Skip to content

SimulationSample scores

Live routing demonstration: Drag the sphere: one star per provider. Pick a name to make it fail.

Pusula 1.0.0, open source under MIT

One needle picks the route.

Keep Claude Code, Codex, Cursor and your other tools exactly as they are. Point their base URL at Pusula and every request is routed across 358 providers (156 with a free tier or no key), with automatic fallback when one is down or rate-limited.

Open dashboard

Runs on your own machine, no account needed

Drag the sphere: one star per provider. Pick a name to make it fail.

X-Pusula-Provider: cc # claude

Drag the sphere: one star per provider. Pick a name to make it fail.

  • 358providers behind one endpoint
  • 156with a free tier or no key
  • 20routing strategies
  • 16auto scoring factors
  • 36tools you already use
  • 110MCP server tools

045° · Network

Pusula reads the map before it moves: each request is scored, sent down the best route, and falls back on its own when a road closes. Here is that motion, drawn on one map.

SimulationA simulation for illustration, not live traffic.
routed requests
0
rerouted
0
  1. Kabul · hub
  • hub
  • provider
  • Incoming request

Sample: each city stands for one provider: Mazar-i-Sharifclaude · Heratopenai · Kandahargemini · Kunduzglm · Jalalabadcerebras · Bamyanollama-local

Pick a city to cut its road. Pusula retries the next provider.

090° · Catalog

From the big labs to free tiers, across 5 kinds of service. 156 have a free tier or need no key; 4 are local CLI bridges.

