Zero-Retention Enterprise Coding Endpoint — Kagi-Hosted Models, Intelligent Auto-Routing, Fixed + Consumption Pricing

kagi2025

Enterprises need a coding AI API that:

Never logs, stores, or trains on their code (contractual zero-day retention)
Routes each request to the optimal model (reasoning vs. speed vs. context) — like NeMo Switchyard
Is fully managed by Kagi — no BYOM infra, no VPC peering, just an API key (like Kagi Assistant today)
Costs predictably — fixed platform fee + token consumption

Kagi already hosts multiple models for Assistant. This feature exposes them as a privacy-first, enterprise-grade coding endpoint with intelligent routing built in.

Additional Context — How It Could Work

1. Privacy-First Guarantees (Contractual, Not Promises)

Guarantee	Implementation
Zero-day retention	Request/response payloads processed in-memory only; never written to disk; audit logs contain metadata only (timestamp, model, tokens, latency — no prompts, no code)
No training use	Contractual: Kagi does not fine-tune, embed, or train on any enterprise payload; upstream model providers (where applicable) receive only the forwarded request with no account identity, no PII
Data residency	EU (Frankfurt), UK (London), US (Virginia) — traffic never leaves selected region
Compliance	SOC 2 Type II, ISO 27001, GDPR DPA, HIPAA BAA (Enterprise tier)

2. Kagi-Hosted Model Pool (Curated, Rotated, Benchmarked)

Kagi operates a private model garden — same infra as Kagi Assistant, expanded for coding:

Model Tier	Examples (Kagi selects/rotates best-in-class)	Tags
Reasoning/Complex Coding	Claude latest, GPT latest, Qwen latest, let me choose	`reasoning`, `coding`, `long-context`
Fast/Everyday Coding	you pick some defaults, and I can cose as well	`fast`, `coding`, `cheap`
Specialist	you pick some defaults, and I can cose as well	`specialist`, `local-feel`
Long-Context/Repo-Scale	you pick some defaults, and I can cose as well	`long-context`, `repo-aware`

Kagi manages model lifecycle — updates, deprecations, benchmarking, capacity- No customer config — just works; advanced users can pin a tier via header

3. NeMo Switchyard–Style Auto-Router (Built-In)

Every request to model: auto (default) is classified and routed in <10 ms:

Signal	Used For Routing
Task type (coding / refactor / test-gen / doc / debug / chat)	Lightweight classifier
Complexity score	Heuristic + optional learned router (trained on anonymised task patterns only)
Latency budget	Header `X-Max-Latency-MS` or org default
Cost ceiling	Header `X-Max-Cost-USD` or org default
Context size	Auto-detected → routes to long-context model if needed
Session affinity	Conversation ID carries routing state (multi-turn consistency)

Routing rationale returned in every response headers:

X-Routed-Model: claude-latest
X-Routing-Rationale: "complex-reasoning, context-12k, latency-budget-3000ms"
X-Latency-MS: 1847
X-Cost-USD: 0.0042
X-Tokens-In: 3421, X-Tokens-Out: 847

One-click override via header: X-Force-Model: fast-tier or reasoning-tier

4. API Surface (OpenAI-Compatible + Coding Extensions)

# Standard chat/completions (works with Cursor, Continue, Cline, Codex, custom agents)
curl -X POST https://api.kagi.com/v1/coding/chat/completions \
  -H "Authorization: Bearer $KAGI_API_KEY" \
  -H "X-Max-Latency-MS: 3000" \
  -d '{"model": "auto", "messages": [...], "temperature": 0.2}'

# Coding-specific endpoints
curl -X POST https://api.kagi.com/v1/coding/edits \
  -d '{"model": "auto", "input": "...", "instruction": "Add unit tests"}'

curl -X POST https://api.kagi.com/v1/coding/completions \
  -d '{"model": "auto", "prompt": "...", "suffix": "..."}'

# Routing dry-run (returns selected model + rationale without calling model)
curl -X POST https://api.kagi.com/v1/coding/route \
  -d '{"messages": [...]}'

5. Pricing — Fixed + Consumption (Predictable, No Per-Seat)

Component	Price
Platform fee	as you have it today / month (includes: private endpoint, router, dashboard, audit logs, DPA, SLA 99.9%)
Consumption	as a consumer, I can top up, manage per user, per group, per org
Volume discounts	you decide..

6. Enterprise Dashboard

Usage analytics: tokens/model, cost/model, latency p50/p95/p99, routing distribution
Audit log export (SIEM-ready: timestamp, model, tokens, latency, routing decision — no payloads)
Team/org RBAC (SSO/SAML/OIDC, project-level budgets/quotas)
Routing profile presets (Kagi-maintained: "max-privacy", "max-speed", "max-reasoning", "balanced")
Note - Claude Teams and OpenAI don't offer this unless we are forced to sign up to enterprise edition, which does not make sense at all. I get what they are doing, but it is not nice.
---

Priority / Impact

Critical differentiator — no competitor offers managed NeMo-style routing + contractual zero-retention + Kagi-hosted models in one package
Massive enterprise TAM: every regulated org (finance, healthcare, defence, IP-heavy) building internal AI coding workflows
Natural extension of Kagi's privacy-first brand + existing Assistant model infra
High-margin recurring revenue: platform fee + consumption scales with adoption

Implementation Phases

Phase	Scope	Target
1	Private coding endpoint (single reasoning model, zero-retention, OpenAI-compat) + fixed+consumption billing	Q4 2026 would be nice 🙂
2	Multi-model pool + task classifier + basic routing (task-type → model tier)	Q1 2027
3	Full NeMo-style router (cost/latency/context constraints, session affinity, learned router) + observability	Q2 2027
4	SOC 2 Type II, HIPAA BAA, routing presets, enterprise SSO	Q3 2027

Why Kagi Wins This

Trust capital: Kagi's brand = privacy-first, user-aligned, no-ads, no-tracking
Existing model infra: Assistant already hosts multiple models with routing — extend, don't rebuild
NeMo Switchyard is Apache 2.0 — can embed/extend the Rust proxy directly for the routing layer
Enterprise sales motion: Kagi already sells Ultimate/Team plans; this is the logical "Enterprise AI" upsell

Bottom line: Enterprises are desperate for a coding AI endpoint that doesn't harvest their IP. Give them Kagi-hosted models, intelligent routing, contractual zero-retention — and charge a fair fixed+consumption price. They will queue up.

z64

We appreciate the suggestion, however, bear in mind that we have zero infrastructure prepared for a system like this. It would be a massive investment for us - in time, resources, and money - that we would need to allocate against everything else we are already trying to do with Search, Assistant, and more. So, I hope you understand that we're not going to change the direction of the entire company from a feedback ticket like this.

This is not to say "no" or "never", but to set expectations that it's not going to happen anytime soon, even if we wanted to do it. The best thing to do would be to get interested parties that you are referring to to comment here and to contact our support directly to express interest in helping fund some of these efforts in some way. If there is actual demonstrated demand for this, then we could consider making some serious plans.

I will say though it would be quite difficult as there are plenty of other companies already working on this who have a much, much further head start on it than us. In my personal opinion / perspective, at this rate, it is more likely that we would partner with one of these other companies than doing it ourselves; but still, quite a long shot.