WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=pAnLpiAG6Es NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/pAnLpiAG6Es.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:12.559 --> 00:00:13.280 Cool. s2 00:00:13.280 --> 00:00:14.160 Hello everyone. s3 00:00:14.160 --> 00:00:16.240 Welcome to our talk. s4 00:00:16.240 --> 00:00:18.560 It's like one of the last talks of uh the conference. s5 00:00:18.560 --> 00:00:19.520 So, thanks for coming. s6 00:00:19.520 --> 00:00:23.279 U this talk is called MCP doesn't suck, your agent does. s7 00:00:23.279 --> 00:00:24.160 My name is Han Chern. s8 00:00:24.160 --> 00:00:26.800 I'm the founder and CEO of Appify. s9 00:00:26.800 --> 00:00:30.400 And I don't know if you've seen the these ones around here. s10 00:00:30.400 --> 00:00:31.439 Maybe I'll zoom here. s11 00:00:31.439 --> 00:00:34.559 Actually, we went great lengths to uh bring people to this talk. s12 00:00:34.559 --> 00:00:38.480 Uh didn't quite work out. s13 00:00:38.480 --> 00:00:40.640 But anyway, thanks for coming. s14 00:00:40.640 --> 00:00:46.079 So, as you know, uh MCP is a standard to securely access tools and resources, right? s15 00:00:46.079 --> 00:00:56.879 Basically, standard for agentto tool interaction and it's been introduced by by Antropic about almost two years ago and actually it took the world by storm, right? s16 00:00:56.879 --> 00:01:14.720 And actually last year MCP became the darling of the AI world like really like like a lot of people build their servers uh I think like now there's like 10 to 15,000 servers right a lot of clients adopted it and actually it became a standard you know like used by cloud or chpt to plug tools and connectors s17 00:01:14.720 --> 00:01:16.880 to your to your AI agents. s18 00:01:16.880 --> 00:01:24.479 There was like even so many MCP uh servers that that people had to create registries of MCP servers and there were so many registries that people have to create even MCP s19 00:01:24.479 --> 00:01:26.560 server registry registries. s20 00:01:26.560 --> 00:01:28.320 Pretty wild times, right? s21 00:01:28.320 --> 00:01:32.880 So when MCP connects agents with tools, everybody loves MCP, right? s22 00:01:32.880 --> 00:01:33.759 Right. s23 00:01:33.759 --> 00:01:35.840 Well, not exactly. s24 00:01:35.840 --> 00:01:39.600 Actually, especially last year, there have been a lot of hate about MCP. s25 00:01:39.600 --> 00:01:46.159 Like for example, Antropic admits that MCP sucks from to MCP is the wrong abstraction. s26 00:01:46.159 --> 00:01:49.280 Antropic is trying so hard to fix MCP. s27 00:01:49.280 --> 00:01:50.560 MCP was a mistake. s28 00:01:50.560 --> 00:01:54.240 Long live CLIs, right? s29 00:01:54.560 --> 00:01:56.399 MCP is dead in the water. s30 00:01:56.399 --> 00:02:01.280 Open Claw has shown me that API and CLI will will win. s31 00:02:01.280 --> 00:02:06.399 There is a reason I didn't add MCP support to Open Claw except via MCorter MCP to CLI converter. s32 00:02:06.399 --> 00:02:08.479 Oh, that's interesting. s33 00:02:08.479 --> 00:02:10.399 MCPS are mostly useless. s34 00:02:10.399 --> 00:02:13.120 I'll die on this hill if I have to. s35 00:02:13.120 --> 00:02:15.760 Every MCP could have been a deterministic CLI. s36 00:02:15.760 --> 00:02:18.239 That's what I'm going to do next year. s37 00:02:18.239 --> 00:02:22.879 And from Gary Tan who's going to speak today as well, MCP sucks honestly. s38 00:02:22.879 --> 00:02:23.680 Oh my god. s39 00:02:23.680 --> 00:02:26.879 And uh Peter levels, thank god MCP is dead. s40 00:02:26.879 --> 00:02:28.319 So I'm like what's happening? s41 00:02:28.319 --> 00:02:30.160 Why are everybody hitting MCPS, right? s42 00:02:30.160 --> 00:02:31.680 This is crazy. s43 00:02:31.680 --> 00:02:35.120 So first like let's look at the problem like what are people talking about? s44 00:02:35.120 --> 00:02:40.959 So the most slighted problem was like hey MCP eats too much context and actually that's true. s45 00:02:40.959 --> 00:02:58.480 I mean the first like way how agents implemented MCP was very naive right they were like hey MCP gives you tools so there is like if you have 10 MCP servers you just like 10 MCP tools so there's like 100 tools and agents would register all these tools in the context right away so before you actually even s46 00:02:58.480 --> 00:03:13.599 ask any question MCP would like eat like 100 like you know tools in in your context maybe like onethird of your context would be gone without actually like doing any work and honestly this is pretty crazy I mean if you think about it right and then you would be calling tools s47 00:03:13.599 --> 00:03:22.000 basically adding to the context the results would be added back to the context