358 of 358

free tier or no keylocal CLI bridge

SimulationAgentic coding, two tool callsClaude Code

  • 1min.AI
  • 360 AI
  • 9router
  • Adapta.org (Adapta One Web)
  • Adobe Firefly (Image/Video)
  • AgentRouter
  • Agnes AI
  • Agnes AI (China)
  • AI Horde
  • AI/ML API
  • AI21 Labs
  • AINative Studio
  • Aion Labs
  • Alibaba (China)
  • Alibaba Cloud Model Studio
  • Alibaba Token Plan
  • Amazon Bedrock
  • Amazon Q
  • Ant Ling / Ring (inclusionAI)
  • Anthropic
  • Antigravity
  • Antigravity CLI
  • AnyAPI AI
  • AnySearch
  • Api.airforce
  • Arcee AI
  • Arena (Free)
  • AssemblyAI
  • Augment (Auggie CLI)
  • Auriko
  • Auto (Zero-Config)
  • AWS Polly
  • Azure AI Foundry
  • Azure OpenAI
  • b.ai
  • Baichuan
  • Baidu (ERNIE)
  • Baidu Qianfan
  • Baseten
  • BazaarLink
  • Black Forest Labs
  • Blackbox AI
  • Blackbox Web (Subscription)
  • BluesMinds
  • Brave Search
  • BytePlus ModelArk
  • Bytez
  • Cartesia
  • Cerebras
  • Charm Hyper
  • Chat Oripe
  • ChatAnywhere
  • ChatGPT Web (Clean Room)
  • ChatGPT Web (Codex)
  • Cheaper Inference
  • Chenzk API
  • Chutes.ai
  • Clarifai
  • Claude Code
  • Claude Web
  • Cline
  • ClinePass
  • CLIProxyAPI
  • CloudCode.ONE
  • Cloudflare AI Playground
  • Cloudflare Workers AI
  • CodeBuddy CN
  • Codestral
  • Codex Cloud
  • Cohere
  • ComfyUI
  • Command Code
  • Conol (Unofficial/Experimental)
  • Context7 (library docs)
  • Coze
  • CrofAI
  • Cursor API
  • Cursor IDE
  • Dahl
  • Databricks
  • DataRobot
  • DeepAI
  • Deepgram
  • DeepInfra
  • DeepSeek
  • DeepSeek Web
  • Devin
  • Devin CLI
  • Devin CLI Agentic Bridge
  • Devin Desktop
  • DGrid
  • Dify
  • DigitalOcean
  • DIT.ai
  • Docker Model Runner
  • Dola Web (ByteDance)
  • Doubao
  • DuckDuckGo AI Chat
  • DXNT / DX Token
  • Electron Hub
  • ElevenLabs
  • Empower
  • EURouter
  • Exa Search
  • Factory
  • Fal.ai
  • FastRouter
  • Featherless AI
  • FenayAI
  • Firecrawl
  • Fireworks AI
  • Fish Audio
  • Free.ai
  • FreeAIAPIKey
  • Freebuff
  • FreeInference
  • FreeModel.dev
  • FreeTheAi
  • FriendliAI
  • g4f.space — Gemini
  • g4f.space — Groq
  • g4f.space — NVIDIA
  • g4f.space — Ollama
  • g4f.space — Pollinations
  • Galadriel
  • Gemini (Google AI Studio)
  • Gemini Web (Free)
  • GigaChat (Sber)
  • GitHub Copilot
  • GitHub Enterprise Copilot
  • GitLab Duo
  • GitLab Duo PAT
  • Gitlawb Opengateway (GMI Cloud)
  • Gitlawb Opengateway (MiMo)
  • Gladia
  • GLM Coding
  • GLM Coding (China)
  • GLM Thinking
  • GoAPI
  • Google Jules
  • Google Programmable Search
  • GreenPT
  • Grok Build
  • Grok Web (Subscription)
  • Groq
  • Haiper
  • HelixMind
  • Helyx AI
  • Heroku AI
  • Huancheng Public API
  • HuggingChat (Free)
  • HuggingFace
  • HyperAgent (Unofficial/Experimental)
  • Hyperbolic
  • IBM watsonx.ai Gateway
  • Ideogram
  • iFlytek Spark
  • Inception
  • Inference.net
  • Inner.ai (Subscription)
  • InternLM (Intern-S1)
  • Inworld
  • Jina AI (Foundation API)
  • Jina Reader (r.jina.ai)
  • Kenari
  • KIE.AI
  • Kilo Code
  • Kilo Gateway
  • Kimi
  • Kimi (Legacy Moonshot API)
  • Kimi Code API Key
  • Kimi Code CLI
  • Kimi Web
  • Kiro AI
  • Lambda AI
  • LaoZhang AI
  • Lemonade Server
  • Leonardo AI
  • Linkup Search
  • Liquid AI
  • LiteRouter
  • llama.cpp
  • Llamafile
  • LlamaGate
  • LLM Gateway
  • LLM.Kiwi
  • LLM7.io
  • LM Studio
  • Logfare
  • LongCat AI
  • Lyceum
  • Magnific
  • Maritalk
  • MaxAI
  • MegaNova AI
  • Meta Llama API
  • Microsoft 365 Copilot (BizChat)
  • Microsoft Copilot Web
  • Minimax (China)
  • Minimax Coding
  • Mistral
  • Mixedbread AI
  • Mixlayer
  • MLX Gemma 26B
  • MLX Qwen 3.8 27B
  • MNN AI
  • Modal
  • ModelScope
  • MonsterAPI
  • Morph
  • Muse Code (Meta)
  • Muse Spark Web (Meta AI)
  • Naga AI
  • Naga.ac
  • NanoGPT
  • NaraRouter
  • Naver CLOVA Studio
  • NavyAI
  • Nebius AI
  • Nimble Search
  • NLP Cloud
  • Nomic
  • Notion AI Web (Unofficial/Experimental)
  • Nous Research
  • Novita AI
  • nScale
  • Nube.sh
  • NVIDIA NIM
  • NVIDIA Triton
  • OCI Generative AI
  • OfoxAI
  • Ollama
  • Ollama Cloud
  • Ollama Search
  • oobabooga
  • OpenAdapter
  • OpenAI
  • OpenAI Codex
  • OpenAI Codex (App-Server)
  • OpenCode Free
  • OpenCode Go
  • OpenCode Zen
  • Openference
  • Openference API
  • OpenRouter
  • OpenVecta
  • Opper
  • OrcaRouter
  • OVHcloud AI
  • Perplexity
  • Perplexity Agent
  • Perplexity Search
  • Perplexity Web (Pro/Max)
  • PiAPI
  • Pioneer AI
  • PLaMo
  • PlayHT
  • Poe
  • Poe Web (Subscription)
  • Poixe AI
  • Pollinations AI
  • Poolside
  • Predibase
  • PromptQL (Unofficial/Experimental)
  • PublicAI
  • Qiniu
  • Qoder
  • Qwen Cloud
  • Qwen Cloud Token Plan
  • Recraft
  • Regolo AI
  • Reka
  • Requesty
  • Rev AI
  • Routeway
  • Runway
  • SambaNova
  • SAP Generative AI Hub
  • Sarvam AI
  • Scaleway AI
  • SD WebUI
  • SEA-LION
  • SearchAPI
  • SearXNG Search
  • SeekAi
  • Segmind
  • SenseNova
  • Serper Search
  • SiliconFlow
  • Snowflake Cortex
  • Soniox
  • SparkDesk
  • Speechmatics
  • Speka AI
  • Stability AI
  • StepFun
  • SumoPod
  • Synthetic
  • t3.chat (Pro/Free)
  • TabiToken
  • Tavily Search
  • Tencent AI Studio (Free)
  • Tencent Hunyuan
  • Tencent Yuanbao (Free)
  • TheB.AI
  • TinyCMS Web (Free/Sub)
  • TinyFish Fetch
  • Together AI
  • Token Kiosk
  • TokenReply
  • TokenRouter
  • Topaz
  • Trae
  • Typhoon
  • UC (uncensored.com)
  • UC Direct (uncensored.com)
  • Udio
  • UncloseAI
  • UnoRouter
  • Upstage
  • v0 (Vercel)
  • v0 Vercel Web (Code Gen)
  • Venice Web (Privacy)
  • Venice.ai
  • Veo AI Free
  • Vercel AI Gateway
  • Vertex AI
  • Vertex AI Partners
  • vLLM
  • Void AI
  • Volcengine
  • Volcengine Ark Agent Plan
  • Volcengine Ark Coding Plan
  • Voyage AI
  • Wafer AI
  • Weights & Biases Inference
  • Writer
  • X Search (Grok)
  • X5Lab
  • xAI (Grok)
  • xAI OAuth (Grok)
  • Xiaomi MiMo
  • Xiaomi MiMo Token Plan
  • XInference
  • xKiro
  • Xquik X Search
  • Yi (01.AI)
  • Yolo-Auto
  • You.com Search
  • Z.AI
  • Z.ai Web
  • ZCode (GLM Coding Plan)
  • Zed Hosted Models
  • Zed IDE
  • ZenMux
  • ZenMux Free (Web)
  • ZeroLimitAI
  • Zylo API

