シャドーイング練習: Running AI Agents in Secure Sandboxes with E2B & Docker MCP | Docker’s AI Guide to the Galaxy - 動画で英語スピーキングを学ぶ

Hi there, welcome to the next episode of AI Guide to the Galaxy.

My name is Olek and I work with the DevRel team at Docker.

Today we are in the real actual studio.

We are here at the Docker Captain Summit in Istanbul.

And it is my honor to introduce my next guest, who is a Docker captain, Jonas from E2B.

Hi, how are you doing?

Good, thank you.

It's great to be here.

It's been two days of very productive conversations about Docker, Docker products, roadmaps

and we're taking this chance to record this video face -to -face.

We'll see how it goes.

Leave it a comment if you think this is a great idea and we should do it more often.

So, recently, E2B and Docker announced this partnership.

And, well, our viewers mostly are familiar with Docker and Docker products.

So maybe, can you tell me a little bit about what E2B is and what do you offer?

So, E2B is an open -source infrastructure for AI sandboxes.

So if you want to run agents

that need access to like a full operating system and you want to do that securely and sandbox and isolate it, that's what E2B is for.

Right.

Sandboxes are a very hot topic nowadays.

Everyone is doing them from Cloudflare to all the coding assistant

shipping some sort of guardrails in form of seed belt or mangos and stuff like that.

So E2B, is E2B like a local sandbox, or is it the cloud machines?

It's cloud -based, and it's full virtual machines.

So the agent can get full access to a full Linux VM, and it's all in the cloud.

So you can also scale up like thousands of sandboxes in front of there.

That is awesome.

I can imagine the effort that went into creating the infrastructure of that.

Definitely not easy.

This is very, very cool.

So ETV provides the sandboxes and, well, Docker has a number of initiatives.

Most notably for this partnership was the MCP catalog,

the trusted catalog of MCP servers that you can plug and use from your AI assistants.

And maybe the supplementing technologies like MCP toolkit and the MCP gateway.

So can you maybe explain the partnership and what the integration is in a few words?

Yeah.

So with a partnership, Docker has this amazing catalog of like pre -configured MCP servers with like all the configuration you need.

And with the partnership, we bring the MCP servers that are already trusted and in the catalog to E2B sandboxes.

So when you run an agent in a sandbox, the agent has like automatic access to all of those 200 MCP servers.

Right.

Awesome.

When you say when you run a sandbox, what does it mean?

Can you maybe show an example?

How do I, as a developer or as a person running my agent, how do you create the sandboxes?

What's the experience look like?

So usually you would create a sandbox through an SDK that's either JavaScript or Python.

And then when you create a sandbox, which is just sandbox .create here, you can configure a few different MCP servers.

So here, this is fully typed.

You could, for example, do EXA, which is a search engine kind of, And then also the configuration is typed and you can provide,

for example, the API key for that.

Right.

So Docker has this catalog of MCP servers,

right, which are technically just Docker images with software that is an MCP server running inside them.

And there is a, like, surrounded technology, which is in case of headless machines is the MCP gateway Docker image

that acts as a configuration entry point in the proxy.

to the MCP servers.

So when I provide this configuration, so this is the API to create sandboxes with E2B.

When I provide this configuration, the library, the E2B library will spin up the gateway within the sandbox.

Yes, that is correct.

Is it like completely within the sandbox or is it an external...?

Yeah, the gateway runs completely in the sandbox, so isolated in that.

And then the MCP servers also run in the same sandbox.

Okay.

And the gateway is also accessible to the outside if you want, like authenticated, of course.

So if you have an agent running locally, it could use the isolated MCP servers in a sandbox as well.

Right.

Okadoke.

So let's look at this.

Walk me through this example more.

So we configure in sandbox and we create this.