again and basically the context would just like grow very long and the wrote over time you lose accuracy s48 00:03:22.000 --> 00:03:32.799 it cost a lot of money and so on it's very ineffective way how to do that but that's not the problem of MCP that's problem of of the harness and if you look at the MCP specification s49 00:03:32.799 --> 00:03:43.519 what it says about like how you should design the harness like it says absolutely nothing it's up to the implementation right so it's your job when you're building your agent to make sure you use MCP effectively. s50 00:03:43.519 --> 00:03:46.879 It's not a problem of the protocol, right? s51 00:03:47.040 --> 00:03:49.120 So what are the solutions of this? s52 00:03:49.120 --> 00:03:54.000 So first solution like fairly naive is like hey let's split the context into sub agents. s53 00:03:54.000 --> 00:04:01.439 So if there's some task you can basically delegate it to a new sub agent it own context. s54 00:04:01.439 --> 00:04:03.680 So basically it doesn't pull you the main context window. s55 00:04:03.680 --> 00:04:06.959 I mean that's great but you still have to pay money for those tokens. s56 00:04:06.959 --> 00:04:08.080 It doesn't go away. s57 00:04:08.080 --> 00:04:11.680 And the problems like u before are still still around. s58 00:04:11.680 --> 00:04:25.919 For example, if the MCP tool uh returns some some sensitive results like password, basically it stays in your context and it can be abused by other other tool calls or basically like u application, right? s59 00:04:25.919 --> 00:04:37.199 So basically context is is is a really bad place to uh pass sensitive value or like large data and sub agents only push that problem a little further but still don't remove it. s60 00:04:37.199 --> 00:04:39.759 So what's the solution number two? s61 00:04:39.759 --> 00:04:42.320 Uh I think it was end of the last year. s62 00:04:42.320 --> 00:04:48.000 I think first uh Antropic and then cursor introduced something which is called progressive tool discovery. s63 00:04:48.000 --> 00:04:52.400 And uh I mean it's it's so simple that it hurts even that like people have to you know like do this. s64 00:04:52.400 --> 00:05:00.240 But uh the idea was like hey if we put all the tools into the context it just consumes too much tokens. s65 00:05:00.240 --> 00:05:04.560 So how about we put those tools into the context progressively only when you need them right? s66 00:05:04.560 --> 00:05:16.639 So for example, Antropic uh in cloth introduced this like tool just called like tool search tool and basically that tool helps you find other tools and only add those to the context when needed, right? s67 00:05:16.639 --> 00:05:17.680 And it kind of makes sense. s68 00:05:17.680 --> 00:05:23.280 I why would you like add all the 100 tools in your context all the time if you rarely need them? s69 00:05:23.280 --> 00:05:30.800 Like typically you just like need maybe one or two and actually this way you can save like huge amount of context right away. s70 00:05:30.800 --> 00:05:35.120 basically makes it more effective, cheaper, you know, faster to run and so on. s71 00:05:35.120 --> 00:05:36.560 So it kind of makes sense, right? s72 00:05:36.560 --> 00:05:38.160 So like every everybody should be doing it. s73 00:05:38.160 --> 00:05:45.120 No. And then solution number three also introduced uh end of last year by Cloudflare. s74 00:05:45.120 --> 00:05:46.400 It's called code mode, right? s75 00:05:46.400 --> 00:06:05.680 And the big idea is like hey let's not treat MCP tools as you know like functions like that you you know you would have to the context but treat them as as a code and actually models are really good at writing and and analyzing code because you know they can use tools like grab they can find the s76 00:06:05.680 --> 00:06:14.720 right you know definition like they can navigate the code pretty well and that way they can navigate also the two descriptions and information about the tools and so on. s77 00:06:14.720 --> 00:06:32.880 So it's it's a pretty like simple idea very straightforward that so let's convert tools MCP tools and servers into into code and treat it as code right pretty simple but extremely effective right because it it turns out models are better at writing and calling code code than than than s78 00:06:32.880 --> 00:06:48.080 calling tools because cool tool calling is an artificial construct basically that we have to teach the the LLMs to do it's like it doesn't exist like in a in a real world like in real training data like this those has to like syn syn synthetized and put into the model. s79 00:06:48.080 --> 00:06:53.840 But unfortunately uh even cloud implementation of code mode is like very difficult