which model id to use.

Continue users mirror Cursor steps: custom OpenAI base URL, sk_sec_ key, model cipher. Screenshot both IDEs if your team splits across tools — UI labels differ even when endpoints match.

Error handling in production

Handle 429 responses and credit exhaustion gracefully. Retry with backoff; never infinite loops on billing errors. Log request ids, not full prompts. Redact user content in observability stacks.

Rotate keys if a log aggregator accidentally ingested secrets. User-facing errors should link to a status page, not echo prompt fragments — UX details matter during incidents. Subscribe to Secrypt status updates if your product depends on Cipher. User trust during outage is easier when you communicate proactively instead of blaming the model.

Circuit breakers should stop runaway agents from draining credits while retrying the same toxic prompt. Pair technical backoff with paging rules when daily allowances exhaust before Pacific morning standup.

Security review talking points

Lead with facts: no training on conversations, keys are secrets, data minimization is required, subprocessors are listed in policy. Hosted inference means plaintext processing — document that honestly for risk registers. Offer local fallback for highest-classification data when reviewers ask about custody.

Bring a printed policy excerpt to review — verbal claims fade; paper persists. Walk reviewers through Settings live. Offer reviewers a live call walkthrough of Settings export and clear.

Demos reduce back-and-forth email chains asking whether users can actually delete threads without opening a support ticket.

Ship a migration in a day

Morning: create sk_sec_ key, point staging at Cipher, run regression tests. Afternoon: canary ten percent of traffic, watch error rates and credit burn. Evening: flip production base_url and revoke legacy keys in the team channel. m.

is predictable. Keep one engineer on the old endpoint for twenty-four hours if your SLA allows quick revert. Celebrate cutover in release notes internally. Cultural signal that privacy migration finished matters as much as the technical diff.

End the day by revoking legacy API keys in the shared team channel. Public revocation signals cutover completion better than a merged pull request nobody noticed.

Related guides

Try Secrypt

Secrypt is private, uncensored AI chat. No training on your messages. Open a thread when you need discretion more than theater.

Open Secrypt chat   View pricing