So that will send the commands to the cloud ETB premises

and give me a virtual machine and the connection to

that right yeah and this because we configured mcp this will also automatically start the gateway

and like pre -pull the docker images for example for those mcp servers this is very interesting actually

so the docker catalog is just a number of um mcp catalog sorry those are docker images

and then uh they don't have any particular configuration externally

because it's the catalog resides and is sort of managed by the toolkit

or the mcp gateway uh

which well you configure them either from the user experiences like form

or using environment properties

and docker secrets uh in the in the in the catalog for in the gateway form like

when you just turn on containers uh how do you get the api like

that did you create it manually or uh so we take the catalog

which is like a huge yml file and then auto generate the types of

that and then in the background we map it back to the catalog

so what actually gets used in the end is the catalog right

but the catalog is quite complex

and like in docker desktop it's also mapped to like a nicer ui

and here we just map it to a different oh

that is awesome so but it's a one -to -one mapping in the end right

so like you get the api for both javascript and python uh

sort of for the price of one yeah yeah

so we just take the catalog and then auto generate everything

that is awesome okay and i can see you put entropic api key

so we're gonna use cloud yes

so here we just run cloud like the actual cloud code inside the sandbox

and we add the mcp gateway here the first command so

when you do sandbox .command .run this actually runs like a command in the sandbox

sort of like docker exec yeah but in the for the sandbox okay um

100

and then we can start i need to wrap here um we then start cloud with the task here

101

and the yolo mode because we're in a sandbox and we don't really, yes.

102

It cannot mess up our machine.

103

And in this case, we have a very simple use case where we just search for interesting research papers

104

and then summarize it and turn it into, like, a wide -coded summary in, like, an HTML page.

105

And then, because, again, we're in the sandbox, we can really do stuff we wouldn't do on our laptop.

106

We can expose it as an HTTP server here and, like, host a website.

107

I can run that.

108

so i hope that actually works

109

so this will probably take however long cloud decides to do its research right okay

110

so it created the mcp gateway yes and it prints the url for us

111

so yes so so for example

112

if you have cloud desktop running you could also connect your

113

cloud desktop to the a remote sandbox with the gateway running inside of it.

114

Right, right.

115

I will need to provide the authorization token, but like I can create, come on, connect to that from external things.

116

Okay, so now what is happening?

117

So now Claude is processing our request and we configured the MCP gateway.

118

So it has access to tools from the servers that we configured, which is the Archive and...

119

Archive and DuckDuckGo.

120

DuckDuckGo just to have a sad engine really right

121

and then it will use those servers

122

and generate on us some information in the index hml page

123

and then we can access that

124

because we spin up um yeah interactively a python server yeah and that's of course something you would not really

125

usually want to do this is the demo purposes uh demo so uh you would probably make it

126

make your clod write it to the database

127

and then you access it securely or put the files into your s3 bucket

128

or anywhere else okay this takes a little bit of time uh can you tell me

129

so here you have clod within the sandbox yes do you package clod by default

130

or is it something that uh just accidentally works here yeah uh

131

so when you create so sandboxes are created from templates

132

and templates are really like the root file system of a of the virtual machine

133

so you can create a new template

134

and then install whatever you want in there

135

so this is example code of how you create a template you can start from the mcp template

136

which already has the gateway installed

137

and a few tools like npm python is already installed in there and then you can extend it however you want.

138

So we have, for example, apt install.

139

So if there's, for example, if you want to install curl, you would just add curl here.

140

Or if you want to do, if you want to clone a certain repository, you could do that here as well.

141

And this is a template which is not the actual instance of the sandbox, but this is the file system.

142

Yes sort of projection yes so it will be used yeah we're starting the sandbox

143

and each template has a name and then

144

when you create the actual sandbox you just reference that name

145

and it's based on that so in this example everything would have curl installed

146

or if you do the right so in sort of

147

an analog in the docker universe terms this would be like

148

sort of building a docker image yes except specifying the entry point

149

because the sandbox entry point is probably controlled by correct yeah okay

150

and there's also start commands which is kind of like an entry point

151

so if you have a service

152

that you want to start automatically you can also do

153

that okay but this is very similar to docker image

154

and we even have support to use docker images as templates oh you can create an etb .docker file usually

155

and then this is converted to a template as well do you like the the api

156

or docker files more for creating templates the api

157

because it's better for like programmatic access So if you have a use case where you provide coding agents for your users,

158

for example, programmatic is usually the better way.

159

Yeah, I agree.

160

I agree.

161

I have experience with test containers.

162

So it's a very similar story there, which it's an open source library,

163

and they expose programmatic API to configure and manage containers within Docker context.

164

So I can totally see the appeal of running the API.

165

Right, so we have the results from Claude.

166

So you can see it found an interesting paper, Foundations of Large Language Models.

