Show HN: Building a Deep Research Agent Using MCP-Agent

Post by cyberninja15 » Fri, Sep 12, 2025, 7:45 PM UTC

diggan wrote: post #8

haniehz wrote: post #3

Earlier quoted context omitted.

based on the article, it seems like a good reasoning model like gpt5 or opus 4.1 might be good choices for the planner. I wonder if the gpt oss reasoning models would do well

Personally been using GPT-OSS-120b locally with reasoning_effort set to `high` and it blows pretty much every other local model out of the water, but takes a lot of time for it to eventually do a proper content reply. But for fire-and-forget jobs like "Create a well-researched report on X from perspective Y" it works really well.

Post by saqadri » Fri, Sep 12, 2025, 8:10 PM UTC

Zetaphor wrote: post #11

diggan wrote: post #7

I gotta say, having white blurry blobs of something in the background floating behind white/grey text maybe wasn't the best design-choice out there. None the less, I tried to find the actual APIs/service/software used for the "search" part, as I've found that to be the hardest to actually get right (at least for as-local-as-possible usage) for my own "Deep Research Agent". I've experimented with Brave's search API wh…

Self host an instance of SearXNG[1] either locally or on a remote server with a simple docker container and use its JSON API [2]. You have to enable the JSON API in the config manually [3]. [1] https://docs.searxng.org/admin/installation-docker.html#inst... [2] https://docs.searxng.org/dev/search_api.html [3] https://github.com/searxng/searxng/discussions/3542

Post by reachableceo » Fri, Sep 12, 2025, 8:46 PM UTC

luckydata wrote: post #10

koakuma-chan wrote: post #5

Earlier quoted context omitted.

Gemini 2.5 Pro is also a great reasoning model, I still prefer it over GPT 5

Gemini is great, it's just incredibly clumsy at tool use and that's why it fails so often in practice. I'm looking forward to the next version, it will for sure address it, it's a big issue internally too (I'm a recent xoogler).

Post by luckydata » Sat, Sep 13, 2025, 1:17 AM UTC

PantaloonFlames wrote: post #15

luckydata wrote: post #10

Earlier quoted context omitted.

Gemini is great, it's just incredibly clumsy at tool use and that's why it fails so often in practice. I'm looking forward to the next version, it will for sure address it, it's a big issue internally too (I'm a recent xoogler).

Can you elaborate on “clumsy at tool use”?

Post by jimmySixDOF » Sun, Sep 14, 2025, 3:45 PM UTC

saqadri wrote: post #9

diggan wrote: post #7

I gotta say, having white blurry blobs of something in the background floating behind white/grey text maybe wasn't the best design-choice out there. None the less, I tried to find the actual APIs/service/software used for the "search" part, as I've found that to be the hardest to actually get right (at least for as-local-as-possible usage) for my own "Deep Research Agent". I've experimented with Brave's search API wh…

Haha, I didn't have control on the blog website, just the content. The readme and code is the ultimate source of truth (and easier to read): https://github.com/lastmile-ai/mcp-agent/blob/main/src/mcp_a... So the core idea is the Deep Orchestrator is pretty unopinionated on what to use for searching, as long as it is exposed over MCP. I tried with a basic fetch server that's one of the reference MCP servers (with a si…