Integrating Dust Agents with Your API Using MCP Protocol

Maxime Laurenty · 2025-05-12 17:32:23.130211+00

Hi, I’m POC-ing Dust and new to agents in general. I want to provide to a Dust agent access to an API that we have developed, let it choose which endpoint to call and provide its output (nicely formatted). I’ve the impression that it’s what the MCP protocol is there for, but it’s not clear to me: • how I could setup our app to act as an MCP server (do I just need a `/list` endpoint with a json schema of all our endpoints?) • how would I connect our MCP server to Dust? I’ve seen but on “add Tools” I don’t see the “add MCP server” button Is this the right approach or is there a simpler way to do it? Should I build something myself, for example by creating an app, using the block “curl” to call `/list`, passing it to a block “chat” to decide which endpoint to call and then adding a “curl” again?

Integrating Dust Agents with Your API Using MCP Protocol | Dust Community

7 comments

· Sorted by Oldest

David Ebbo
·
·
You don't see the option because it's still behind a feature flag. However, my take is that if your endpoint is not already an MCP server, the easiest approach is to just use a curl block to call it as it exists today. You could make it a long term goal to make it an MCP server, but it's more work than just adding a /list endpoint. With the curl approach, I would stay away from the chat block. Instead, create two apps: one that calls list and gets all the schema back, and one that invokes actions. I did something quite similar and it works well.
🙏1
Maxime Laurenty
·
·
Thanks for the hint. I temporarily added an openAPI doc as data source and created an app to perform the endpoints call. The agent understands basic things (like when I tell him “retrieve data X for vessel Y” and there is an endpoint “/get-X”) but I’m getting weird errors when I ask less straightforward questions. (the dust app seems to still be used properly based on its logs, so it’s happening in the agent after having called the dust app) I’ll try to tweak it but given how systematic this error is, I’m a bit confused.
David Ebbo
·
·
This has come up with Claude before, and I'm not sure if the cause was identified. e.g. here: https://dustcommunity.slack.com/archives/C06SHT0F20G/p1745499157134259. Can you try with GPT4 instead?
Maxime Laurenty
·
·
Interesting, With prompts :
@market-intel-api what are the specs of vessel 34 ? @market-intel-api What other information can you provide me about this vessel? Any workpackage for example?
I get the above error on Claude. On GPT4o:
I got a black screen (twice) when testing from the agent preview page.
No black screen when testing from the main chat page (but it chose poorly the parameters for the query)
Weirdly enough with:
@market-intel-api Can you retrieve the specs for vessel 31 ? @market-intel-api Can you provide me the workpackages for it? @market-intel-api can you retrieve all information about this vessel?
Claude manages to answer properly, and selects the good API parameters. 🙃
David Ebbo
·
·
I would think that with proper agent instructions, it could do all the steps at once. In my case, my instructions were something like: "Use the list_tools app to get a list of available tools and their schema. You can then use any of those tools via the invoke_tool app." Then when I ask anything, it goes "let's first call the list tool to find what's available" and then "the XYZ tool looks like a great fit for this task" etc... Also, by looking at your apps' logs, you can see exactly how they were invoked by the agent (choose API instead of Design in the App Logs page).
Maxime Laurenty
·
·
The app API logs show that calls are all valid, GPT just forgot an important one (vesselId). My instructions are in the idea you suggest: ” You are an assistant that is tasked to call Spinergie market-intelligence API and display it in nice visualization. For this, your reasoning should be: 1) read the documentation of the API provided in your data sources 2) based on the request and the API doc, decide which endpoint(s) to call and with which parameters. You can select as many endpoints as you wish. If you find no suitable endpoint, ask for more clarification and potentially state which endpoints you are hesitating with. 3) use the dust app “use_market_intel_api” to make the API calls. It expects to receive: - an endpoint with a leading / - an (optional) parameterBlock with parameters formatted as a query string (examples: “?windfarmId=10&vesselId=4,339,5228" ; “?managerId=40” ; “”) 4) state which endpoint(s) you have called 5) if needed, generate react code that will be rendered in the web browser directly. This can be used to help visualize the data returned by the dust app. Endpoints about “trend” or “aggregated” probably needs to be visualized as graphs (lines), while endpoints like “spec” don’t need it. 6) provide the raw dust app output in the end. If you get an error (from the dust app or from anything else), provide all the info that was passed to you or partially answered to debug easily. ” Anyway, I think this is good enough for this particular agent for now. Thanks for the tips!
David Ebbo
·
·
Yeah, it can be hit and miss. Your prompt looks solid!
🙏1