167

Okay.

168

And then also did some research about the authors.

169

And then hopefully if it worked, published it at this URL.

170

I think it says sandbox closed successfully.

171

Yeah, okay.

172

We were too slow.

173

The sandbox had a lifetime of five minutes and we talked too long.

174

So it's already gone.

175

uh but i can run it again

176

if you want no you don't have let's let's run let

177

okay let's show it we're gonna edit it in post yeah uh no uh

178

so i believe you

179

that it works i actually have seen it work i've built

180

a couple of examples uh i've mostly tried python myself to figure out pieces

181

which you know because you help me debug issues there but

182

so i i know that it works

183

and i know you can expose those uh sandboxes uh over the internet

184

if you want to like interactively sort of look uh for the demos

185

so you said you're not really you won't do this um during the normal

186

like usage of sandboxes um well like

187

if you can share some details like what is the common patterns how people use EGB sandboxes.

188

Is it common to package AI systems within the sandbox

189

and run it as a sort of isolated cloud machine for the unit of computation

190

or do people connect from external AI systems?

191

What do you see happening and also what do you personally think is a more interesting way going forward?

192

So I think one of the most common use cases is like the true coding agents

193

that you want to run completely in the background

194

so for example if you have like if you've seen cursor backbot which like

195

partially runs code or like rabbit which like um does pull request reviews

196

and like fixes code already um then usually like the agents run actually completely in the sandbox

197

just to have it fully isolated and

198

but of course there's also benefits to do it like partially in the sandbox partially outside

199

because um well it's it's a bit easier to orchestrate probably

200

if you run partially outside um

201

and i didn't quite understand what you meant with the no interesting so

202

if you look at the home page for etb yeah uh

203

well it's like there is a lot of examples uh here for technologies

204

that you can connect if you scroll a little bit down right

205

so we have uh there were some here for example right

206

and i understand that those could be synthetic examples but

207

if you click on the open ai for example right

208

so here the

209

which i assume is a classic example you run open ai externally in your in your python code

210

and then you connect it to sandbox just to run the code oh yeah

211

that is a very common use case as well for example

212

a lot of the chat applications have some way of running code right like

213

if you give it a csv file you like figure out

214

how many columns i have in there they write python code

215

and then actually run the python code

216

and then of course only like the actual code execution is in the sandbox right

217

and very short -lived usually and this is like a major use case Um, and you can, yeah, I, yeah.

218

Right.

219

So what I hear is there is no particular preference, uh, from the sort of, uh, user point of view, like if they want to connect external clients, it's absolutely fine.

220

If they want to put it into the sandboxes and have it to run that more like independent, longer, uh, end -to -end processes, that is also all good

221

and fine i think it also depends on who what do you want to isolate

222

and what do you want to like protect against

223

because in the end like the sandboxes provide a lot of

224

isolation right yeah this is very very cool i'm very very

225

happy uh with the was was how this partnership turned out it's really easy to configure the mcp servers um

226

so i know that docker uh mcp toolkit recently added uh the remote server abilities

227

so like if you can connect to for example notion uh server

228

and from idea token uh but the notion server is not run locally in the container

229

but it just connects to notion api uh would

230

that work with this uh with the integration like this as well

231

that is already supported

232

so you can already for example notion is a remote one uh you can connect it like this

233

but for now mostly pre -configured ones so those that are actually in the catalog you can use

234

right okay okay we can also now see the actual demo

235

result okay it's just real oh no we went too fast

236

you probably need to increase the time out somewhere in the code just say it say

237

that like run it like for 10 minutes or something yeah right okay

238

so you can do remote uh a lot of mcp servers

239

uh are created i know with the sort of idea

240

that their end users will be uh will be individual uh individuals who is their particular data access rights

241

and there's this whole OAuth integration story for mcp servers uh how does

242

that work or i assume there is not it's not important for programmatically created tasks um yeah

243

so how does it

244

so right now oauth doesn't work in the mcp gateway with

245

e2b right i think docker has already some auth support for like two

246

or three mcp servers as far as i know like a very little subset um

247

but i do see that oauth is

248

very important

249

and then we will have to figure it out how also

250

it fits into like the programmatic access use case yeah yeah

251

and i don't think we can ignore it just

252

because it's like a programmatic use case um

253

because in the end this is like user data and the user needs to give access to that

254

um for now it's just api keys yes so like in the docker toolkit there are some servers like for For example,

255

Github that you can configure with OAuth, right?