to use. s80 00:06:53.840 --> 00:06:57.680 It's like very like tied to their platform and it's it's not really straightforward you know. s81 00:06:57.680 --> 00:07:15.280 So not a lot of people are actually using code mode you know uh except few examples right so still most agents are living in the dark ages and this they don't they don't support these features you know and most mcp clients and it's a pity because the protocol has evolved a lot s82 00:07:15.280 --> 00:07:28.880 a lot of new things but the clients don't support it so how are CLI different you know why people compare CLI to MCP well actually agents never load the full CLI into context it's not like the agent would explore s83 00:07:28.880 --> 00:07:34.000 all the all the commands of the CLI tool like find a help and put it in the context. s84 00:07:34.000 --> 00:07:38.160 No, no, it does it it it's doing it progressively by default. s85 00:07:38.160 --> 00:07:49.599 So the agent uh is calling the CLI tool only when needed and sometimes actually it knows the tools by heart like so for example the basic Linux commands the agents already know because they have seen it training data s86 00:07:49.599 --> 00:08:02.080 since you know forever but also they can like learn about the tools uh the CLI tools from help because there is help and actually uh agents run CLI tools s87 00:08:02.080 --> 00:08:11.759 uh as code by default right so in order to run CLI tools you have to have some runtime like sand sandbox or machine that you can like run the commands in. s88 00:08:11.759 --> 00:08:20.479 So and when you like invoke these like CLI tools I mean it's code mode by by default you you run it as a code it's like bash code but it's still code s89 00:08:20.479 --> 00:08:30.800 so you know it's basically like code mode like uh is available to to CLI from scratch while MCP only had to like sort of like grow into it you know s90 00:08:30.800 --> 00:08:43.360 and this is important like agents know the shell by heart I mean Linux shell Linux or Unix has been around from like 1969 when uh Ken Thompson and Dis Richie actually created Linux s91 00:08:43.360 --> 00:09:03.440 and when you look like the basic terminal that black box of like 80 columns and I don't know 25 rows it's like really like condensed representation of what's happening in the computer right like like every bite every every character there is is optimized you know like over 40 years to really convey only the most important information right s92 00:09:03.440 --> 00:09:12.800 so it's really really like optimized you know for for people but it turns out like agents like this as And agents actually know shell by heart. s93 00:09:12.800 --> 00:09:16.560 They really know like how to do call call C cli commands, how to pipe them. s94 00:09:16.560 --> 00:09:20.720 They really know like how to use a shell well because they have seen it in so much training data. s95 00:09:20.720 --> 00:09:26.399 Plus for for the AI labs, you can actually synthesize like infinite amount of training data from using shell, right? s96 00:09:26.399 --> 00:09:33.600 You can just like pipe different commands together, explore the commands, extract information from your manual pages and basically feed this all to the models. s97 00:09:33.600 --> 00:09:36.320 So they know the shell really really well. s98 00:09:36.320 --> 00:09:41.120 But CLI are a local black box with no standard outend transport protocol. s99 00:09:41.120 --> 00:09:43.839 So basically I mean they are literally a black box. s100 00:09:43.839 --> 00:09:47.200 It's a black terminal thing right and you don't see what's happening inside. s101 00:09:47.200 --> 00:10:01.920 So for example if you wanted to like you know instrument like these like CLI uh in your enterprise you would have to sort of like in introspect the protocol and like uh oh are they using API or websockets or you know you don't know you cannot inject their credentials into that. s102 00:10:01.920 --> 00:10:07.440 So basically CLI are great for local interface but for remote access MCP is better. s103 00:10:07.440 --> 00:10:13.040 I mean I have yet to see uh some agent like using CLI connectors it doesn't exist because it doesn't make sense. s104 00:10:13.040 --> 00:10:14.560 You have MCP connectors. s105 00:10:14.560 --> 00:10:24.560 So how about we use MCP for standard remote access and CLI for local agent access or local agent interface with all the goodies like uh that I described. s106 00:10:24.560 --> 00:10:34.640 Well, that way you can provide all the protocol features of MCP through a simple tool call that all the agents already know which is called bash right. s107 00:10:34.640 --> 00:10:44.320 So basically the full complexity of MCP is