AI Gateways (Zen, Kimi, Kilo)

One-key multi-model gateways as first-class NIKI providers.

AI Gateways: Zen, Kimi & Kilo

Three single-key gateways work out of the box: one key unlocks many models behind an OpenAI-compatible endpoint. Any other OpenAI-compatible endpoint works too: point a named provider’s base_url at it (unknown provider names still error, so typos can’t silently route to the wrong endpoint).

OpenCode Zen (`zen`)

Models from MiniMax, Zhipu (GLM), Moonshot (Kimi) and others, including free tiers, behind one key from opencode.ai:

[providers.zen]
base_url      = "https://opencode.ai/zen/v1"
api_key       = "..."                    # Or set OPENCODE_API_KEY env var
default_model = "kimi-k2.5"

Kimi Code (`kimi`)

Moonshot’s coding gateway (docs). Speaks both OpenAI-compatible chat completions and the Anthropic Messages API; NIKI uses the OpenAI-compatible surface:

[providers.kimi]
base_url      = "https://api.kimi.com/coding/v1"
api_key       = "sk-..."                 # Or set KIMI_API_KEY env var
default_model = "k3-256k"

KiloCode Gateway (`kilo`)

Kilo’s unified inference API (docs): hundreds of models behind one key, addressed as vendor/model:

[providers.kilo]
base_url      = "https://api.kilo.ai/api/gateway"
api_key       = "..."                    # Or set KILO_API_KEY env var
default_model = "anthropic/claude-sonnet-4.5"

Setup

niki init --interactive       # pick one provider, guided, with key validation hints
export OPENCODE_API_KEY=...   # runtime provider construction reads supported env vars
niki providers check          # the `check` subcommand is required
niki doctor                   # reports env/keyring status

All three also work with zero config: set the provider’s environment variable and niki chat auto-detects it. niki auth can store a keyring entry, but runtime provider construction does not currently resolve keyring values.

AI Gateways (Zen, Kimi, Kilo)