256

You can also configure it with the usual token, which is much, much simpler and maybe offers you more control for the integrations, because in the OAuth case,

257

you really need to have a user in the loop and they need to review the OAuth consent screen and like,

258

it makes no sense to show a consent screen if there is no end user present in the loop.

259

You might as well go with the token

260

and then the predefined rights uh

261

so i see i like i honestly assume allows is not

262

the most important feature for for the programmatic use cases like this um

263

but specifically with remote servers most of them have some sort of allows i just realized

264

that we have a sandbox skill in the demo so it would never work um

265

the fourth trial see this is an example how easy you can change your code

266

that runs this integration and how convenient it is to work with e2b sandboxes that actually

267

if you think about it it's quite a sophisticated piece of

268

turtles all the way down we run some javascript code

269

and it creates a virtual machine and then inside

270

that virtual machine it runs docker engine i assume

271

because it needs to run containers and it runs a container for mcp gateway

272

and it needs to run the containers for individual mcp servers uh and all of that surprisingly works

273

and at the same time surprisingly fast no it's actually pretty

274

fast spinning up a container is fast i assume you have

275

a lot of effort put into optimizing the startup for the sandboxes uh specifically

276

because you mentioned the short -lived uh interactions with them as well um

277

so this is all glorious

278

and nice um uh are there any plans to work on

279

other uh programming languages to support uh in the api

280

or is javascript and python sort of the largest chunk of the ecosystem

281

that you are aiming at currently i i'm this is not

282

really my field like i i don't really work on the sdks but

283

in ai right now i think javascript

284

and python probably make up 99 of all use cases i would love to see go personally

285

and i know that some customers actually build their own sdks for

286

that yeah you have the sort of a rest api

287

or something yeah you can you can use the api

288

and yeah yeah the sdk is open source

289

so you can also just like reverse engineer the sdk it's except for like the template template system it's not

290

that complicated yeah

291

so it would be fairly easy to for the basic operations

292

like creating a sandbox you could just use the api this

293

is an exercise for the curious watcher dropped the script from

294

the javascript sdk also the entire um project is open source yeah

295

so there's also like the open api specs for example right

296

can i can i run the backend myself on my premises

297

or yes you can it's it's not super easy to save for us because it's quite a lot of

298

stuff involved to make it work um so it's not like a docker run and then

299

i i think it's just the log file now oh yes okay um

300

so we've got our vibe coded landing page here well

301

that is quite awesome so you get the summary

302

and you get the information about the authors and

303

so the information about the authors what what the what was that was using the

304

search engine yeah this is awesome this is very easy to

305

use you can you can all i can ultimate those tasks i can put them in github actions

306

and the only and the only thing

307

that i need to configure to get access to the sandbox

308

i just need the etb api key correct yeah uh

309

and they will npm library or like a python library yeah okay this is awesome uh

310

what is your favorite mcp server oh the catalog because you said

311

that you're sort of parse the situation uh

312

or what's your least favorite okay what is what is one

313

that you hit with the most passion okay i need i need to think for

314

that i think my most favorite one

315

is probably dr go because it doesn't need an api key

316

and i used it so much for testing okay um so

317

and it always works like i never had a case where

318

i mean it's a very simple mcp server there's one tool

319

but it does its job

320

so i love it perfect for testing yeah it's it's it's a perfect

321

and i've probably used it like a thousand times in the

322

last week the one i hate the most you don't have

323

to answer we don't have to go on camera with this you can just tell me

324

that it's at last and later oh actually it's it's probably the ones

325

that need the most configuration yeah where there could be same defaults like i don't have a specific one

326

but sometimes it's like you need 20 different fields you need to configure

327

but you actually only need two

328

and i think that's a pain yes yes i think so

329

which is actually frankly that is the case i think with the atlassian mcp server

330

that we have in the catalog

331

because you need to have the key for maybe confluence

332

and then jira and some urls for both of them

333

and some additional configuration uh it's not

334

so easy from the lick uh it's probably very convenient

335

when you need to use it in anger in real life but like

336

if you need to create a demo yeah uh you will you kind of stumble on this uh

337

which is a minor use case it's not it's not actually like an actual complaint uh

338

and sorry the docker hub mcp server i'm not just saying

339

that because i'm a docker captain

340

but i've used it a lot um to find documentation about other docker images right it's really good