hidden behind single tool call bash and you don't need to worry about the sessions authorization all out nothing. s108 00:10:44.320 --> 00:10:52.560 So uh ladies and gentlemen let me introduce you uh MCPC which is like our universal CLI client for MCP. s109 00:10:52.560 --> 00:11:07.760 uh we this started as a hobby project uh during the uh December so or also known as winter of cloth and u from a hobby project it actually like u you know evolved into probably the the most uh featurerich s110 00:11:07.760 --> 00:11:10.320 uh MCPC like cline on on the market. s111 00:11:10.320 --> 00:11:14.160 So you can install it like through npm or ban and how it works right. s112 00:11:14.160 --> 00:11:25.680 So the design goes for MCPC where hey we really want to support support like everything the MCP protocol has to has to offer you know task resources proons basically everything s113 00:11:25.680 --> 00:11:30.480 right and maximum compatibility to make it really easy to run anywhere anytime. s114 00:11:30.480 --> 00:11:32.079 MCP is like super lightweight. s115 00:11:32.079 --> 00:11:33.120 There is no LLM. s116 00:11:33.120 --> 00:11:34.959 It's just a a wrapper. s117 00:11:34.959 --> 00:11:40.720 It's a CLI wrapper over the MCP protocol that sort of like abstracts away the protocol thing but nothing else. s118 00:11:40.720 --> 00:11:51.760 It's very it's supposed to be easy to use for both agents and humans obviously and it needs to support code mode all the way and actually every command in MCPC s119 00:11:51.760 --> 00:11:59.760 has like a option to run with D-JSON which returns just like pure JSON representation of the data again MCP specification compliant. s120 00:11:59.760 --> 00:12:10.160 So actually you can compose the different like uh CLI calls together with tools like jq pipe them together and really build like a sequences of of of of tools s121 00:12:10.160 --> 00:12:12.720 uh of tool calls in code. s122 00:12:12.720 --> 00:12:19.040 So I'll show you how it looks like uh just a quick demo u we still have some time. s123 00:12:19.040 --> 00:12:22.880 So here is the basic interface of mcpc right. s124 00:12:22.880 --> 00:12:33.519 So it has help actually the help like we optimized a lot the help to kind of like make it really easy for agents to pick up right away without any external skills. s125 00:12:33.519 --> 00:12:34.480 Right. s126 00:12:34.480 --> 00:12:43.600 So uh first let me connect to uh a local server. s127 00:12:43.600 --> 00:12:51.200 Oh no no no. Uh so it supports stddio which is like the local processes uh running on your computer. s128 00:12:51.200 --> 00:13:08.480 So here I'm connect to MCP server called file system FS and when I connect I can suddenly like see the server name protocol capabilities tools and available commands right so for example I can get a list of the commands of the of the file system server s129 00:13:08.480 --> 00:13:20.560 I can get them get them in JSON format so again this is like something you can use as a code because it's JSON right and then I can also connect uh to remote MCP servers s130 00:13:20.560 --> 00:13:22.399 So first I need to uh login. s131 00:13:22.399 --> 00:13:26.000 So let's say appy mcp server. s132 00:13:27.120 --> 00:13:30.959 I it opens a browser where I run authentication. s133 00:13:30.959 --> 00:13:36.240 So just like pick some uh pick some account. s134 00:13:36.240 --> 00:13:44.399 So now uh the MCPC is is is logged in and basically it's it securely save the credentials into your local OS keychain. s135 00:13:44.399 --> 00:13:48.800 So we can now use them like uh to connect securely to to different resources. s136 00:13:48.800 --> 00:13:54.320 So uh let me connect to Appify MCP server. s137 00:13:54.959 --> 00:13:56.160 I hope this works. s138 00:13:56.160 --> 00:13:56.639 All right. s139 00:13:56.639 --> 00:14:00.240 So now I get like the basic information about the server including instructions. s140 00:14:00.240 --> 00:14:10.079 You know actually MCP protocol has instructions where server kind of explains what it does but most clients still don't support this basic primitive right which is kind of crazy. s141 00:14:10.079 --> 00:14:16.240 MCPC are supported of course then it can list the tools available commands and so on. s142 00:14:16.240 --> 00:14:21.519 So now when I run MCPC I see I have like three MCP sessions. s143 00:14:21.519 --> 00:14:23.600 Actually MCPC purses those sessions right? s144 00:14:23.600 --> 00:14:28.560 So basically if I go away it keeps those sessions alive and I can come back to them later right? s145 00:14:28.560 --> 00:14:42.240 So basically it keeps a state for you and actually you can just connect or set up once and then like let your code cloud