ferryllm

Universal LLM protocol middleware for OpenAI, Anthropic, Claude Code, and OpenAI-compatible backends.

ferryllm is a Rust gateway that normalizes client and provider LLM protocols into one internal representation. Use it as a Claude Code bridge, a private model gateway, or an embeddable adapter layer.

What It Does

• Accepts OpenAI-compatible chat requests at POST /v1/chat/completions
• Accepts Anthropic-compatible messages at POST /v1/messages
• Rewrites model names with exact and prefix routing rules
• Forwards to OpenAI-compatible or Anthropic backend adapters
• Preserves tool calls and SSE streaming behavior
• Keeps prompt-cache keys stable while stripping transport metadata
• Maps reasoning control through the IR and provider adapters

Why ferryllm

Most gateways end up as an N x M matrix: every client protocol needs custom code for every provider protocol.

ferryllm uses N + M routing instead:

Client protocol -> ferryllm IR -> provider protocol

That makes it easier to:

• put Claude Code behind a stable backend
• expose one gateway to multiple client protocols
• keep cache behavior predictable
• add new providers without rewriting every client path

Fast Start

Install from crates.io:

cargo install ferryllm

Run from source:

git clone https://github.com/caomengxuan666/ferryllm.git
cd ferryllm
cargo run --features http --bin ferryllm -- serve --config examples/config/codexapis.toml

Set the provider key and start the server:

export CODX_API_KEY="your-api-key"
RUST_LOG=info ferryllm serve --config examples/config/codexapis.toml

Smoke test the Anthropic-compatible endpoint:

curl -s http://127.0.0.1:3000/v1/messages \
  -H 'content-type: application/json' \
  -H 'authorization: Bearer local-test-token' \
  -d '{"model":"cc-gpt55","max_tokens":64,"messages":[{"role":"user","content":"hello"}]}'

Claude Code Bridge

Claude Code sends Anthropic-format requests. ferryllm can receive those requests, rewrite the model, and forward them to an OpenAI-compatible backend.

Claude Code
  -> POST /v1/messages, model = claude-*
  -> ferryllm Anthropic entry
  -> unified IR
  -> route match: claude-
  -> rewrite backend model: gpt-5.4
  -> OpenAI-compatible backend

Start ferryllm:

export CODX_API_KEY="your-api-key"
RUST_LOG=ferryllm=info,tower_http=info \
  ferryllm serve --config examples/config/codexapis.toml

Point Claude Code at ferryllm:

ANTHROPIC_API_KEY=dummy \
ANTHROPIC_BASE_URL=http://127.0.0.1:3000 \
claude --bare --print --model claude-opus-4-6 \
  "Reply with exactly one short word: pong"

Expected output:

pong

See docs/claude-code.md for persistent Claude Code and cc-switch setup.

Configuration

ferryllm uses TOML config. Secrets stay in environment variables.

[server]
listen = "0.0.0.0:3000"
request_timeout_secs = 120
body_limit_mb = 32
default_reasoning_effort = "medium"
# Optional. Uncomment to cap in-flight requests.
# max_concurrent_requests = 128
# Optional. Uncomment to cap total requests per minute.
# rate_limit_per_minute = 600
# Optional non-streaming upstream resilience. Streaming requests are not retried.
# retry_attempts = 2
# retry_backoff_ms = 100
# circuit_breaker_failures = 5
# circuit_breaker_cooldown_secs = 30

[logging]
level = "info"
format = "text"

[auth]
enabled = false
# api_keys_env = "FERRYLLM_API_KEYS"
# Optional per-client caps, keyed by the authenticated API key.
# per_key_rate_limit_per_minute = 120
# per_key_max_concurrent_requests = 8

[metrics]
enabled = true

[prompt_cache]
auto_inject_anthropic_cache_control = true
cache_system = true
cache_tools = true
cache_last_user_message = true
openai_prompt_cache_key = "ferryllm"
# openai_prompt_cache_retention = "24h"
debug_log_request_shape = true
relocate_system_prefix_range = "0..1"
log_relocated_system_text = false
strip_system_line_prefixes = ["x-anthropic-billing-header:"]

[[providers]]
name = "codexapis"
# Default path for controlling reasoning effort today.
type = "openai_responses"
base_url = "https://codexapis.com"
api_key_env = "CODX_API_KEY"
# If you want the legacy Chat Completions path instead, switch this back to:
# type = "openai"

[[routes]]
match = "cc-gpt55"
match_type = "exact"
provider = "codexapis"
rewrite_model = "gpt-5.4"

[[routes]]
match = "claude-"
provider = "codexapis"
rewrite_model = "gpt-5.4"

[[routes]]
match = "gpt-"
provider = "codexapis"

[[routes]]
match = "grok-"
provider = "codexapis"

[[routes]]
match = "*"
provider = "codexapis"
rewrite_model = "gpt-5.4"

Check a config without starting the server:

ferryllm check-config --config examples/config/codexapis.toml

To route OpenAI-compatible upstream calls through the Responses API instead of Chat Completions, build with the optional feature and use provider type openai_responses:

cargo build --release --features http,prompt-observability,openai-responses --bin ferryllm

[[providers]]
name = "codexapis"
type = "openai_responses"
base_url = "https://codexapis.com"
api_key_env = "CODX_API_KEY"

See examples/config/codexapis-responses.toml.

Reasoning Effort

Set the default model reasoning depth in TOML:

[server]
default_reasoning_effort = "medium"

Valid values are none, low, medium, high, xhigh, and x_high.

This default is applied only when the client request does not already include an explicit reasoning or thinking control. For Claude Code today, changing this in TOML is the practical way to control the forwarded OpenAI-compatible reasoning.effort.

Run with debug logging and look for reasoning=effort=... in the outbound request-shape log to confirm what ferryllm sent upstream.

API Surface

Prompt Cache

ferryllm keeps prompt-cache keys stable while stripping transport metadata and normalizing the prompt prefix.

With prompt-observability enabled, ferryllm logs prompt-cache usage and exposes it through /metrics.

For Claude Code deployments, the important knobs are:

• relocate_system_prefix_range
• strip_system_line_prefixes
• openai_prompt_cache_key
• default_reasoning_effort

See docs/prompt-caching.md and docs/reasoning-control.md.

Architecture

src/
  adapter.rs        Adapter trait
  ir.rs             Unified request, response, content, tool, and stream types
  router.rs         Exact and prefix model routing
  server.rs         Axum HTTP server
  config.rs         TOML config loader and validator
  entry/            Client protocol translators
  adapters/         Backend provider adapters

More detail: docs/architecture.md.

Load Testing

ferryllm ships a benchmark-style load tester for local mock-upstream testing:

MOCK_DELAY_MS=20 cargo run --example mock_openai_upstream --features http
cargo run --release --example load_test --features http -- \
  --preset mock-anthropic \
  --requests 10000 \
  --concurrency 512

See docs/load-testing.md.

Documentation

• Chinese README
• Architecture
• Claude Code setup
• Configuration
• Compatibility notes
• Deployment
• Load testing
• Prompt caching and token observability
• Reasoning control

Roadmap

• More provider adapters, including Gemini
• Weighted and latency-aware provider pools
• Hot-reload configuration
• Richer Prometheus metrics labels
• Per-key quota and usage accounting hooks
• Packaged Docker images and deployment templates

License

MIT. See LICENSE.