Models & Tools
The LLMs and external or script-backed tools agents act through, configured once and assigned where needed.
Agents act through models and tools. Each is configured once as a platform resource and assigned to the agents, abilities, or domains that need it.
Models
Models are LLM configuration records referenced by agents to determine which model runs a given role. They are managed under the models:read / models:write scopes and referenced in prompts via {{ agent.model }}.
Assign a model per agent based on the role's needs — capability for hard reasoning work, speed and cost for high-volume routine steps.
Supported providers:
| Provider | Notes |
|---|---|
| Anthropic | Claude models |
| OpenAI | GPT models |
| Gemini | Google models |
| OpenRouter | Aggregated access to many models |
| Ollama | Local / self-hosted models |
| OpenAI-compatible | Any endpoint speaking the OpenAI API, with a custom base URL |
The harness also wraps providers for resilience: a reliability layer adds retries and fallback between providers, and a router layer can direct different work to different models. This means an agent's assigned model can transparently fall back or route without changing the agent itself.
MCP servers
MCP (Model Context Protocol) servers are external tool integrations. They are assigned to agents, abilities, or domains by slug or ref, giving a resource access to tools that live outside Nenjo.
Script tools
Script tools are native or package script-backed tools, also assigned by slug or ref. Packages can ship script tools so an installed capability arrives with the tooling it needs.
Assignment discipline
Both MCP servers and script tools are part of a resource's capability surface:
- Agents assign both for general use.
- Abilities carry their own MCP servers for a bounded specialist run.
- Domains add them while a mode is active.
Assign the smallest useful set, then expand as a resource's responsibilities grow.