135° · Routing

A combo is a list of models plus a strategy. Set the model to auto and Pusula scores every candidate for the request in front of it; pick another strategy and you decide the order yourself.

How auto scores a candidate, 16 factors

Default weights from the scoring engine, as shipped. Every factor is measured per request from live data; the weights decide how much each one votes.

score = Σ wᵢ·xᵢ

Incoming request · Agentic coding, two tool calls · 3.1k tokens

  1. claude0.764
  2. openai0.737
  3. glm0.722
  4. cerebras0.691
  5. gemini0.687
  6. ollama-local0.653

Σ w = 1.000

Drag a point to retune. The whole page then routes with your weights.

Weights sum to 1.0 and can be retuned per combo. Factors with no default weight still gate decisions such as cache reuse.

Slide a factor to retune. The whole page then routes with your weights.

Σ w = 1.000

Weights sum to 1.0 and can be retuned per combo. Factors with no default weight still gate decisions such as cache reuse.

SimulationReal default weights. Candidate measurements are a simulation.

20 strategies

Each one answers a different question: cheapest, fastest, most quota left, most context, or a panel that argues it out.

priority
weighted
round-robin
context-relay
fill-first
p2c
random
least-used
cost-optimized
reset-aware
reset-window
headroom
quota-weighted
strict-random
auto
lkgp
context-optimized
cache-optimized
fusion
pipeline
priority

First healthy target in your order

SimulationPick a target to fail it

First healthy target in your order
Spread traffic by the weights you set
Take turns, one request each
Hand a long session to the next account with a summary
Use one target until it is exhausted, then the next
Pick two at random, keep the healthier one
Any healthy target, then fall back to the rest
The target with the fewest requests so far
Cheapest per token first
Balance quota against 5-hour and weekly resets
Spend quota before its window closes
Most remaining quota first
In proportion to leftover quota, divided by in-flight load
Shuffle the deck, use each target once
Score every candidate on 16 factors, self-healing
The last provider that answered well
Fit the prompt to the largest context window
Keep a reusable prompt prefix on the account that cached it
Ask a panel in parallel, let a judge merge
Chain models, each step feeding the next

180° · Resilience

Failures are isolated at the smallest scope that explains them. A rate-limited model locks only that model, a bad key cools only that connection, and the whole provider is bypassed only after repeated upstream errors.

Simulation
provider408 · 500 · 502 · 503 · 504
Circuit breaker for the whole provider. Opens only on provider-level failures (408, 500, 502, 503, 504); combo routing skips it until a half-open probe succeeds.
connection
Cooldown for one key or account. Exponential backoff, upstream Retry-After hints respected, sibling keys for the same provider keep serving.
model
Lockout for one model on one connection. A 429 on a per-model quota, or a 404 for a model your local server does not have, leaves the other models on that connection serving.