341

because you get the like true source of the documentation

342

and not if use web search like you get a lot of wrong information and on dock hub it's usually right

343

so yeah docker hub could be the source of information right very very good uh

344

so how do people get started with e2b

345

and maybe docker mcp catalog on e2b sandboxes

346

and where can they maybe meet the community

347

and ask questions to how to use this integration where do they go

348

so if you want to get started with E2B in general, there's a very generous free tier.

349

And if you want to try out MCP, there's a quick start guide with TypeScript and Python symbols that you can run just like that.

350

And we also have a lot of examples in our cookbook.

351

So for example, if you want to run cloud code or research agents or browser base for like web automation,

352

like we have a lot of examples for that right and

353

if you have questions there's a e2b discord

354

that you can join i think that should also be referenced somewhere somewhere on the e2b page

355

i'm sure it's uh somewhere there we're gonna find it

356

and we're gonna put the links into description of the videos

357

so header uh no thank you so much for joining us

358

and explaining how it looks and what is it to be

359

and how to use the sandboxes it was a pleasure stay tuned for the next episodes

360

and please check out the etub and docker partnership

361

and what you can do build uh running your agent securely in isolated sandboxes,

362

um, with enabled by default YOLO mode, which is fun thing to say.

363

Thank you so much.

364

See you there.

✨ おすすめ動画

78 文

このレッスンの語彙とスピーキングのポイント

このC1レベルのスピーキングレッスンは、動画「Running AI Agents in Secure Sandboxes with E2B & Docker MCP」を教材にしています。繰り返し出てくる語は次のとおりです：docker, mcp, sandbox, server, example。この動画には、シャドーイング用の文が364文、単語が4291語あります。音声の長さは25:17です。話す速さは1分あたり約170語で、日常会話に近い自然なペースです。英語の頻出3,000語に含まれる単語は79%だけなので、語彙は難しめです。

この動画の重要語彙

動画の中で特に難しい単語15語を、発音と意味つきで紹介します。

単語	発音	意味
sandbox 名詞	/ˈsæn(d).bɑks/	砂場, 砂箱
catalogue 名詞	/ˈkæt.əˌlɔɡ/	カタログ, 目録
configure 動詞	/kənˈfɪɡə(ɹ)/	設定する
template 名詞	/ˈtɛm.plɪt/	型板, 雛形
configuration 名詞	/kənˌfɪɡ.əˈɹeɪ̯.ʃən/	構成, 配置
container 名詞	/kənˈteɪ.nəɹ/	容器
isolate 動詞	/ˈaɪ.sə.leɪt/	切り離す, 隔絶する
demo 名詞	/ˈdɛm.oʊ/	デモ
toolkit 名詞	/ˈtulˌkɪt/	ツールキット
token 名詞	/ˈtoʊkən/	トークン
curl 名詞	/kɜɹl/	巻き毛
desktop 名詞	/ˈdɛsktɒp/	机の上
expose 動詞	/ɪkˈspoʊz/	晒す, 暴く
authorization 名詞	/ˌɔːθəɹaɪˈzeɪʃən/	許可, 認可
premise 名詞	/ˈpɹɛm.ɪs/	前提

動画に出てくる句動詞

単語	意味
check out 動詞	試験する, 検査する
try out 動詞	試す, 試してみる

繰り返し練習したい文

動画に出てくる短く完結した文です。日常会話でそのまま使えます。

Hi, how are you doing?
We'll see how it goes.
When you say when you run a sandbox, what does it mean?
Can you maybe show an example?
What's the experience look like?

