Assemble working agents visually from your skills — no code required.
Agent Builder is a developer platform for building, testing and running AI agents. Write the goal and tasks, attach skills, connect tools over MCP, let one agent call others as sub-agents, publish a version, and see what every run did and what it cost.
- Agents follow versioned skills from your library
- Tools reached through MCP connectors
- Agents can call other agents as sub-agents
- OpenAI, Anthropic and Google models
Three kinds of agent. Pick by what the job needs.
Start with the simplest type that does the work. You can add skills, tools and connectors later.
Conversational assistant
Smart Agent
Answers from its instructions and its own knowledge. No external tools or data connections.
- Custom instructions and tone
- Guardrails and boundaries
- Persona
Connected to your systems
Expert Agent
Reaches your business systems, knowledge bases and custom tools, and follows the skills you attach.
- MCP connectors
- Skills and knowledge bases
- Custom tools
Task automation
Flow Agent
Given a goal, runs a workflow with tools and skills. It keeps no conversation memory between runs.
- Goal-driven execution
- Tool and skill orchestration
- Stateless runs
Multi-agent
Agents can call other agents.
Add any agent in your workspace as a sub-agent. The lead agent calls it like a tool during a conversation and hands it one specialist job. Each sub-agent runs in its own context, with its own skills, tools and model.
What every agent is made of
- Goal and guardrailsWhat the agent is for and what it must not do.
- TasksThe sequence of steps it works through, with memory variables.
- SkillsReusable instruction sets from your library, in the order you set.
- ToolsHTTP tools for real APIs, or simulated tools while you design.
- Sub-agentsOther agents it can delegate to.
- KnowledgeCollections it can reference in a conversation.
- VoiceA phone number, if the agent takes calls.
- ModelThe model and its settings, chosen per agent.
Build, test, deploy, observe.
Four steps in one place, and what you learn in the last one goes back into the first.
- 01
Build
Write the agent’s instructions, attach skills, choose its tools and its model. Every edit is saved as a version.
- 02
Test On request
Chat with a draft before you publish. Agent Testing runs an agent against a set of test cases and scores the results.
Explore Agent Testing - 03
Deploy
Publish a version and put the agent to work: in a chat widget on your site, or called from your own code through the API.
- 04
Observe
Agent Insights shows each conversation, tool call and decision, with tokens and cost per agent.
Explore Agent Insights
The library your team writes is the library your agents follow.
An agent is its instructions plus three things you can reuse across every agent you build.
Skills are the method
A skill is a written procedure: how one piece of work is done. Your agents follow the same library your team already uses in Claude and ChatGPT, so the method is written once.
Browse the skill library
Connectors are the reach
An MCP connector is how an agent reads from and acts on a system: a database, a document store, a SaaS tool. You choose which tools each agent may call.
Browse MCP connectors
Plugins are the bundle
A plugin packages the skills and connectors for one job, so an agent can start with a working set and you edit from there.
Browse Claude plugins
Choose the model. Keep the record.
A model console is built for the first API call. Agent Builder is for the agents you keep running: which version answered, what it called, and what it cost.
Models
OpenAI, Anthropic and Google models. Pick the model for each agent and change it later without rewriting the skills it follows.
Versions
Agents and skills are versioned. Publish when a version is ready, compare it with the one before, and restore an earlier one if a change does not hold up.
Sessions
Open any conversation and read what the agent was asked, which tools it called and what it answered.
Cost
Usage is counted per agent, so you can see what each one costs to run instead of one bill per API key.
Built for more than one team.
Workspaces
Keep agents, skills and credentials apart by team, client or project. One workspace does not read another.
Roles
Decide who can view, edit and publish in each workspace. Builders are the paid seats; viewers and the people who use an agent are not.
A record of each run
Sessions are kept for review, with the agent version that produced them.
HIPAA, ISO 27001 and SOC 2 ready. Compliance audits underway.
Review security and governanceWe run our own product on it.
Customer-reported result at our founding design partner. Sachin also serves as Agentman's VP, Agentic Healthcare. Results vary by practice.
Medman, our agents for independent medical practices, is built and run on the same platform you sign up for. One result from a practice that uses it:
Explore Medman“We cut 50 hours of monthly verification calls down to 7.”
Free to start. Pay per builder as the team grows.
Free
One builder, up to 3 agents and 200 agent credits a month.
Team
Unlimited agents and 5,000 agent credits per builder a month. Viewers and agent users are free. 14-day trial.
Enterprise
Custom agent credits and pricing for larger organizations, with a dedicated support team.
Partner
For integrators and agencies that build agents for their clients.
Questions developers ask before they sign up.
What is Agent Builder?
Agent Builder is a developer platform for building, testing and running AI agents. You write an agent’s instructions, attach skills from your library, give it tools through MCP connectors, publish a version, and review what each run did and what it cost.
Which models does Agent Builder support?
OpenAI, Anthropic and Google models. You choose the model for each agent, and you can change it later without rewriting the skills the agent follows.
How do agents use skills and connectors?
A skill is a written procedure: how one piece of work is done. An agent follows the skills you attach to it, from the same library your team uses in Claude and ChatGPT. An MCP connector is how the agent reaches a system, such as a database, a document store or a SaaS tool. A plugin bundles skills and connectors for one job.
How are agents versioned?
Agents and the skills they follow are versioned. You publish a version when it is ready, compare it with the one before, and restore an earlier version if a change does not hold up.
Can I test an agent before it goes live?
You can chat with a draft in Agent Builder before you publish it. Agent Testing, which runs an agent against a set of test cases and scores the results, is available on request.
What does Agent Builder cost?
There are four plans. Free is $0. Team is $29 per builder per month, with a 14-day trial. Enterprise and Partner are priced per agreement. The Agent Builder pricing page lists what each plan includes.
What is the security status?
HIPAA, ISO 27001 and SOC 2 ready. Compliance audits underway. Our SOC 2 and ISO 27001 audits are underway, and neither has issued a report yet. Before any document is shared, we walk your security team through how it is handled.
Build the first agent today.
Start on the Free plan with three agents. Bring the team in when the first one earns its place.
No credit card for the Free plan.