code or codex use it without you sort of having to worry about it and sharing the configuration between these different agents. s146 00:14:42.240 --> 00:14:46.000 Basically the agent just just just use the same setup. s147 00:14:46.000 --> 00:14:49.760 And uh I was mentioning the progressive tool discovery, right? s148 00:14:49.760 --> 00:14:59.680 Uh so MCPC has this command called grap and I can for example like look for all tools or servers that contain the word find, right? s149 00:14:59.680 --> 00:15:12.720 And I see there is like the the the FS the fast system MCP server has like three tools that match this this uh this the search string and API has four right. s150 00:15:12.720 --> 00:15:23.279 So I can you know try another command for example I can try to run a tool called ampify regular browser which is like one of our tools to search web s151 00:15:23.279 --> 00:15:35.600 and actually MCP protocol has like this new feature called asynchronous task again most clients don't support it but here I just like add task and I run this task asynchronously that means like it starts on the server s152 00:15:35.600 --> 00:15:44.639 and I can like you know do locally other things you know and then just like pick up the results later and you know check the progress and so So we can see it's running s153 00:15:44.639 --> 00:15:49.680 in the background and then eventually I can get the results or I can actually detach you know and do something else. s154 00:15:49.680 --> 00:15:58.240 So again like MCPC is one of the few clients that supports uh it supports uh asynchronous task which is pretty cool and you can get it right away in your s155 00:15:58.240 --> 00:16:00.639 in your uh browser. s156 00:16:00.639 --> 00:16:09.440 Okay, I'll detach and for example okay I'm running out of time a little bit so I'll continue with the presentation there. s157 00:16:10.320 --> 00:16:29.040 So we did some demos here and actually you can use the the d-JSON command to uh return everything in JSON and then pipe these together to create like basically like like shell scripts that run MCB servers in the background without wasting your your your context tokens. s158 00:16:29.040 --> 00:16:37.040 By the way, recently we added support also for X4 X42 because it turns out uh there are not too many like tools to manage your local wallets. s159 00:16:37.040 --> 00:16:39.839 U so we added it to to MCPC as well. s160 00:16:39.839 --> 00:16:42.959 So because like we want we are we are supporting X42 now. s161 00:16:42.959 --> 00:16:47.680 So uh it's actually one of the coolest tools like you can use for X42 now. s162 00:16:47.680 --> 00:16:59.120 uh there are like special commands like for example proxies for sandboxing you know I will not go into detail there but also we wanted to to to understand what is the performance of the MCP s163 00:16:59.120 --> 00:17:10.720 MCPC and versus like native CLI right so we will we build this like new uh framework called connector evals and like typically like evals or you know this like benchmarks s164 00:17:10.720 --> 00:17:23.039 uh compare like different agents like hey for example like terminal bench is comparing whether codeex is better than cloud code and so But we sort of like flipped it and we built a framework to to compare like how different connectors s165 00:17:23.039 --> 00:17:24.799 compare with each other, right? s166 00:17:24.799 --> 00:17:30.080 Is it is it more effective to UCLI or MCP or MCPC or whatever tool. s167 00:17:30.080 --> 00:17:39.200 So here are a few results actually and you can see that like the this is I think uh these all these tests are using cloud code with sonet 5 s168 00:17:39.200 --> 00:17:51.440 and you can see that for example this is this chart the x x is is the time how long did it take to finish the task and y chart is the cost of tokens right and you can see that like mcpc s169 00:17:51.440 --> 00:18:00.240 and cli are actually performing pretty similarly while row mcp finish faster for whatever reason but actually consume more tokens. s170 00:18:00.240 --> 00:18:01.760 It's like another test. s171 00:18:01.760 --> 00:18:04.559 Again, MCPC and CLI are fairly comparable. s172 00:18:04.559 --> 00:18:07.840 Raw MCP like performs wars. s173 00:18:08.000 --> 00:18:10.640 And there's another one again similar result, right? s174 00:18:10.640 --> 00:18:13.280 So, actually these tests are still like fairly early. s175 00:18:13.280 --> 00:18:18.080 Uh but like uh yeah, be happy if you come and check it out and um contribute. s176 00:18:18.080 --> 00:18:22.320 So, please stop saying CLA is better than MCP because MCP plus CLI is the best. s177 00:18:22.320 --> 00:18:25.200 Thank you very much for your attention.