AI gateway · pure Go · single binary

A thousand streams of light,
one gateway.

Firefly is an OpenAI-wire-compatible reverse proxy that carries the models you use behind a single endpoint: high-concurrency SSE streams, key rotation, and load balancing that keeps traffic flowing in the dark.

# one command, then it simply glows
$ curl -fsSL https://raw.githubusercontent.com/dickymuliafiqri/firefly/main/install.sh | sh
$ firefly -config-dir=/etc/firefly
listening on 0.0.0.0:8080 · embedded dashboard at /

Built for the hard parts.

Every firefly carries its own lantern. Here is how Firefly keeps traffic flowing when providers flicker, keys run dry, or hosts go dark.

Virtual Combos

Group many upstream models under one public name and let Firefly pick the lane. Choose least_inflight, round_robin, or a failover chain, so the pool absorbs bursts and dead providers without clients ever noticing.

  • – per-request least-inflight routing
  • – automatic reroute when a lane fails
  • – one model name, many upstreams
ClientFireflyOpenAIGoogleAnthropicMistral

Speaks to your providers.

One endpoint, wire-compatible with OpenAI, in front of your providers. Adapters translate, authenticate and stream natively, with no glue code and no env juggling.

OpenAI
Google
Anthropic
xAI
Meta
Mistral AI
DeepSeek
Ollama

Backed by builders, built in the open.

Firefly is free and open-source under the MIT license. Maintained with care, contributions welcome.

Sponsors

Individual supporters

Ayu Lestari· ML EngineerBima Prasetyo· Platform EngineerSinta Maharani· DevOps EngineerRizky Ramadhan· Backend DeveloperNadia Putri· AI ResearcherDimas Angga· Site Reliability EngineerAyu Lestari· ML EngineerBima Prasetyo· Platform EngineerSinta Maharani· DevOps EngineerRizky Ramadhan· Backend DeveloperNadia Putri· AI ResearcherDimas Angga· Site Reliability Engineer

Used by

Nimbus AI· high-concurrency inferenceKopiko Labs· multi-region inferenceDatamaya· streaming analyticsNimbus AI· high-concurrency inferenceKopiko Labs· multi-region inferenceDatamaya· streaming analyticsNimbus AI· high-concurrency inferenceKopiko Labs· multi-region inferenceDatamaya· streaming analyticsNimbus AI· high-concurrency inferenceKopiko Labs· multi-region inferenceDatamaya· streaming analytics
Awanpay· fraud-detection copilotsTerraworks· field-agent assistantsLangit Labs· research eval harnessAwanpay· fraud-detection copilotsTerraworks· field-agent assistantsLangit Labs· research eval harnessAwanpay· fraud-detection copilotsTerraworks· field-agent assistantsLangit Labs· research eval harnessAwanpay· fraud-detection copilotsTerraworks· field-agent assistantsLangit Labs· research eval harness

The names above are placeholders showing how this wall will look. Your team could be the first real one here.

Single binary · MIT licensed

Light up your gateway.

Drop Firefly in front of your OpenAI, Anthropic, or local Ollama instances. No database required, zero runtime dependencies, and high concurrency powered by Go.

# run via docker or grab the binary
docker run -p 8080:8080 ghcr.io/dickymuliafiqri/firefly:latest