David Ebbo
·
·
You don't see the option because it's still behind a feature flag. However, my take is that if your endpoint is not already an MCP server, the easiest approach is to just use a curl block to call it as it exists today. You could make it a long term goal to make it an MCP server, but it's more work than just adding a /list endpoint. With the curl approach, I would stay away from the chat block. Instead, create two apps: one that calls list and gets all the schema back, and one that invokes actions. I did something quite similar and it works well.
🙏1
Maxime Laurenty
·
·
Thanks for the hint. I temporarily added an openAPI doc as data source and created an app to perform the endpoints call. The agent understands basic things (like when I tell him “retrieve data X for vessel Y” and there is an endpoint “/get-X”) but I’m getting weird errors when I ask less straightforward questions. (the dust app seems to still be used properly based on its logs, so it’s happening in the agent after having called the dust app) I’ll try to tweak it but given how systematic this error is, I’m a bit confused.
David Ebbo
·
·
This has come up with Claude before, and I'm not sure if the cause was identified. e.g. here: https://dustcommunity.slack.com/archives/C06SHT0F20G/p1745499157134259. Can you try with GPT4 instead?
Maxime Laurenty
·
·
Interesting, With prompts :
@market-intel-api what are the specs of vessel 34 ? @market-intel-api What other information can you provide me about this vessel? Any workpackage for example?
I get the above error on Claude. On GPT4o:
I got a black screen (twice) when testing from the agent preview page.
No black screen when testing from the main chat page (but it chose poorly the parameters for the query)
Weirdly enough with:
@market-intel-api Can you retrieve the specs for vessel 31 ? @market-intel-api Can you provide me the workpackages for it? @market-intel-api can you retrieve all information about this vessel?
Claude manages to answer properly, and selects the good API parameters. 🙃
David Ebbo
·
·
I would think that with proper agent instructions, it could do all the steps at once. In my case, my instructions were something like: "Use the list_tools app to get a list of available tools and their schema. You can then use any of those tools via the invoke_tool app." Then when I ask anything, it goes "let's first call the list tool to find what's available" and then "the XYZ tool looks like a great fit for this task" etc... Also, by looking at your apps' logs, you can see exactly how they were invoked by the agent (choose API instead of Design in the App Logs page).
Maxime Laurenty
·
·
The app API logs show that calls are all valid, GPT just forgot an important one (vesselId). My instructions are in the idea you suggest: ” You are an assistant that is tasked to call Spinergie market-intelligence API and display it in nice visualization. For this, your reasoning should be: 1) read the documentation of the API provided in your data sources 2) based on the request and the API doc, decide which endpoint(s) to call and with which parameters. You can select as many endpoints as you wish. If you find no suitable endpoint, ask for more clarification and potentially state which endpoints you are hesitating with. 3) use the dust app “use_market_intel_api” to make the API calls. It expects to receive: - an endpoint with a leading / - an (optional) parameterBlock with parameters formatted as a query string (examples: “?windfarmId=10&vesselId=4,339,5228" ; “?managerId=40” ; “”) 4) state which endpoint(s) you have called 5) if needed, generate react code that will be rendered in the web browser directly. This can be used to help visualize the data returned by the dust app. Endpoints about “trend” or “aggregated” probably needs to be visualized as graphs (lines), while endpoints like “spec” don’t need it. 6) provide the raw dust app output in the end. If you get an error (from the dust app or from anything else), provide all the info that was passed to you or partially answered to debug easily. ” Anyway, I think this is good enough for this particular agent for now. Thanks for the tips!
David Ebbo
·
·
Yeah, it can be hit and miss. Your prompt looks solid!
🙏1