read-only demo

Videos YBYUvGOuotE

MCP = Mega Context Problem - Matt Carey

index_state ready data_status ok

AI Engineer· published 2026-04-25· 0:22:41· en-US· indexed 2026-08-10 19:53

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:05, 1 of 1 keyframes kept
  2. Shot 1, 0:05 to 0:09, 1 of 1 keyframes kept
  3. Shot 2, 0:09 to 0:14, 1 of 1 keyframes kept
  4. Shot 3, 0:14 to 0:42, 1 of 1 keyframes kept
  5. Shot 4, 0:42 to 1:08, 1 of 1 keyframes kept
  6. Shot 5, 1:08 to 1:33, 0 of 1 keyframes kept
  7. Shot 6, 1:33 to 1:58, 1 of 1 keyframes kept
  8. Shot 7, 1:58 to 2:23, 0 of 1 keyframes kept
  9. Shot 8, 2:23 to 2:37, 0 of 1 keyframes kept
  10. Shot 9, 2:37 to 3:07, 1 of 1 keyframes kept
  11. Shot 10, 3:07 to 3:37, 1 of 1 keyframes kept
  12. Shot 11, 3:37 to 4:06, 0 of 1 keyframes kept
  13. Shot 12, 4:06 to 4:29, 1 of 1 keyframes kept
  14. Shot 13, 4:29 to 5:02, 0 of 1 keyframes kept
  15. Shot 14, 5:02 to 5:36, 0 of 1 keyframes kept
  16. Shot 15, 5:36 to 6:04, 0 of 1 keyframes kept
  17. Shot 16, 6:04 to 6:32, 0 of 1 keyframes kept
  18. Shot 17, 6:32 to 6:59, 0 of 1 keyframes kept
  19. Shot 18, 6:59 to 7:27, 0 of 1 keyframes kept
  20. Shot 19, 7:27 to 7:52, 0 of 1 keyframes kept
  21. Shot 20, 7:52 to 8:20, 0 of 1 keyframes kept
  22. Shot 21, 8:20 to 8:48, 0 of 1 keyframes kept
  23. Shot 22, 8:48 to 9:16, 1 of 1 keyframes kept
  24. Shot 23, 9:16 to 9:44, 0 of 1 keyframes kept
  25. Shot 24, 9:44 to 10:11, 0 of 1 keyframes kept
  26. Shot 25, 10:11 to 10:39, 1 of 1 keyframes kept
  27. Shot 26, 10:39 to 11:00, 0 of 1 keyframes kept
  28. Shot 27, 11:00 to 11:26, 0 of 1 keyframes kept
  29. Shot 28, 11:26 to 11:53, 1 of 1 keyframes kept
  30. Shot 29, 11:53 to 12:19, 0 of 1 keyframes kept
  31. Shot 30, 12:19 to 12:45, 0 of 1 keyframes kept
  32. Shot 31, 12:45 to 13:12, 0 of 1 keyframes kept
  33. Shot 32, 13:12 to 13:19, 1 of 1 keyframes kept
  34. Shot 33, 13:19 to 13:25, 1 of 1 keyframes kept
  35. Shot 34, 13:25 to 13:51, 0 of 1 keyframes kept
  36. Shot 35, 13:51 to 14:17, 0 of 1 keyframes kept
  37. Shot 36, 14:17 to 14:43, 0 of 1 keyframes kept
  38. Shot 37, 14:43 to 14:52, 0 of 1 keyframes kept
  39. Shot 38, 14:52 to 15:33, 0 of 1 keyframes kept
  40. Shot 39, 15:33 to 16:00, 0 of 1 keyframes kept
  41. Shot 40, 16:00 to 16:28, 0 of 1 keyframes kept
  42. Shot 41, 16:28 to 16:55, 0 of 1 keyframes kept
  43. Shot 42, 16:55 to 17:22, 0 of 1 keyframes kept
  44. Shot 43, 17:22 to 17:49, 0 of 1 keyframes kept
  45. Shot 44, 17:49 to 18:17, 0 of 1 keyframes kept
  46. Shot 45, 18:17 to 18:44, 0 of 1 keyframes kept
  47. Shot 46, 18:44 to 19:08, 0 of 1 keyframes kept
  48. Shot 47, 19:08 to 19:34, 0 of 1 keyframes kept
  49. Shot 48, 19:34 to 20:01, 0 of 1 keyframes kept
  50. Shot 49, 20:01 to 20:27, 0 of 1 keyframes kept
  51. Shot 50, 20:27 to 20:29, 1 of 1 keyframes kept
  52. Shot 51, 20:29 to 21:15, 0 of 1 keyframes kept
  53. Shot 52, 21:15 to 21:26, 1 of 1 keyframes kept
  54. Shot 53, 21:26 to 21:51, 0 of 1 keyframes kept
  55. Shot 54, 21:51 to 22:17, 1 of 1 keyframes kept
  56. Shot 55, 22:17 to 22:26, 0 of 1 keyframes kept
  57. Shot 56, 22:26 to 22:41, 1 of 1 keyframes kept