Profile

Breaker defaults by connection profile

ProfileDegradesOpensProbes againKey cooldown base
after 5 failuresafter 8 failuresafter 60s5s
after 7 failuresafter 12 failuresafter 30s3s
—after 2 failuresafter 15s2s

Every default can be overridden with an environment variable.

  • modelmodel 2
  • connection#2
  • providerclaude · closed

provider · claude

  • closed
  • degraded
  • open
  • half-open

0 / 8

after 8 failures

—

Probes again

provider408 · 500 · 502 · 503 · 504
Circuit breaker for the whole provider. Opens only on provider-level failures (408, 500, 502, 503, 504); combo routing skips it until a half-open probe succeeds.
connection
Cooldown for one key or account. Exponential backoff, upstream Retry-After hints respected, sibling keys for the same provider keep serving.
model
Lockout for one model on one connection. A 429 on a per-model quota, or a 404 for a model your local server does not have, leaves the other models on that connection serving.

225° · Tools

36 CLIs, editors and agents are configured out of the box. Change one base URL, keep your workflow.

Supported tools36

Base URL for your client

http://localhost:20128/v1

autoclaude

Simulation
pusula /v1
# Base URL for your client
curl http://localhost:20128/v1/chat/completions \
  -H "Authorization: Bearer $PUSULA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Hello"}]}'

200Hello! How can I help you today?

X-Pusula-Provider: cc

X-Pusula-Latency-Ms: 475

X-Pusula-Request-Id: d4bef405-044e-4b01-a499-619ace2f7bf2

Anything that speaks the OpenAI, Anthropic or Gemini API works too: the same /v1 endpoint accepts all three formats and translates between them.

270° · Compression and MCP

Prompts are compressed before they leave, answers are cached when they repeat. Each response header tells you what happened.

48 tokens

Semantic cache

lossless lossy

POST /v1/chat/completions
HTTP/1.1 200 OK
X-Pusula-Cache: MISS
X-Pusula-Compression: stacked; source=active-profile; tokens=48->19; rules: …

lossless lossy

SimulationIllustration with example header values. The engine stack is real; the reduction shown is not a measurement.
12 compression engines
Stack them per combo or per request. Measured savings range from 15% to 95% of prompt tokens depending on the workload.
Semantic cache
A repeated or near-identical request is answered from the cache without an upstream call.
Savings you can audit
Tokens and cost saved are reported per response and summed in the dashboard.

The same gateway speaks MCP to your editor and A2A to other agents, so routing, quota and health become something an agent can ask about.

stdionpx pusula --mcp

SSEGET /api/mcp/sse

Streamable HTTPPOST /api/mcp/stream

tools
110
scopes
33
transports
3

MCP server

110 tools across 33 scopes over 3 transports (stdio, SSE, streamable HTTP). Plug it into Claude Code, Cursor or any MCP client.

SimulationOne light per MCP tool. Calls are simulated; positions are not a map of the real tools.

A2A agentJSON-RPC 2.0

A JSON-RPC 2.0 agent card other agents can discover and call, with skills for routing, quota, discovery, cost and health.

POST /a2a/.well-known/agent.json

POST /a2a
→ message/stream · skill: cost-analysis
← completed
→ message/send · skill: health-report
← completed
→ tasks/get
← completed
Simulation

Protocol reference in the docs

315° · Dashboard

Providers, combos, keys, usage, costs and health in one local page. Everything stays on the machine that runs Pusula.

localhost:20128/dashboardscreenshot
Pusula dashboard showing connected providers, request volume and costs
Credentials are stored with AES-256-GCM; the key is generated on first run.
Keys, usage and history live in a database on disk, in your data directory.
API and dashboard share port 20128. Put it behind your own domain or keep it on localhost.
Nothing to sign up for. Pusula talks to the providers you connect and to no one else.
Open dashboard

No build, no config file. The first run creates the data directory and the encryption key for you.

  1. Run it

    Needs Node 22 or newer. The command starts the API and the dashboard on one port.

  2. Connect providers

    Sign in with OAuth, paste an API key, or pick a free tier. Create an API key for each of your clients.

  3. Point your tools at it

    Set the base URL to http://localhost:20128/v1 and use the model name auto, or the name of a combo you built.

npx pusula
terminal
● Pusula is running
  API        http://localhost:20128/v1
  Dashboard  http://localhost:20128/dashboard

Set the course.

SimulationSimulated traffic, not live data