I am really impressed with your demo video, particularly the analytics, logs, and test suite features.
I'm a target customer where I have a few curious customers, but I'm not fully ready to roll it out yet across the customer base. One thing that's stopping me, what does credits mean on your pricing page? And what is the pay as you go price after you hit your limit? I would need to be able to budget this before I deploy.
The second piece - we already have a CLI, which is great for terminal based agents and what we will continue to recommend. What we really want, and I think is what you are offering, is basically an easier way to deploy a 'remote connector' to use Claude lingo so that normal users with the claude/chatgpt app can just use our MCP. Can you point me to guidelines or the right place in your open source templates to understand how I would best handle auth (or the tradeoffs in each) during the initial build phase of the MCP server?
thank you so much we have put a lot of love into our product!
pricing: we offer several credit based products that have different cost - requests, build minutes, eval runs, checklist all have different unit cost the breakdown is here https://docs.manufact.com/dashboard/billing thanks for pointing out that it should be more transparent!
This is cool. I was skeptical of MCP's until I made one recently. They're essentially the exact same as 1) giving your agent a CLI tool or REST API and 2) pointing it there in an AGENT.md/CLAUDE.md. Agents are great at using built-for-human CLI tools and IMO they don't need anything purpose-built for agents. The key difference, which ends up being a usability win for non-technical users, is that the MCP bundles 1 and 2 - harnesses inject the MCP tool descriptions on every session after install. Of course, that's also why you need to be careful about context bloat when using/building them
+1 !! Thanks for putting it so clearly, also CLIs and REST have their space in some applications, I think MCP is a way to organize them and distribute them better
MCP is something people think of as a "panacea" to all AI issues. I think people are beginning to realize it is just one, albeit important, part of a successful AI architecture.
First of all, people have been saying this for a long time. It’s nothing new, at least half a year, maybe closer to a year.
Secondly, it’s not even that important, it’s the tool calling itself that’s important. MCP servers are just a convenient way to interact with remote services when a command line utility for the same would be inconvenient.
One issue i saw with scoped is the offline_access scope that causes frequent reauthenticate request from the client. for example codex has this bug (https://github.com/openai/codex/issues/20503). many servers solve this with some workarounds or increasing the token life. In v2 of mcp-use coming end of the month there should be a builtin way to deal with buggy clients so the server remain connected.
You may want to have a look at Skybridge, a TypeScript framework designed to build MCP servers and MCP Apps, with a recent emphasis on making authentication easy.
Not for every situation. CLI is great for coding agents (and I'd agree, far better in most cases than MCP). But it requires some execution runtime somewhere to actually run. So for app use cases where you don't want to build out your own tools for every integration, MCP can be a solid option.
Also, in my opinion, it's much easier to build a good MCP interface than it is to build a good CLI interface - and afaik there's support for MCP tools to return things like images from an MCP tool call directly to the calling LLM that is a bit tricker to do via CLI.
CLI certainly is better than local MCP. But nowadays, most MCPs are remote and the comparison fall short, at the notable exception of `gh` in a coding environment. But having CLI already authenticated is not guaranted either!
MCP makes a lot, lot more sense when you think of it as as a auth standard and not a comparison with CLIs. It obviously does more than just auth, but having standardised auth (which CLIs definitely do not) is the real 'killer' feature.
I am so tired of people repeating this. Usually, this results from conflating two uses of MCP: local, which can indeed be replaced by CLI (and you can argue which one is better), and remote, which is entirely different, and there is no way to replace it with a CLI (note that you are making an implicit assumption that a CLI tool can be used at all, which is not always the case).
Please don't repeat this. It's like saying that apples are dead and oranges are the future.
> Tell me you don't understand what you're saying without telling me you don't understand what you're saying.
Please don't cross into personal attack, regardless of how wrong someone is or you feel they are.
Your comment would be fine without that last swipe, and even better if you had gone on to say what the two purposes are. Then we could learn something from it.
Love it guys! Also working deep in the MCP space we’ve used manufact right since the start. Great product, team and new release! Congrats
I am really impressed with your demo video, particularly the analytics, logs, and test suite features.
I'm a target customer where I have a few curious customers, but I'm not fully ready to roll it out yet across the customer base. One thing that's stopping me, what does credits mean on your pricing page? And what is the pay as you go price after you hit your limit? I would need to be able to budget this before I deploy.
