kitty-bridge 0.1.7

Kitty Bridge — launch coding agents through a local API bridge

Kitty Bridge

A thin launcher for coding agents that routes requests through a local API bridge to any upstream Chat Completions provider.

What it does

kitty sits between your coding agent and the upstream API:

Agent (codex / claude / gemini / kilo) → kitty bridge → Chat Completions → upstream provider

The bridge translates between the agent's wire protocol (Responses API, Messages API, or Gemini API) and a standard Chat Completions API, so you can use any compatible provider without the agent needing native support.

Install

pip install -e .

Requires Python 3.10+.

Quick start

# First-time setup — interactive wizard creates a profile
kitty setup

# Launch Codex CLI through the bridge
kitty codex

# Launch Claude Code through the bridge
kitty claude

# Launch Gemini CLI through the bridge
kitty gemini

# Launch Kilo Code through the bridge
kitty kilo

# Use a specific profile
kitty my-profile codex

# Start a standalone OpenAI-compatible API server (bridge mode)
kitty bridge
kitty my-profile bridge

Commands

Command	Description
kitty setup	Interactive wizard to create your first profile
kitty profile	Manage profiles (create, delete, set default, list)
kitty doctor	Diagnose installation issues
kitty cleanup	Remove stale bridge values from agent config files
kitty codex	Launch Codex CLI through the bridge
kitty claude	Launch Claude Code through the bridge
kitty gemini	Launch Gemini CLI through the bridge
kitty kilo	Launch Kilo Code through the bridge
kitty bridge	Start a standalone OpenAI-compatible API server
kitty <profile>	Launch default agent with a named profile
kitty <profile> codex	Launch Codex with a named profile
kitty <profile> claude	Launch Claude Code with a named profile
kitty <profile> gemini	Launch Gemini CLI with a named profile
kitty <profile> kilo	Launch Kilo Code with a named profile
kitty <profile> bridge	Start bridge server with a named profile
kitty --version	Print version
kitty --help	Print help

Supported agents

• Claude Code — Anthropic's coding agent (Messages API)
• Codex CLI — OpenAI's coding agent (Responses API)
• Gemini CLI — Google's coding agent (Gemini API)
• Kilo Code — Open-source coding agent

Supported providers

Provider	Type ID	Notes
Z.AI (regular)	zai_regular	Standard Z.AI API
Z.AI (coding plan)	zai_coding	Z.AI Coding Plan endpoint
MiniMax	minimax
Novita	novita
Ollama	ollama	Local LLM deployment (OpenAI-compatible)
OpenAI	openai
OpenRouter	openrouter	Multi-provider router
Fireworks	fireworks	Fire Pass endpoint
Anthropic	anthropic	Direct Anthropic Messages API
AWS Bedrock	bedrock	Uses boto3 SigV4 auth
Azure OpenAI	azure	Azure-specific endpoint format
Google Vertex AI	vertex	Requires project/location config

How it works

Agent launch mode

1. kitty resolves your profile (provider, model, API key)
2. Starts a local HTTP bridge on a random port
3. Configures the agent to talk to the bridge
4. The bridge translates requests to Chat Completions format and forwards them to your provider
5. Responses are translated back to the agent's native format
6. When the agent exits, kitty restores the agent's config files to their original state

Bridge mode

Start a standalone OpenAI-compatible API server without launching an agent:

# Use default profile
kitty bridge

# Use a specific profile
kitty my-profile bridge

The server exposes:

• POST /v1/chat/completions — Chat Completions API (streaming and non-streaming)
• POST /v1/messages — Anthropic Messages API (for Claude Code)
• POST /v1/responses — OpenAI Responses API (for Codex)
• POST /v1/gemini/generateContent — Gemini API (for Gemini CLI)
• GET /healthz — Health check

Balancing profiles

Balancing profiles distribute LLM calls across multiple providers using random selection across healthy members:

# Create a balancing profile via the profile menu
kitty profile
# → "Create balancing profile" → select 2+ member profiles

# Use a balancing profile
kitty my-balancer codex
kitty my-balancer bridge

Each request is routed to a randomly selected healthy provider member. Useful for cost optimization, rate limit distribution, and resilience across multiple providers.

