> ## Documentation Index > Fetch the complete documentation index at: https://praison.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # HostedAgent Compute Runtimes > Run the whole agent loop in Docker, E2B, Modal, Daytona, Fly.io, or Tenki — one class, one string HostedAgent hosts the whole agent loop in any registered compute runtime — pick one with a single string. ```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} graph LR H[🌐 HostedAgent
provider=...] --> R[🔀 Backend Registry] R --> A[☁ī¸ anthropic] R --> D[đŸ“Ļ docker] R --> E[đŸ–Ĩī¸ e2b] R --> M[⚡ modal] R --> Y[🗄ī¸ daytona] R --> F[✈ī¸ flyio] R --> T[🌩ī¸ tenki] classDef input fill:#8B0000,stroke:#7C90A0,color:#fff classDef registry fill:#189AB4,stroke:#7C90A0,color:#fff classDef provider fill:#10B981,stroke:#7C90A0,color:#fff class H input class R registry class A,D,E,M,Y,F,T provider ``` ## Quick Start ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents import Agent from praisonai import HostedAgent agent = Agent(backend=HostedAgent(provider="anthropic")) agent.start("Summarise today's meeting notes.") ``` ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents import Agent from praisonai import HostedAgent, HostedAgentConfig agent = Agent(backend=HostedAgent( provider="daytona", config=HostedAgentConfig(model="gpt-4o-mini", name="RemoteDev"), )) agent.start("Set up a Python project and run pytest.") ``` *** ## Which provider should I pick? The scenario picks the string. ```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} graph TB Q{What do you need?} Q -->|Anthropic's managed cloud| AN[provider=anthropic] Q -->|My own Docker daemon| DK[provider=docker] Q -->|Persistent dev workspace| DY[provider=daytona] Q -->|Instant cloud VM| E2[provider=e2b] Q -->|Serverless burst| MO[provider=modal] Q -->|Fly Machines| FL[provider=flyio] Q -->|Tenki cloud| TK[provider=tenki] classDef question fill:#F59E0B,stroke:#7C90A0,color:#fff classDef answer fill:#10B981,stroke:#7C90A0,color:#fff class Q question class AN,DK,DY,E2,MO,FL,TK answer ``` *** ## How it resolves — the backend registry `HostedAgent(provider=...)` is a factory: it looks the provider up in `ManagedBackendRegistry` and returns the resolved backend. ```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} sequenceDiagram participant User participant HostedAgent participant Registry participant Backend User->>HostedAgent: HostedAgent(provider="daytona") HostedAgent->>Registry: is_available("daytona")? Registry-->>HostedAgent: True HostedAgent->>Backend: resolve + build Backend-->>User: backend instance ``` `anthropic` is a builtin; every registered compute runtime — `docker` included as of [PR #4077](https://github.com/MervinPraison/PraisonAI/pull/4077) — resolves through the same registry and runs on one generic backend. *** ## Provider matrix | Provider | Extra install | Env var | Persistent | Best for | | ----------- | ---------------------------------- | ------------------- | :--------: | ------------------------- | | `anthropic` | (builtin) | `ANTHROPIC_API_KEY` | ❌ | Anthropic-hosted loops | | `docker` | `pip install "praisonai[docker]"` | (local daemon) | ❌ | Local isolation | | `e2b` | `pip install "praisonai[e2b]"` | `E2B_API_KEY` | ❌ | Instant cloud VM | | `modal` | `pip install "praisonai[modal]"` | Modal auth | ❌ | Serverless burst | | `daytona` | `pip install "praisonai[daytona]"` | `DAYTONA_API_KEY` | ✅ | Persistent dev workspaces | | `flyio` | Fly.io token | `FLY_API_TOKEN` | ❌ | Fly Machines | | `tenki` | `pip install "praisonai[tenki]"` | Tenki auth | ❌ | Tenki cloud | *** ## Custom image / template Since [PR #4077](https://github.com/MervinPraison/PraisonAI/pull/4077) the generic backend honours `image=` — one string maps to each provider's native concept. | Provider | `image=` honoured as | | --------------------------- | -------------------------------------------------------------- | | `docker` | Docker image ref | | `e2b` | E2B template name (unless `metadata["e2b_template"]` is set) | | `modal` | Registry image ref via `Image.from_registry(...)` | | `daytona`, `flyio`, `tenki` | Not applicable (uses provider-native workspace/image handling) | ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents import Agent from praisonai import HostedAgent, HostedAgentConfig # Same string across providers — each maps it to the right native concept. agent = Agent(backend=HostedAgent( provider="modal", config=HostedAgentConfig(model="gpt-4o-mini"), image="ghcr.io/my-org/my-runtime:latest", )) agent.start("Do something in my custom image.") ``` The default image is `python:3.12-slim`; `image=` is honoured only when set to something other than that default. *** ## Daytona notes Two Daytona defects were fixed in [PR #4074](https://github.com/MervinPraison/PraisonAI/pull/4074). **Memory unit.** `ComputeConfig` takes `memory_mb` (megabytes). Daytona's native unit is GiB, so the provider converts automatically — `memory_gib = max(1, round((memory_mb or 1024) / 1024))`. The default `ComputeConfig()` now provisions 1 GiB instead of asking for 1024 GiB. Sub-1024 MB requests floor to 1 GiB (never 0). **File uploads.** Uploading a local file copies the file's *contents* to the remote destination (`upload_file(content, remote_path)`). Earlier versions swapped the arguments and wrote the path string into a file named after the content. *** ## What HostedAgent will NOT accept `openai`, `gemini`, `ollama`, and `local` are LLM routing names, not runtimes. HostedAgent raises `ValueError("... not yet available. Use LocalAgent ...")`. ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonai import LocalAgent, LocalAgentConfig agent = LocalAgent(config=LocalAgentConfig(model="openai/gpt-4o")) agent.start("Draft a release note.") ``` *** ## Common Patterns ### Swap the provider without touching agent code The provider is one string; the agent stays the same. ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} from praisonaiagents import Agent from praisonai import HostedAgent, HostedAgentConfig def build(provider: str) -> Agent: return Agent(backend=HostedAgent( provider=provider, config=HostedAgentConfig(model="gpt-4o-mini"), )) agent = build("daytona") agent.start("Run the test suite.") ``` ### Pick the provider per environment Read the provider from an env var to switch cloud vs local. ```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}} import os from praisonaiagents import Agent from praisonai import HostedAgent provider = os.environ.get("HOSTED_PROVIDER", "anthropic") agent = Agent(backend=HostedAgent(provider=provider)) agent.start("Summarise the changelog.") ``` *** ## Best Practices `provider="anthropic"` needs only `ANTHROPIC_API_KEY` — no extra install, no daemon. Reach for a compute runtime when you need your own isolation, persistence, or a specific cloud. Pass the provider as a plain string from config or an env var. Every registered runtime shares the same `HostedAgent(provider=..., config=...)` shape, so switching clouds never touches agent code. `openai`, `gemini`, `ollama`, and `local` are not runtimes. Use `LocalAgent(config=LocalAgentConfig(model="..."))` for those, and add `compute="..."` when you want cloud-sandboxed tools with a local loop. `daytona` is the only runtime in the matrix that persists between runs — pick it when you want a dev workspace to survive across sessions. Remember `memory_mb` is megabytes; the provider converts to GiB. *** ## Related The base class and its Anthropic path. Short form: `Agent(run_on="daytona")`. Register your own runtime via entry points. `provider="docker"` runs the whole loop in a container you own.