The second piece - we already have a CLI, which is great for terminal based agents and what we will continue to recommend. What we really want, and I think is what you are offering, is basically an easier way to deploy a 'remote connector' to use Claude lingo so that normal users with the claude/chatgpt app can just use our MCP. Can you point me to guidelines or the right place in your open source templates to understand how I would best handle auth (or the tradeoffs in each) during the initial build phase of the MCP server?
thank you so much we have put a lot of love into our product!
pricing: we offer several credit based products that have different cost - requests, build minutes, eval runs, checklist all have different unit cost the breakdown is here https://docs.manufact.com/dashboard/billing thanks for pointing out that it should be more transparent!
auth: we published a few blogs https://manufact.com/blog/oauth-mcp https://manufact.com/blog/authentication
+ lots of auth templates you can start from https://manufact.com/templates
This is cool. I was skeptical of MCP's until I made one recently. They're essentially the exact same as 1) giving your agent a CLI tool or REST API and 2) pointing it there in an AGENT.md/CLAUDE.md. Agents are great at using built-for-human CLI tools and IMO they don't need anything purpose-built for agents. The key difference, which ends up being a usability win for non-technical users, is that the MCP bundles 1 and 2 - harnesses inject the MCP tool descriptions on every session after install. Of course, that's also why you need to be careful about context bloat when using/building them
Yeah MCPs are easier to distribute.
In my experience, they also work better with dumber models than CLIs (which saves money)
+1 !! Thanks for putting it so clearly, also CLIs and REST have their space in some applications, I think MCP is a way to organize them and distribute them better
Looks great! Congratulations on the launch, guys!
stopped using mcp and mostly using skills now. can't understand what this product does or how it could help me.
MCP is something people think of as a "panacea" to all AI issues. I think people are beginning to realize it is just one, albeit important, part of a successful AI architecture.
First of all, people have been saying this for a long time. It’s nothing new, at least half a year, maybe closer to a year.
Secondly, it’s not even that important, it’s the tool calling itself that’s important. MCP servers are just a convenient way to interact with remote services when a command line utility for the same would be inconvenient.
congrats guys! how does the automated testing using ChatGPT/Claude clients work?
we have so many problems with MCP related to auth, scopes etc... how do you guys help solve that?
One issue i saw with scoped is the offline_access scope that causes frequent reauthenticate request from the client. for example codex has this bug (https://github.com/openai/codex/issues/20503). many servers solve this with some workarounds or increasing the token life. In v2 of mcp-use coming end of the month there should be a builtin way to deal with buggy clients so the server remain connected.
with our SDK we provide many adapters to popular authentication providers which basically provision oauth on your server in one line of code, here the docs https://docs.mcp-use.com/typescript/server/authentication/in...
also we ship templates for you to get started https://manufact.com/templates
You may want to have a look at Skybridge, a TypeScript framework designed to build MCP servers and MCP Apps, with a recent emphasis on making authentication easy.
TypeScript, for all its benefits, still feels like a toy or project language. I'd love to see a Rust, C++, or even a Go library for such purpose.
I'd love for people with experience to break me of this negativity towards TypeScript. Anybody?
Can you get into the specifics of the problems you currently experience with MCP? I'd love to learn more.
Huge congrats to the team on the official launch of Manufact!
we've been using manufact for months now. couldn't build mcps another way.
MCP is a deadend. CLI use is the future.
Not for every situation. CLI is great for coding agents (and I'd agree, far better in most cases than MCP). But it requires some execution runtime somewhere to actually run. So for app use cases where you don't want to build out your own tools for every integration, MCP can be a solid option.
Also, in my opinion, it's much easier to build a good MCP interface than it is to build a good CLI interface - and afaik there's support for MCP tools to return things like images from an MCP tool call directly to the calling LLM that is a bit tricker to do via CLI.
MCP is great for docs and stuff, also saves tokens and reduces errors if you have something complicated you're abstracting over
- agents have old/inaccurate knowledge and it's nice to have up to date docs: https://awslabs.github.io/mcp/servers/aws-documentation-mcp-...
- geting agents to do apple builds and stuff is much easier with: https://github.com/getsentry/XcodeBuildMCP
- also for searching stuff like pdfs/epubs it's nice to have a place that's easy/fast for an agent to go to: https://github.com/nburns/doc-search-mcp
none of these strictly requrie mcp, but it is still a useful abstraction/shared convention
CLI certainly is better than local MCP. But nowadays, most MCPs are remote and the comparison fall short, at the notable exception of `gh` in a coding environment. But having CLI already authenticated is not guaranted either!
Could you elaborate on this thought? MCP vs CLI feels very much like a Apples/Oranges comparison, without additional context.
MCP makes a lot, lot more sense when you think of it as as a auth standard and not a comparison with CLIs. It obviously does more than just auth, but having standardised auth (which CLIs definitely do not) is the real 'killer' feature.
I am so tired of people repeating this. Usually, this results from conflating two uses of MCP: local, which can indeed be replaced by CLI (and you can argue which one is better), and remote, which is entirely different, and there is no way to replace it with a CLI (note that you are making an implicit assumption that a CLI tool can be used at all, which is not always the case).
Please don't repeat this. It's like saying that apples are dead and oranges are the future.
cached thought. running CLIs is impractical and expensive in many environments and a hell of a lot less secure than using MCP
They serve two different purposes.
Edited*
> Tell me you don't understand what you're saying without telling me you don't understand what you're saying.
Please don't cross into personal attack, regardless of how wrong someone is or you feel they are.
Your comment would be fine without that last swipe, and even better if you had gone on to say what the two purposes are. Then we could learn something from it.
If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.
Apologies, I appreciate you posting this!