Where Tencent Cloud AIGate is heading. This is a living, high-level plan: it sets the direction only — how each item is implemented is decided when work starts. Day-to-day work is tracked in issues, and shipped work is recorded in CHANGELOG.md.
The current foundation: a dual-format data plane (OpenAI + Anthropic), tiered token quotas with usage accounting, a credential vault with BYOK, signed webhook events, and idempotent import plus self-service keys. See the changelog for details.
Three directions for the next release, all about making model calls more resilient and the gateway easier to extend:
- Model fallback — when a call cannot be served by the primary model, the gateway falls back to a backup automatically: transient upstream failures no longer break the request.
- Auto-routing — choose among a key's models by configurable policy instead of a fixed binding order: cost, latency or capability preferences become configuration.
- Plugin support — the plugin capability serves auto-routing: routing strategies are provided as plugins, so custom selection logic can be added without changing the gateway core.
- Finer-grained quota management — more flexible quota settings for users and groups.
- Observability — metrics for requests, errors and latency, ready for dashboards and alerting.
- Quality-aware routing — pick models by measured quality, not only cost or latency.
- Multi-instance coordination — consistent quotas and routing state across replicas.
- Web UI — AIGate is a pure API service by design.
- MCP proxying and session management — those belong to adjacent layers.
- Issues are the source of truth for ongoing work; this file is revised during release planning.
- API and configuration stay backward compatible within v0.x unless a change is explicitly flagged in the changelog.
- Feedback and design discussions are welcome via the project's issue tracker.