What you can build

Full Claude Code capabilities. Real workspaces. No wrappers.

When you build a NextAgent agent, you're configuring a real Claude Code CLI process pointed at a real directory on a machine you own. The entire Claude Code toolchain is available: Bash, Read, Write, Edit, Glob, Grep, WebFetch, MCP servers, custom skills — everything.

There's no proprietary agent framework to learn. If you can build it with Claude Code in a terminal, you can deploy it through NextAgent to everyone on your team.

Agents point at existing project directories. NextAgent doesn't manage workspace contents, CLAUDE.md files, or MCP configs — you control those directly on the machine, the same way you would for yourself.

File system access

Agents can read, write, and edit files within the configured workspace directory.

Shell execution

Run build commands, tests, scripts — anything Bash can do, scoped to the workspace.

MCP servers

Any MCP server configured in the workspace .mcp.json is auto-enabled at spawn.

Custom skills

Deploy SKILL.md files to agent workspaces. Inject consistent workflows across your team's agents.

Web access

WebFetch and WebSearch available. Combine with MCP for richer integrations.

Persistent memory

Agents maintain conversation history and memory files across sessions, stored on your hardware.

Deployment

Build on your machine. Ship to the org in one click.

The handoff from "I built something" to "the whole team is using it" is intentionally minimal. No packaging, no container builds, no deployment pipelines.

1

Build and test locally

Write your CLAUDE.md, configure MCP servers, test with Claude Code directly in your terminal. Build it until it works the way you want.

2

Create an agent in NextAgent

Point the agent at the machine and workspace directory you've been developing in. Give it a name your colleagues will recognize. Choose the model.

3

Admin assigns it to the team

An org admin assigns the agent to specific users or the whole org. From that moment, it appears in their chat interface — no installation, no setup on their end.

4

Iterate like you normally would

Update the workspace on your machine and the changes are live immediately. No redeploy step. The next conversation your colleague has picks up the updated agent.

Who does what
You

Build the agent on the org machine using Claude Code and your normal dev workflow

You

Register the agent in NextAgent — workspace path, model, name, description

Admin

Reviews and assigns the agent to team members or the whole org

Team

Opens a chat tab and starts using it — no install, no keys, no training

Models

Pick the model that fits the task. Change it any time.

NextAgent manages inference routing. Model selection is per-agent — set it once when you create the agent, adjust later without rebuilding anything.

Fireworks

Open-weight, ZDR

The default path for most agents. Fast, private, zero data retention. Recommended for agents that process sensitive org data.

DeepSeek V4 Pro Kimi K2.6 + more
OpenRouter

Broad catalog, ZDR routes

Access to hundreds of models through a single path. NextAgent routes through ZDR-only providers. Good when you need a specific capability a frontier model provides.

100+ models ZDR-only routing
Anthropic

Claude via subscription

If you have a Claude Max subscription, credentials stay on your machine. NextAgent dispatches the agent without ever seeing your auth. Full Claude capabilities, your existing plan.

Claude Opus 4 Claude Sonnet 4 Claude Haiku 4

Admins control which providers and models are available to each role. You build against what the org has authorized — no billing surprises, no policy exceptions needed for standard use.

Configuration

Standard Claude Code configuration. Nothing proprietary.

Everything that controls agent behavior lives in the workspace on your machine: CLAUDE.md for system context, .mcp.json for MCP servers, skills/ directories for reusable workflows, per-agent permission profiles for tool access.

NextAgent adds one thing: a .nextagent/ directory in the workspace that holds the na CLI tool and env configuration. It doesn't touch your existing Claude Code setup.

The na CLI gives agents access to publish artifacts to S3, run searches, and integrate with org-level NextAgent features from within an agent run. Install it once per workspace; it auto-updates.

workspace layout
my-agent-workspace/
CLAUDE.md # system context + instructions
.mcp.json # MCP server definitions
.claude/settings.local.json
skills/ # reusable workflow .md files
.nextagent/ # added by NextAgent
bin/na # na CLI tool
env # org keys, inference config
... your project files
Permissions

Scope what agents can do. Lock it down as tightly as you need.

Per-agent permission profiles control which Claude Code tools the agent can invoke. A read-only research agent should never run Bash. An ops agent might need it. Set the profile once; NextAgent enforces it at every spawn.

Read-only profile

Allows Read, Glob, Grep, WebFetch, WebSearch. Suitable for research, Q&A, and documentation agents.

Standard profile

Adds Bash and file-write tools. Suitable for most coding and task agents.

Custom profile

Specify allowed tools explicitly. Combine with MCP tool allowlists for precise control.

Service key access

Control which external services (search APIs, publish endpoints) each agent can reach via the na CLI.

Documentation

Full reference for agent configuration, permission profiles, the na CLI, MCP integration, and the NAEP protocol that connects your machine to the control plane.

View docs →
Get started

Provision a machine. Build your first agent. Ship it today.

The free tier includes one machine and five users. Enough to build, test, and put a real agent in front of your team before you decide on anything else.

Get started free How it works →