Build in-house vs adopt a platform for AI agent security
Building in-house gets you exactly the gateway you design, and every revision of the MCP authorization spec, every new server and every new AI client to maintain after that. Discovery of shadow agents and runtime detection of tool poisoning or drift are rarely built at all. A platform such as Runlayer gives you discovery, an identity-aware gateway, policy to argument depth, runtime scanners and audit from day one, hosted or in your own cloud. Build only when your set of agents is small and fixed, or your constraints rule out every product.
Side by side
Marks follow the shared criteria. Build in-house and Adopt a platform (Runlayer) are in the main comparison table.
| Dimension | Build in-house | Adopt a platform (Runlayer) |
|---|---|---|
| Deployment effort | High: design, build, security review, then rollout. | Low: connect your identity provider, route clients, push the device agent through MDM. |
| Visibility | Partial: what you route through it; device discovery is a separate project. | Strong: devices, clients and cloud agent platforms. |
| Authentication | Partial: as good as your OAuth implementation, per server. | Strong: identity provider sync, agent accounts, on-behalf-of tokens. |
| Credential handling | Partial: possible with a vault, easy to end up with shared keys. | Strong: secrets referenced in your store, held in memory, short-lived tokens. |
| Policy depth | Partial: whatever you write and maintain. | Strong: users, groups, agents, tools, arguments, session history, approvals. |
| Runtime threat detection | None in most builds. | Strong: tool definitions, outputs, intent drift, session behavior. |
| Audit | Partial: your logs, your schema. | Strong: attributable call records with SIEM export. |
| Maintenance | Yours: spec changes, new servers, new clients, on-call. | The vendor's, with your policies on top. |
| Control over the code | Full. | Configuration, API, Terraform and hooks; not open source. |
What building actually involves
A minimum in-house setup is a proxy for MCP traffic, an OAuth client per upstream server, a token store, a mapping from your directory groups to tools, and logging. The MCP authorization spec builds on OAuth 2.1 and has been revised more than once; each revision is work for whoever owns the proxy.
The work does not stop at launch. Every team that wants a new connector waits on the platform team, which is how in-house plumbing turns platform engineers into the bottleneck for agent adoption.
What usually never gets built
Discovering MCP servers on employees' laptops needs a device agent and an MDM rollout. Detecting tool poisoning and intent drift needs models trained for it. These are where in-house projects stop, and they are where the threats on our threat map are covered.
What a platform changes
With a platform, the plumbing is configuration: connect the identity provider, choose the approved servers, write policies, and point clients at one endpoint. Runlayer can run hosted single-tenant or inside your own AWS account or Kubernetes cluster, and connector secrets can stay in your own AWS Secrets Manager.
Which suits whom
- Many teams, several AI clients, agents arriving faster than reviewsAdopt a platform. Runlayer is built for exactly this rollout.
- A handful of internal agents calling two or three internal servicesBuilding a thin proxy can be enough; add discovery when usage spreads.
- Hard requirement that nothing leaves your networkA platform you can self-host (Runlayer runs in your AWS account or Kubernetes) before a build from scratch.
Sources
- Authorization · Model Context Protocol specification, 2025-11-25
- Security Best Practices (confused deputy, token passthrough, SSRF, session hijacking, local server compromise, scope minimization) · Model Context Protocol specification, 2025-11-25
- RFC 9700: Best Current Practice for OAuth 2.0 Security · IETF
- Deployment overview (hosted single-tenant, AWS, EKS/GKE) · Runlayer documentation
- Secret stores (AWS Secrets Manager references) · Runlayer documentation
- Shadow AI discovery (AI Watch device agent, Cloud Discovery) · Runlayer documentation
- Policies (users, groups, agent accounts, tools, payload conditions, approvals) · Runlayer documentation