57 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
336
whisperx 336
chunks
40
from 336 cues
keyframes
18
kept of 57 captured
frames with text
18
378 lines read
chapters
0
from the source metadata
keyframe bytes
5.2 MB
word timings on 336 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done 2026-08-10 17:31 1m 35s
stt done 2026-08-10 17:33 25s
chunk done 2026-08-10 17:33 0s
text_embed done 2026-08-10 19:53 1s
keyframe done 2026-08-10 17:33 1m 49s
ocr done 2026-08-10 17:35 6s
frame_embed done 2026-08-10 19:53 3s

Frames, and what the machine read

  • 0:03 #0 done2 line(s)

    shot 0·sharpness 658.4

    1. AlEngineer0.98
    2. EUROPE1.00
  • 0:08 #1 done2 line(s)

    shot 1·sharpness 829.1

    1. PRESENTINGSPONSOR1.00
    2. Google DeepMind1.00
  • 0:13 #2 done3 line(s)

    shot 2·sharpness 907.7

    1. PLATINUM SPONSORS0.98
    2. # Braintrust0.96
    3. WorkOS OpenAI0.95
  • 0:20 #3 done8 line(s)

    shot 3·sharpness 566.2

    1. AlEngineer0.98
    2. Every APl is a0.99
    3. EUROPE1.00
    4. Tool for Agents0.97
    5. Matt Carey1.00
    6. EU/ACC1.00
    7. AlEngineer1.00
    8. EUROPE1.00
  • 1:00 #4 done10 line(s)

    shot 4·sharpness 2186.7

    1. If a dog wore pants would he wear them1.00
    2. like this1.00
    3. or1.00
    4. like this?1.00
    5. AIE1.00
    6. 1.00
    7. 1.00
    8. Engineering the future of Al1.00
    9. AlEngineer1.00
    10. 20260.98
  • 1:11 #5 skipped

    shot 5·duplicate of #4

  • 1:48 #6 done26 line(s)

    shot 6·sharpness 1705.5

    1. From Bundled to Shared0.99
    2. AIE1.00
    3. BEFORE MCP0.98
    4. REMOTEMCP1.00
    5. App A0.99
    6. App B1.00
    7. 1.00
    8. 1.00
    9. 1.00
    10. get_woather0.97
    11. get_woather0.95
    12. Agent B0.97
    13. Agent C0.97
    14. send_email0.98
    15. send_omail0.93
    16. MCP Server -80 tokens0.92
    17. Every app reimplements the same integrations1.00
    18. got_weather0.91
    19. send_email0.94
    20. get_calendar1.00
    21. "Share tools with agents you've never met*0.98
    22. 4/211.00
    23. AlEngineer0.97
    24. AlEngineer1.00
    25. EUROPE1.00
    26. 20260.99
  • 2:01 #7 skipped

    shot 7·duplicate of #6

  • 2:24 #8 skipped

    shot 8·duplicate of #3

  • 2:46 #9 done41 line(s)

    shot 9·sharpness 1751.0

    1. From Bundled to Shared1.00
    2. BEFORE MCP1.00
    3. REMOTE MCP1.00
    4. App A0.98
    5. App B0.91
    6. got_weather0.96
    7. got_woather0.96
    8. Agent A0.94
    9. Agent C0.99
    10. send_onail0.95
    11. send_onail0.95
    12. AIE1.00
    13. 1.00
    14. 1.00
    15. Every app reimplements the same integrations1.00
    16. get_weather0.96
    17. MCP Server1.00
    18. send_email1.00
    19. -200K tokens0.99
    20. get_calendar1.00
    21. list_users0.96
    22. delete_user1.00
    23. 1.00
    24. 1.00
    25. 1.00
    26. 1.00
    27. delete_dns_record0.97
    28. kv_put0.99
    29. List_r2_buckets0.98
    30. r2_get_object1.00
    31. r2_put_object1.00
    32. create_dl_database1.00
    33. List_dl_databases0.98
    34. get_zone0.95
    35. List_zones0.99
    36. purge_cache1.00
    37. get_analytics1.00
    38. "Share tools with agents you've never met*0.95
    39. Engineering the future of Al0.99
    40. AlEngineer1.00
    41. 20260.99
  • 3:33 #10 done18 line(s)