この動画の文法

話し手がよく使っている文型を、動画の実際の表現とともに紹介します。

文型	動画での表現
受動態 be + 過去分詞 — 誰がするかより、何が起きるかに焦点を当てる	are already trusted · is typed · are created
関係詞節 who / which + 節 — 人や物について情報を加える	sandbox, which is · EXA, which is · right, which are
現在完了形 have/has + 過去分詞 — 過去の出来事が今も関係している	It's been · have let · have seen

注意したい発音

話し手はdon't, we're, I'mなど、短縮形や弱形を39回使っています。聞こえたとおりの短い形で発音しましょう。

「th」の音: authorization /ˌɔːθəɹaɪˈzeɪʃən/, authenticate /ɔˈθɛn.tɪ.keɪt/, synthetic /sɪnˈθɛtɪk/
「sh」と「zh」の音: configuration /kənˌfɪɡ.əˈɹeɪ̯.ʃən/, authorization /ˌɔːθəɹaɪˈzeɪʃən/, documentation /ˌdɑkjəmɛnˈteɪʃən/, computation /ˌkɑmpjʊˈteɪʃn̩/, projection /pɹəˈd͡ʒɛkʃən/
長い単語（アクセントの位置に注意）: configuration /kənˌfɪɡ.əˈɹeɪ̯.ʃən/, programmatic /ˌpɹəʊ.ɡɹəˈmæt.ɪk/, authorization /ˌɔːθəɹaɪˈzeɪʃən/, automobile /ˈɔː.tə.məˌbil/, documentation /ˌdɑkjəmɛnˈteɪʃən/

日本語話者が苦手な音：

/r/ と /l/ の区別 — /l/ は舌先を歯茎につけ、/r/ はどこにもつけない: securely /səˈkjʊɹ.li/, curl /kɜɹl/, surprisingly /səˈpɹaɪzɪŋli/, guardrail /ˈɡɑɹdˌɹeɪl/, programmatically /pɹoʊɡɹəˈmætɪk(ə)li/
/v/ — /b/ にならないように、上の歯を下唇に当てる: viewer /ˈvju.ɚ/, vibe /vaɪb/
/f/ — 「フ」ではなく、上の歯と下唇で出す: configure /kənˈfɪɡə(ɹ)/, configuration /kənˌfɪɡ.əˈɹeɪ̯.ʃən/, confluence /ˈkɑnˌflu.əns/, specify /ˈspɛs.əˌfaɪ/, frankly /fɹæŋkli/

この動画での練習方法

まず声を出さずに動画を最後まで聞き、知らない単語をメモします。
まず0.75倍速で一文ずつシャドーイングし、慣れてきたら通常の速度に戻します。
自分の声を録音して元の音声と比べます。sandbox, catalogue, configureなどの単語に特に注意しましょう。

シャドーイングとは？英語上達に効果的な理由

シャドーイング（Shadowing）は、もともとプロの通訳者養成プログラムで開発された言語学習法で、多言語習得者として知られるDr. Alexander Arguelles によって広く普及されました。方法はシンプルですが非常に効果的：ネイティブスピーカーの英語を聞きながら、1〜2秒の遅延で声に出してすぐに繰り返す——まるで「影（shadow）」のように話者を追いかけます。文法ドリルや受動的なリスニングと異なり、シャドーイングは脳と口の筋肉が同時にリアルタイムで英語を処理・再現することを強制します。研究により、発音精度、抑揚、リズム、連音、リスニング力、そして会話の流暢さが大幅に向上することが確認されています。IELTSスピーキング対策や自然な英語コミュニケーションを目指す方に特におすすめです。

シャドーイングのやり方: ステップ別の完全ガイドを読む →

シャドーイング練習: Running AI Agents in Secure Sandboxes with E2B & Docker MCP | Docker’s AI Guide to the Galaxy - 動画で英語スピーキングを学ぶ

How I deploy serverless containers for free

What AI can (and can't) currently do in a real lab | Jennifer Doudna

OpenAI’s lead engineer was secretly a spy...

Understand Microsoft Entra ID | AZ-104 | Episode 2

Armchair Architects: Resiliency in the Cloud – Core Patterns

What is Docker? Docker container concept explained || Docker Tutorial 1

Cleo Abram Answers The Web's Most Searched Questions | WIRED

GitHub Copilot now controls your command line...

Amazon Athena explained in 120 seconds💡

NVIDIA GTC 2025 - Keynote Highlights by NVIDIA CEO Jensen Huang

Most devs don't understand what agents are

JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro)

このレッスンの語彙とスピーキングのポイント

この動画の重要語彙

動画に出てくる句動詞

繰り返し練習したい文

この動画の文法

注意したい発音

この動画での練習方法

シャドーイングとは？英語上達に効果的な理由