Build agents
Nexus runs the agent loop so that your application does not have to. You declare what the agent is — its tools, prompts, approval rules, model and hooks — and submit inputs; Nexus drives the model rounds, delivers tool calls to executors, parks calls for approval, compacts context, persists every turn and task, records usage and streams events back.
What Nexus does for you
- Durable conversations with inputs, turns, candidate answers, forks and per-conversation access.
- A round driver that calls the model, expands tool calls into tasks, waits for results and continues, with retries, fallback models and failure policies.
- Tool delivery to the agent application, runners and tool providers through one inbox protocol with exclusive claims and write-once results.
- An approval stage every tool call crosses, evaluated from your rules against opaque tool input.
- Compaction when context will not fit, with the kernel's summarizer or your own.
- Memory, an opaque store and uploads as durable state.
- Orchestration primitives: delegated branches, child conversations, messages between conversations, waits, authored DAG steps and your own code tool.
- Schedules, lifecycle hooks, usage records and retention.
What your application owns
- Prompt meaning and the system prompt, persona and summarizer documents.
- Model selection, fallback, reasoning settings and tiers.
- Which tools exist, what they mean, and the approval mode and rules that govern them.
- Orchestration strategy, memory extraction and ranking, user interface and external channels.
rho is one such application: an always-on personal agent and a coding agent, built only on these APIs. Your application can run beside it on the same Nexus with its own identity, prompts and rules, and rho's source is the worked example for every page in this section. Build your own agent orients a new integration; this section is the detail.
Choose how deep to go
| You want to | Use |
|---|---|
| A conversation reply from a declared agent | A Workspace, a Conversation and an input with kind: direct_reply — see the quickstart |
| Tools that run on a machine | A runner selected for the conversation, with its tools declared on the agent |
| Tools your application serves itself | Announce them on the application's own executor address and claim them from its inbox |
| Parallel or delegated work | Orchestration: delegate_task, spawn, send, wait, authored steps or a code tool |
| A model call with no conversation | A direct model call |
| Work on a timer | Schedules or a queued input with deliver_at |
The API families
| Family | Base path | Who calls it |
|---|---|---|
| Agent API | /agent_api/v1 | Agent applications on the member plane; runners, tool providers and applications on the executor transport plane |
| Platform API | /api/v1 | Humans for personal settings; owners and administrators under /api/v1/admin |
| OAuth | /oauth/* | Application login and device connection |
Start with Credentials and planes, then the quickstart. The Ruby SDK wraps the same endpoints with typed resources, credential rotation and durable event following.