Profiles

A profile binds a provider, model, and API key together. Profiles are stored in ~/.config/kitty/profiles.json.

# Create profiles interactively
kitty setup
kitty profile

# Use a specific profile
kitty my-profile claude

# Set a default profile (used when no profile is named)
kitty profile  # → "Set default profile"

A balancing profile is a list of regular profiles. Requests are distributed by random healthy-member selection.

Pre-flight validation

Before launching, kitty validates your API key by making a lightweight test request to the upstream provider. If the key is expired or invalid, kitty reports the error immediately and exits.

# Skip validation (e.g. in air-gapped environments)
kitty --no-validate my-profile claude

Cleanup

kitty uses a three-layer cleanup strategy to restore agent config files:

1. finally block — normal path after agent exits
2. atexit handler — runs on sys.exit(), unhandled exceptions, and SIGTERM
3. kitty cleanup — manual fallback for SIGKILL or kernel OOM

# Manually restore stale agent config
kitty cleanup

Architecture

src/kitty/
├── bridge/          # HTTP bridge server + protocol translation
│   ├── server.py    # aiohttp-based bridge with random healthy-backend balancing
│   ├── engine.py    # Shared translation primitives
│   ├── responses/   # Responses API translation (Codex)
│   ├── messages/    # Messages API translation (Claude Code)
│   └── gemini/      # Gemini API translation
├── cli/             # Command-line interface
│   ├── main.py      # Entry point
│   ├── router.py    # Argument routing
│   ├── launcher.py  # Bridge + child process orchestration
│   ├── doctor_cmd.py
│   ├── setup_cmd.py
│   └── profile_cmd.py
├── credentials/     # API key storage (file backend)
├── launchers/       # Agent-specific adapters
│   ├── claude.py    # Claude Code adapter (env vars + settings.json)
│   ├── codex.py     # Codex CLI adapter
│   ├── gemini.py    # Gemini CLI adapter
│   ├── kilo.py      # Kilo Code adapter
│   └── discovery.py # Binary discovery (PATH + fallbacks)
├── profiles/        # Profile management
│   ├── schema.py    # Profile, BalancingProfile, BackendConfig
│   ├── store.py     # JSON persistence with type discriminator
│   └── resolver.py  # Profile lookup and default resolution
├── providers/       # Upstream provider adapters
├── tui/             # Terminal UI components
└── types.py         # Shared types

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check .

# Type check
mypy src/kitty

FAQ

"API Error: Unable to connect to API (ConnectionRefused)"

The coding agent is trying to connect to a bridge server that isn't running. Two common causes:

Stale config from a crashed session. If kitty was killed with SIGKILL, the agent's config file may still contain stale values. Fix:

kitty cleanup

Running the agent directly without kitty. Always launch through kitty, or run kitty cleanup first.

"API Error: 401" or "token expired or incorrect"

Your API key has expired or been revoked:

kitty setup

"Prompt exceeds max length" (Z.AI error 1261)

The conversation context has grown beyond the model's context window. Use /clear in the agent to reset the conversation.

"pip's dependency resolver ... langchain/langchain-core requires langsmith<0.4, but you have langsmith 0.4.x"

kitty-bridge does not depend on langsmith. This warning means your current Python environment already has a langchain 0.3.x stack that is incompatible with the installed langsmith version.

If you are staying on langchain 0.3.x, repair the environment first:

python -m pip check
python -m pip install --upgrade "langsmith<0.4"
python -m pip check

Then install kitty again:

python -m pip install -e .

Using a fresh virtual environment per project avoids this class of cross-project dependency conflict.

License

MIT