AI Adapter

A Rust proxy that exposes an OpenAI Responses API interface, translating to various upstream backends (DeepSeek, OpenAI, Anthropic, XiaomiMimo) with full streaming SSE support. Built for Codex CLI and similar tools hardcoded to OpenAI.

Codex / Client          AI Adapter :9090          Upstream
─────────────          ─────────────────          ────────
/v1/responses   ──▶    translate/            ──▶  DeepSeek Chat
(Responses API)        deepseek/chat.rs           (Plan A)

/v1/responses   ──▶    translate/            ──▶  DeepSeek Anthropic
(Responses API)        deepseek/anthropic.rs      (Plan B)

/v1/responses   ──▶    translate/            ──▶  OpenAI Chat
(Responses API)        openai/chat.rs

/v1/responses   ──▶    translate/            ──▶  Anthropic
(Responses API)        anthropic/convert.rs

/v1/responses   ──▶    translate/            ──▶  XiaomiMimo Chat
(Responses API)        xiaomimimo/chat.rs         (+ Anthropic)

Features

• Multi-provider: Configure multiple upstream providers in one YAML config, switch by model name per request
• Vendor isolation: src/translate/{deepseek,openai,anthropic,xiaomimimo}/ — each vendor gets its own module
• Bidirectional streaming: Full SSE event translation (Chat ↔ Responses, Anthropic ↔ Responses)
• DeepSeek Plan A/B: Chat (thinking disabled) and Anthropic (tool_use merging) protocols
• Reasoning cache: SQLite-backed reasoning_content store for multi-turn thinking compliance
• State management: SQLite for sessions, reasoning, and request logging; auto-migrates from legacy redb
• Tool calling: Full function call streaming, flat & nested tool format support
• Compaction: POST /v1/responses/compact for Codex CLI context management
• Config-driven: YAML config files with providers and modelAliases, plus CLI/env overrides
• Auto-detect format: format is optional — inferred from base_url (anthropic/openai-chat/responses)
• HTTP/2 upstream: Transparent ALPN negotiation for supported backends
• Health & models API: /health with uptime info, /__/models listing configured aliases
• Error dumps: Saves failed exchanges with auth redaction
• Structured logging: Human-readable to stderr, JSON to daily-rotated files

Quick Start

# Build
cargo build

# Run with a config file
./target/debug/ai-adapter -c ai-adapter.example.yaml

Codex CLI Setup

Configure Codex to use ai-adapter by editing ~/.codex/config.toml:

[model_providers.adapter]
name = "Proxy"
base_url = "http://127.0.0.1:9090/v1"
wire_api = "responses"

Create a profile to use a specific model (e.g. ~/.codex/deepseek.config.toml):

model = "deepseek-v4-pro"
model_provider = "adapter"

Then launch Codex:

codex --profile deepseek --dangerously-bypass-approvals-and-sandbox -s danger-full-access

Profile Examples

Each model gets its own profile file under ~/.codex/:

DeepSeek (~/.codex/deepseek-flash.config.toml):

model = "deepseek-v4-flash"
model_provider = "adapter"

Xiaomi MiMo (~/.codex/mimo.config.toml):

model = "mimo-v2.5-pro"
model_provider = "adapter"

MiniMax (~/.codex/minimax.config.toml):

model = "MiniMax-M3"
model_provider = "adapter"

Launch with:

codex --profile deepseek-flash --dangerously-bypass-approvals-and-sandbox -s danger-full-access
# or
codex --profile mimo ...
# or
codex --profile minimax ...

Configuration

YAML Config File

Note: format is optional and auto-detected from baseUrl when omitted. Set to anthropic, openai-chat, openai-responses, or "auto".

server:
  addr: "0.0.0.0:9090"
  accessLogDir: "~/.ai-adapter/logs"
  reqlog: true
  dropImages: true

providers:
  - name: deepseek
    baseUrl: "https://api.deepseek.com/anthropic"
    format: anthropic   # optional — auto-detected from base_url
    apikey: "${DEEPSEEK_API_KEY}"
    vendor: deepseek
    dropImages: true
    models:
      - deepseek-v4-pro
      - deepseek-v4-flash

modelAliases:
  "gpt-5.5": "deepseek/deepseek-v4-pro"
  "gpt-5.4": "deepseek/deepseek-v4-pro"
  "gpt-5.3-codex": "deepseek/deepseek-v4-pro"
  "*": "deepseek/deepseek-v4-pro"

CLI Options

Flag	Env	Default	Description
-c, --config	—	—	Config file (YAML/JSON)
--base-url	UPSTREAM_BASE_URL	—	Upstream API base URL (single provider mode)
--upstream-format	UPSTREAM_FORMAT	auto-detect from URL	anthropic, openai-chat, openai-responses
--vendor	—	auto	deepseek, openai, anthropic, xiaomimimo, auto
--apikey	UPSTREAM_API_KEY	—	Upstream API key
--prefer-client-key	—	false	Prefer Authorization header over config key
--model	UPSTREAM_MODEL	—	Default model override
-a, --addr	ADDR	0.0.0.0:9090	Server listen address
--model-alias	—	—	Model alias (repeatable, e.g. gpt-5=provider/model)
--reqlog	—	false	Log requests to SQLite
--access-log	—	false	Enable HTTP access log (JSON lines)
--drop-images	—	false	Strip images from requests

See --help for the full list.

API Endpoints

Endpoint	Method	Description
/v1/responses	POST	Responses → upstream Chat/Anthropic
/v1/chat/completions	POST	Chat → upstream
/v1/responses/compact	POST	Context compaction for Codex CLI
/v1/models	GET	Pass-through
/v1/*	*	Catch-all pass-through
/health	GET	Health check (uptime, addr, vendor)
/__/models	GET	List configured model aliases
/__/session	GET	List active sessions (JSON)

Vendor Behavior

Vendor	Protocol	thinking	developer role	tool format	Special handling
DeepSeek	Chat	disabled	→ system	flat	reasoning cache, reasoning_tokens=0
DeepSeek	Anthropic	disabled	→ assistant	nested	tool_use/tool_result merge
OpenAI	Chat	N/A	preserved	nested	standard
Anthropic	Anthropic	N/A	→ assistant	nested	standard
XiaomiMimo	Chat	none	→ system	flat	api-key auth, token-plan support
XiaomiMimo	Anthropic	none	→ assistant	nested	shared sanitization with DeepSeek

Subcommands

# Show version info
./target/debug/ai-adapter version

# Print default config template (yaml/json)
./target/debug/ai-adapter config
./target/debug/ai-adapter config --format json

# List active sessions (requires running server)
./target/debug/ai-adapter session ls

Docker

docker run -d -p 9090:9090 \
  -v ai-adapter-data:/data \
  -e ADDR=0.0.0.0:9090 \
  dyrnq/ai-adapter

Volume / Env	Description
/data (volume)	Persistent data: state DB and logs
DATA_DIR	Data directory (default /data)
ADDR	Listen address (default 0.0.0.0:9090)
AI_ADAPTER_DB	SQLite database path

License

MIT