Agent infrastructure API · checked 28 Sept 2026
DeskFerry vs Claude Managed Agents
Claude Managed Agents is Anthropic's API for running agents in the cloud. Anthropic hosts the agent loop and a sandbox per session; your developers define the agent, wire up tools through MCP servers, and build whatever interface people use. DeskFerry is the other side of build vs buy: a finished app where the person who owns the work writes a brief.
Choose Claude Managed Agents when
- you're building agents into your own product or platform
- you have engineers who want direct control of the model, prompt and tools
- the agent needs a real sandbox: bash, files and code execution
Choose DeskFerry when
- the users are business teams, not developers
- you want Slack, approvals and 1,500+ app connections without building them
- you'd rather pay a flat plan than meter runtime, tokens and searches
01 · What it is
Claude Managed Agents: an API (in beta) for hosted agents
DeskFerry: a finished app for business teams
02 · Who builds
Claude Managed Agents: developers, in code or config files
DeskFerry: whoever owns the task
03 · Priced by
Claude Managed Agents: $0.08 per session-hour + tokens + searches
DeskFerry: $49 / $99 / $299 a month
04 · Tools
Claude Managed Agents: sandbox, MCP servers, Skills, your own
DeskFerry: 1,500+ connected apps
Build vs buy
What 'set up an agent' means on each side
On Managed Agents, an agent is an API object you create once and run in sessions. This is the minimum setup code before your own interface, credentials and scheduling. In DeskFerry, the setup is the brief.
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();
// Once: where tools run
const environment = await client.beta.environments.create({
name: "ops-env",
config: { type: "cloud", networking: { type: "unrestricted" } },
});
// Once: the agent (persisted and versioned)
const agent = await client.beta.agents.create({
name: "Invoice chaser",
model: "claude-opus-5",
system: "Chase overdue invoices politely. Ask before sending.",
mcp_servers: [{ type: "url", name: "billing", url: MCP_URL }],
tools: [
{ type: "agent_toolset_20260401", default_config: { enabled: true } },
{ type: "mcp_toolset", mcp_server_name: "billing" },
],
});
// Every run: a session, then send a message and stream events
const session = await client.beta.sessions.create({
agent: agent.id,
environment_id: environment.id,
vault_ids: [vault.id], // credentials for the MCP server
});Powerful and flexible. You still build the MCP server or pick one, store credentials in a vault, handle the event stream, and give people a way to use it.
Brief
Every Monday, find invoices more than 14 days overdue in Xero, draft a polite reminder to each customer's billing contact, and send them after I approve in #finance.
Allowed apps
Last run
- Found 6 invoices over 14 days overdue
- Matched billing contacts for all 6 from Xero
- Drafted 6 reminders, firmer for the two past 45 days
- 6 drafts waiting for approval in #finance
No code, no hosting, no interface to build. You trade low-level control for a working agent this afternoon.
Feature by feature
Where each one wins
| Feature | Claude Managed Agents | DeskFerry | Edge |
|---|---|---|---|
| Who it's for | Developers building agent products | Business teams automating their own work | Depends |
| Control | Model, system prompt, tools, versions, permissions: all yours | A brief and the apps it may use | Claude Managed Agents |
| Execution sandbox | Per-session container with bash, files and code | No code execution | Claude Managed Agents |
| App connections | MCP servers you configure, plus Skills and custom tools | 1,500+ apps with sign-in | DeskFerry |
| Interface for users | You build it | Web app, Slack app, chat widget | DeskFerry |
| Scheduling | Scheduled deployments via the API | Set a schedule on the agent | Depends |
| Models | Claude models, pinned per agent | Managed for you | Claude Managed Agents |
| Pricing | $0.08 per session-hour, tokens at list price, $10 per 1,000 searches | $49 / $99 / $299 a month | Depends |
Sources: Managed Agents docs · Claude pricing · deskferry.com/pricing · checked 28 Sept 2026
How the bill works
Three meters vs one plan
Managed Agents charges for what a session uses: $0.08 for every hour the session runs, model tokens at the model's list price, and web searches at $10 per 1,000. It's efficient for engineering teams who can watch and tune usage.
DeskFerry wraps the model, the runtime and the app connections into one credit balance on a monthly plan. It's less granular and easier to budget for a team that doesn't read token reports.
The units differ, so compare what each bar covers. A developer-built agent also costs engineering time, which no bar shows.
Straight talk
When to stay on Claude Managed Agents
You're building a product
If agents are part of what you sell, you want the API, versioned configs and full control over the experience. DeskFerry is an end-user app.
The job needs a sandbox
Code execution, file work and long-running tasks in a container are what Managed Agents is built for.
You have the engineers
A team that already builds on the Claude API can go further with Managed Agents than any no-code tool allows.
FAQ
Claude Managed Agents questions
What is Claude Managed Agents?
A beta API from Anthropic for server-hosted agents. You create a versioned agent (model, system prompt, tools, MCP servers, Skills), then run it in sessions. Anthropic runs the agent loop and gives each session its own container for bash, file and code tools.
How is Claude Managed Agents priced?
By usage: $0.08 per hour of session running time, model tokens at list price, and web searches at $10 per 1,000. There's no monthly plan.
Does Claude Managed Agents connect to apps?
Yes, through MCP servers you configure on the agent, with credentials stored in vaults, plus Skills and your own custom tools. It doesn't come with a catalogue of ready-made business app connections the way DeskFerry does.
Can non-developers use Claude Managed Agents?
Not directly. It's an API: someone writes code or config files to define agents, and someone builds the interface people use. DeskFerry is built for non-developers.
Is Managed Agents available on AWS Bedrock or Google Vertex?
Managed Agents is an Anthropic API feature. On Bedrock, Vertex AI and Microsoft Foundry, Anthropic points developers to the Claude API with tool use instead.
Skip the build. Keep the agent.
Brief an agent in plain English, connect the apps, and run it from Slack this afternoon.
