Comparison
Build your own MCP servers, or use actionlake
An alternative to self-built MCP for AI agents
Building your own MCP servers with the official SDKs gives you full control and no per-call fee, but each server must implement the MCP authorization spec and you build caps, approvals and audit yourself. actionlake is an MCP gateway with 995 vendor connectors that already has those controls, billed per upstream call with 100,000 calls a month free.
Checked 22 October 2026 · every fact links its source
Which to choose
It depends on the job.
The Model Context Protocol is an open protocol. Official SDKs exist in ten languages, six of them Tier 1 (TypeScript, Python, C#, Go, Rust and Ruby). A remote server is an OAuth 2.1 resource server: it must publish Protected Resource Metadata, accept only tokens issued for it, and never pass a client's token through to an upstream API.
Choose Self-built MCP if
- The system is internal or proprietary and no gateway covers it.
- Data must not pass through a third party, or must stay in a region actionlake does not run in.
- You need custom logic, resources or prompts inside the server, not just API actions.
- You have the engineering time to run, patch and harden each server.
Choose actionlake if
- You need many business systems connected now, not one server per system.
- You want spending caps per key, approvals and a hash-chained record without building them.
- You want one endpoint with eight tools instead of a tool list per server in the agent's context.
- You want cross-system answers: ask joins two systems on a shared entity in one call.
Credit where due
Where Self-built MCP is stronger
- Control: you decide every tool, every check and where the data goes.
- No third party in the data path, and no per-call fee to a gateway.
- Any system: internal APIs, databases and files, with your own logic.
- Your own compliance boundary: you can certify the server inside your own scope.
The difference
Where actionlake differs
- With your own servers, limits, approvals and records are code you write and maintain per server. actionlake applies them per key across 995 vendor connectors.
- The MCP spec says the protocol cannot enforce consent itself; implementors should build consent and authorization flows. actionlake's caps and approvals are enforced on the server before a request is sent.
- You can combine both: actionlake can draft a connector for your internal API from its documentation, or you keep a custom server beside it in the same client.
Feature by feature
15 things we compared.
Each Self-built MCP cell links the page we read it on. Checked 22 October 2026.
| Feature | actionlake | Self-built MCP |
|---|---|---|
| Hosted MCP server | YesOne endpoint, https://connect.actionlake.com/mcp, over Streamable HTTP; a REST API with the same behavior. | PartlyYou host it. The spec defines stdio and Streamable HTTP transports.Source: modelcontextprotocol.io/specification/latest |
| How a client signs in | YesOAuth 2.1 with dynamic client registration and PKCE, or an actionlake key. The agent holds an actionlake key, never a vendor token. | PartlyYou implement OAuth 2.1 as a resource server: Protected Resource Metadata (RFC 9728) is required, token audience must be validated, and other tokens must not be accepted or passed through.Source: modelcontextprotocol.io/specification/2026-07-28/basic/authorization |
| Apps and actions | Yes995 vendor connectors and 289,000+ actions behind eight tools. An agent can draft, test and sign a connector for an API that is missing. | PartlyYou write each server and each tool. Official SDKs in ten languages.Source: modelcontextprotocol.io/docs/sdk |
| What is billed | YesOne upstream call: a request actionlake sent to a vendor that returned a definitive response. Searching, checking, planning, cached answers and failures you did not cause (a vendor's 401, 403, 429, 5xx or a timeout) are free. | YesNo gateway fee. You pay for hosting, engineering time and the vendors' own API charges.Source: modelcontextprotocol.io/docs/sdk |
| Free allowance | Yes100,000 upstream calls a month, no card. | YesThe SDKs are open source.Source: modelcontextprotocol.io/docs/sdk |
| Limit which actions an agent can use | YesPer key: by action, by service, by verb, or read-only. A key can mint narrower keys, never wider ones. | PartlyScopes are part of the authorization spec; the server enforces them in code you write.Source: modelcontextprotocol.io/specification/2026-07-28/basic/authorization |
| Spending cap per agent key | YesA cap in upstream calls per key. A child key's calls count against each parent. Refused before the request is sent. | PartlyNot part of the protocol; you build it.Source: modelcontextprotocol.io/specification/latest |
| A person approves a write before it runs | YesDestructive actions need an explicit confirmation, and under the key's `human` policy the agent cannot give it, so a person approves. Approval rules make chosen actions, or every write, wait for a named person, who decides in the console or from an emailed signed link. | PartlyThe spec says hosts must get user consent before invoking a tool and that MCP cannot enforce this at the protocol level; a server-side approval step is yours to build.Source: modelcontextprotocol.io/specification/latest |
| Record of each call | YesOne ledger row per run, hash-chained and sealed daily, with the agent's stated reason (kept 90 days). Request and response traces are kept for the plan's window. | PartlyNot part of the protocol; you build it.Source: modelcontextprotocol.io/specification/latest |
| Scan call data for sensitive content | NoSecrets are removed from traces. Call data is not scanned or redacted for personal data. | PartlyNot part of the protocol; you build it.Source: modelcontextprotocol.io/specification/latest |
| One question across two systems | YesThe ask tool fetches from two or more systems, joins on a shared entity such as the customer, and returns the answer, not the rows. | PartlyAny tool can combine sources in code you write.Source: modelcontextprotocol.io/specification/latest |
| Event triggers and webhooks | NoBy design. Agents and schedules start the work; a scheduled plan can ask what changed since the last run. | PartlyWhatever your server implements.Source: modelcontextprotocol.io/specification/latest |
| Self-host | NoHosted service only: actionlake runs in one region in Sydney, Australia, and no self-hosted edition is offered. | YesYou run it wherever you choose.Source: modelcontextprotocol.io/docs/sdk |
| Security certifications | NoNone held yet. SOC 2 Type II, ISO 27001 and a third-party penetration test are in progress. | PartlyInside your own compliance scope.Source: modelcontextprotocol.io/specification/latest |
| Where data is processed | YesOne region: Sydney, Australia (DigitalOcean syd1). Stored in Australia at rest; TLS terminates at Cloudflare's edge. | YesWherever you host it.Source: modelcontextprotocol.io/docs/sdk |
"Not stated" means the vendor pages we read did not describe the feature. It is not a claim that the feature is missing. Checked 22 October 2026.
