Self-Hosting an Agent Platform: What to Prepare First
Before you self-host an agent platform, prepare five kinds of resources: a runtime environment (servers and database sized to your usage), model services (which models and how they are billed), the systems to connect and the account permissions, people to run it (upgrades, backups, monitoring), and data and audit requirements. Trial the business scenario in the cloud first, then prepare self-hosting along these five lines. FIM One by Beijing FIM Network Technology Co., Ltd. offers a free, self-hosted Community edition with the full feature set, deployed with Docker Compose in production and compatible with any OpenAI-protocol model service.
Once a team decides to self-host an agent platform, the first question is often how many servers it needs. Servers are only one item. What decides whether go-live goes smoothly is whether five kinds of resources are ready: a runtime environment, model services, the systems and permissions to connect, people to run it, and data and audit requirements.
A sensible order: trial in the cloud first to confirm one business scenario works, then prepare self-hosting along these five lines. The FIM One website by Beijing FIM Network Technology Co., Ltd. recommends the same: the cloud version is for trial and evaluation, and production can be self-hosted or deployed privately inside your network.
1. Five resources in one table
| Resource | What to prepare | Who decides |
|---|---|---|
| Runtime environment | Servers, database and storage sized to users and workload; separate test and production | IT or operations lead |
| Model services | Which model provider, endpoint and quota, how costs are settled; whether different tasks use different models | Business lead together with IT |
| Systems and permissions | Which systems to connect first, with which accounts, and which read and write operations to open | The data owner of each system |
| People to run it | Who handles upgrades, backups, monitoring and incidents, and who to call when something breaks | Named by the department head |
| Data and audit | Where data lives, how long operation logs are kept, who can view and export them | Compliance or security lead |
2. Runtime: start at pilot size, leave room to grow
There is no need to size for the whole company on day one. Start from the pilot team's headcount and workload, and keep test and production separate. For FIM One, the website says the Community edition deploys with Docker Compose and production can scale out as usage grows. Installation and configuration steps live in the developer docs and are not covered here.
3. Model services: pick a provider, then work out the cost
You self-host the platform; models are usually connected separately. Before go-live, settle three things: which model service to use, how it is billed, and what happens when the quota runs out. FIM One works with any model service that offers an OpenAI-protocol endpoint, can assign different models to different roles, and lets you switch models without changing business workflows. Model fees recur every month after launch; for how the costs break down, see Private Agent Platform Deployment: What the Cost Covers.
4. Systems and permissions: connect one, open only what is needed
Pick one system and connect only the operations one task needs. Use a dedicated account for each system and start with read-only access; operations that create, update or delete data or start approvals should wait for a person to confirm before they run. Write the connection scope as a list and have each system's data owner confirm it.
5. People: self-hosting needs an owner
Self-hosting means upgrades, backups, monitoring and incident handling are yours. Name the people before go-live: how often to upgrade, where backups go and how to restore them, and who takes over when the service fails. No one owning operations is a common problem after self-hosted launches. For weighing self-hosting against cloud, see Private deployment or cloud for AI agents: three factors.
6. Data and audit: write it down before launch
Where data lives, how long operation logs are kept, and who can view and export them should be agreed with the compliance or security lead before launch. Every call the agent makes to a system should leave a record that can be filtered by person, system and time.
FAQ
**Do we need to buy many servers before self-hosting?**No. Start at pilot size, keep test and production separate, and scale out as usage grows.
**Do we have to host the model ourselves?**Not necessarily. The platform is self-hosted; the model can be an external model service or one deployed inside the company. Whether data may leave your network decides which.
**What is the difference between Community and Enterprise?**FIM One Community is free, self-hosted and has the full feature set. Enterprise runs the same codebase and adds compliance tooling, dedicated support, SLA guarantees and architecture reviews.
**How is FIM One licensed?**FIM One uses a source-available license. The full code is public on GitHub and can be read, forked and self-hosted; the terms are in the repository's LICENSE.
**We are not sure about self-hosting yet. What now?**Read Cloud trial and private deployment to see how the two differ, validate the business scenario on the cloud version, then decide how to deploy.
Summary
Before self-hosting an agent platform, prepare five resources: runtime environment, model services, systems and permissions, people to run it, and data and audit. Trial the scenario in the cloud first, then start at pilot size, open only the permissions you need and name an operations owner.