Any model

An agent's model is a single string. Change the string, change the provider, and the rest of your code is unchanged.

from jamjet import Agent

Agent("bot", model="anthropic/claude-sonnet-4-6", tools=[...])
Agent("bot", model="openai/gpt-4o", tools=[...])
Agent("bot", model="gemini/gemini-2.0-flash", tools=[...])
Agent("bot", model="ollama/llama3.2", tools=[...])

Every one of those calls goes through the same governed seam, so the model is swappable without touching governance, tools, or the agent loop.

The model reference format

A model reference is "provider/model". The provider prefix selects the route; the rest is passed through to the provider.

"anthropic/claude-sonnet-4-6"   # provider: anthropic
"openai/gpt-4o"                 # provider: openai
"gemini/gemini-2.0-flash"       # provider: gemini
"ollama/llama3.2"               # provider: ollama

JamJet routes the prefixes anthropic, openai, gemini, and ollama through the seam. Use gemini/ for Google Gemini models. Any other prefix is treated as an OpenAI-compatible endpoint.

Provider keys

The model seam is backed by LiteLLM, which reads each provider's key from the environment. Install the model extra and set the key for whichever providers you use:

pip install 'jamjet[model]'

export ANTHROPIC_API_KEY=sk-ant-...   # anthropic/*
export OPENAI_API_KEY=sk-...          # openai/*
export GEMINI_API_KEY=...             # gemini/*

Local models through Ollama need no API key. The ollama/ prefix targets a local Ollama server at http://localhost:11434.

The governed seam

Application code never calls a provider directly. Every model call enters the Model seam, which runs a middleware chain around the provider call: the model allowlist denies disallowed models, PII redaction rewrites outbound prompts, the budget check fails closed when spend is exhausted, and metering records token and cost usage. The provider is only reached after the before chain passes.

Because this seam is the single path for every call, swapping providers never bypasses governance. An agent capped at budget=0.50 keeps that ceiling whether it runs on Claude, GPT, or a local model. See Governance for the knobs that configure the chain.

Durable model calls go through the sidecar

On the in-process path (agent.run()), the seam runs in your own process. On the durable path (agent.run_durable() and deployed agents), the model calls are made by the Rust engine, so to keep the same governed seam the engine routes them to a model sidecar: a local Python process that wraps jamjet.model.Model.

pip install 'jamjet[sidecar]'
uvicorn jamjet.model.sidecar_server:app --host 127.0.0.1 --port 4280

Then point the engine at it:

export JAMJET_MODEL_SEAM_URL=http://127.0.0.1:4280

The engine probes the sidecar's /health at startup and refuses to start if it is unreachable, so durable model calls never silently bypass the governed seam. If JAMJET_MODEL_SEAM_URL is unset, the engine falls back to native Rust adapters, which is the default for development.

Next steps