Pricing
What 100,000 and 1,000,000 calls a month cost.
Self-built MCP's unit: Your hosting and engineering time. Plan used: not applicable (no list price). actionlake's figure is the cheapest plan that fits the volume, from the published price book; one upstream call counts as one action.
| Calls a month | actionlake | Self-built MCP |
|---|---|---|
| 100,000 | US$0 a monthUS$0.00 per 1,000 callsFree: US$0 a month, 100,000 calls included. | Not publishedNo list price: hosting, engineering and maintenance for each server.Source: modelcontextprotocol.io/docs/sdk, read 22 October 2026 |
| 1,000,000 | US$219 a monthUS$0.22 per 1,000 callsPro: US$149 a month for 750,000 calls, plus 250,000 calls at US$0.28 per 1,000. | Not publishedNo list price: hosting, engineering and maintenance for each server.Source: modelcontextprotocol.io/docs/sdk, read 22 October 2026 |
Caveats
- The per-call cost of a server you run can be very low. The larger cost is building and maintaining auth, limits, approvals and records for each server.
- Vendor API charges apply either way; actionlake does not add them to its price.
How we priced it
- List prices on the vendor's own pricing page on the date shown, annual billing where the vendor offers it.
- The smallest plan or ladder step that covers the volume, plus published overage where that is cheaper.
- Each vendor bills a different unit. The unit is named on every row, and the conversion to one action or one tool call is stated.
- Where no public price covers the volume, the cell says so. Nothing is estimated from a sales quote.
actionlake's figures are the USD list. Searching the catalogue, checking a call, planning, cached answers and failed calls are not billed. See every plan on the pricing page.
Moving over
Migration notes
- Keep custom servers that hold internal logic or reach private systems; connect them beside actionlake in the same client.
- Replace servers that only wrap a public SaaS API with the matching actionlake connector.
- Where a system is missing, point your agent at the API's documentation: actionlake drafts, tests and signs a private connector.
- Move per-server limits into actionlake keys: allowed actions, a monthly cap, and approval for writes.
Questions
Asked about Self-built MCP and actionlake
What does the MCP spec require of a remote server?
A protected server acts as an OAuth 2.1 resource server. It must publish OAuth 2.0 Protected Resource Metadata (RFC 9728), validate that each access token was issued for it, and must not accept or pass through other tokens (modelcontextprotocol.io, revision 2026-07-28).
Which SDKs can I build with?
The MCP project lists official SDKs for TypeScript, Python, C#, Go, Rust and Ruby (Tier 1), Java (Tier 2), and Swift, PHP and Kotlin (Tier 3).
Can I use actionlake and my own MCP servers together?
Yes. An MCP client can connect to several servers. Keep custom servers for internal logic and use actionlake for the business systems it covers.
When is building my own server the better choice?
When the system is internal or unusual, when data must not leave your infrastructure or region, or when you need logic a set of API actions cannot express.
Sources
- MCP specificationmodelcontextprotocol.io/specification/latest · read 22 October 2026
- MCP authorization (revision 2026-07-28)modelcontextprotocol.io/specification/2026-07-28/basic/authorization · read 22 October 2026
- MCP SDKsmodelcontextprotocol.io/docs/sdk · read 22 October 2026
We wrote these pages, so read them with that in mind. Each competitor is stronger somewhere, and each page says where. If a fact here is out of date, tell us at [email protected] and we will correct it.
Give your agent every system, with limits.
Start free - 100,000 calls/month, no card. No signup. Hard caps on every key. Cached answers are free.