Agents — Overview

A NanoFleet agent is a Docker container running a Bun/TypeScript runtime built on the Mastra framework. Each agent has its own identity, memory, and tool access — and connects to the outside world through channels.

Architecture

┌──────────────────────────────────────┐
│             NanoFleet Agent          │
│                                      │
│  ┌──────────┐   ┌──────────────────┐ │
│  │ Identity │   │   LLM loop       │ │
│  │ SOUL.md  │──▶│  (Mastra/Vercel  │ │
│  │ STYLE.md │   │   AI SDK)        │ │
│  │AGENTS.md │   └────────┬─────────┘ │
│  └──────────┘            │           │
│                     ┌────▼────┐      │
│  ┌──────────┐       │  Tools  │      │
│  │  Memory  │       │ native  │      │
│  │ MEMORY.md│       │ + MCP   │      │
│  │HISTORY.md│       └─────────┘      │
│  └──────────┘                        │
│                                      │
│  Transport: HTTP (SSE streaming)     │
└──────────────────────────────────────┘

Native tools

Every agent has access to these tools by default:

Tool	Description
`webSearch`	Search the web
`webFetch`	Fetch a URL and extract content
`readFile`	Read a file from the workspace
`writeFile`	Write a file to the workspace
`editFile`	Edit a file with string replacement
`listDir`	List files in a directory
`execShell`	Execute a shell command (disabled by default)

Note: execShell is disabled by default for security. Enable it by setting SHELL_TOOL_ENABLED=true in the agent’s environment.

Supported models

NanoFleet agents support 600+ models via Mastra’s provider adapters:

Anthropic: claude-sonnet-4-6, claude-haiku-4-5, claude-opus-4-6 — requires ANTHROPIC_API_KEY
Google: gemini-2.0-flash, gemini-2.5-pro — requires GOOGLE_GENERATIVE_AI_API_KEY
OpenRouter: 600+ models (Llama, Qwen, Mistral, DeepSeek…) — requires OPENROUTER_API_KEY. Prefix the model ID with openrouter/: e.g. openrouter/meta-llama/llama-3.3-70b-instruct
MiniMax: minimax-m2.5 — requires MINIMAX_API_KEY. Prefix the model ID with minimax/: e.g. minimax/minimax-m2.5
vLLM: Self-hosted models via OpenAI-compatible API

Set the model with the AGENT_MODEL environment variable.

Agent lifecycle

Agents in NanoFleet have four lifecycle states:

State	Description
`running`	Container is active and accepting requests
`stopped`	Container is stopped, workspace preserved
`error`	Container exited with an error
`deploying`	Container is being created

From the dashboard or API:

Deploy — pull image, create container, start
Stop — stop container (workspace preserved)
Resume — restart a stopped container
Delete — stop + remove container + optionally remove workspace
Upgrade — pull latest image, recreate container, preserve workspace

Usage and cost tracking

Each agent tracks token usage and estimated cost per LLM call. Totals are shown on the Fleet Dashboard card for each agent and in the agent workspace view.

Cost is calculated using per-model pricing at the time of the call. Supported for all Anthropic, Google, and MiniMax models; OpenRouter models use the reported usage from the API response.