    shot 10·sharpness 1809.0

    1. We Filled the Context Window1.00
    2. 2.3M1.00
    3. AIE1.00
    4. tokens in Cloudflare's OpenAPI spec0.99
    5. 1.00
    6. 1.00
    7. CONTEXT WINDOW0.97
    8. 128K tokens1.00
    9. Tool Definitions1.00
    10. Tool Definitions 1 17H0.96
    11. System Prompt 2K0.98
    12. Conversation1.00
    13. Actual Task1.00
    14. The context limit is not an MCP problem. It's an Agent problem.0.96
    15. 5/210.99
    16. Engineering the future of Al1.00
    17. AlEngineer0.99
    18. 20260.93
  • 3:54 #11 skipped

    shot 11·duplicate of #6

  • 4:26 #12 done6 line(s)

    shot 12·sharpness 519.6

    1. Split by D1.00
    2. AlEngineer0.98
    3. EUROPE1.00
    4. EU/ACC1.00
    5. AlEngineer0.99
    6. EUROPE1.00
  • 4:55 #13 skipped

    shot 13·duplicate of #9

  • 5:22 #14 skipped

    shot 14·duplicate of #9

  • 5:53 #15 skipped

    shot 15·duplicate of #6

  • 6:07 #16 skipped

    shot 16·duplicate of #6

  • 6:56 #17 skipped

    shot 17·duplicate of #9

  • 7:08 #18 skipped

    shot 18·duplicate of #9

  • 7:42 #19 skipped

    shot 19·duplicate of #3

  • 8:14 #20 skipped

    shot 20·duplicate of #9

  • 8:23 #21 skipped

    shot 21·duplicate of #9

  • 9:12 #22 done10 line(s)

    shot 22·sharpness 1833.9

    1. But Clients Didn't Implement1.00
    2. AIE1.00
    3. Code Mode Quickly1.00
    4. 1.00
    5. People still couldn't use our API.1.00
    6. Why not?1.00
    7. AlEngineer0.97
    8. AlEngineer1.00
    9. EUROPE1.00
    10. 20261.00
  • 9:22 #23 skipped

    shot 23·duplicate of #22

Transcript

336 cues· 4,098 words· 21,454 chars

  1. 0:16 Hello, everyone.
  2. 0:17 Welcome.
  3. 0:19 Quiet down, quiet down.
  4. 0:20 Hi.
  5. 0:22 Awesome.
  6. 0:27 How is everyone?
  7. 0:28 Yeah, good?
  8. 0:31 Thanks.
  9. 0:33 Want to hear some MCP versus CLI debates?
  10. 0:37 Is that why you all came?
  11. 0:38 Anyway.
  12. 0:39 Hello, my name is Matt.
  13. 0:40 I work on MCP and agents at Cloudflare.
  14. 0:44 Welcome to my talk.
  15. 0:45 It's all about how we can make every API a tool for agents.
  16. 0:48 APIs exist in the wild.
  17. 0:49 How can we connect them to agents and make them do things?
  18. 0:54 So I really love my job because every day I get to decide like if an agent looked like this, would he do this or would he do this?
  19. 1:01 I think it's kind of fun.
  20. 1:03 And we often fluctuate between the two of them.
  21. 1:06 Someone does something that you think is slightly funny and then six months later we're all doing it and claiming it was the best thing in the world.
  22. 1:12 So yeah, it's really good crack.
  23. 1:14 But the main part of the role, I guess, and what I end up doing day to day is how do we give agents hands?
  24. 1:21 How do we let them interact with the outside world?
  25. 1:23 And you're probably familiar with something like this.
  26. 1:26 This is tool calling, function calling.
  27. 1:28 It's been around for a while now.
  28. 1:29 The LLM writes a function.
  29. 1:31 You execute the function.
  30. 1:33 Bash, bash, bash, the weather in London is 18 degrees, it's not, it's like eight and it's freezing.
  31. 1:39 Sad times.
  32. 1:40 And then from there we went from bundled tools to something like shared tools, like people made tools in their agents.
  33. 1:46 And you probably, this is like all like recent history, so everyone's probably aware of this, but before MCP we had like, people would bundle all their tools in their agents and then they would keep them bundled in their agents and then if I was like trying to interact with Gmail or something, I would make loads of tools for Gmail, bundle them with my agents,
  34. 2:03 And that would be it.
  35. 2:03 And the next person would have to do exactly the same thing.
  36. 2:06 And then we ended up with this big explosion of MCP and remote MCP about April last year.
  37. 2:11 And the service providers were like, oh, we can like...
  38. 2:16 give everyone MCP tools, and then everyone can use the same standardized tools, and we just make it once, and we provide it as another surface for people to consume our API.
  39. 2:24 Maybe there's a CLI, there's an API, maybe there's like, I don't know, GraphQL API, and there's now MCP as like another surface.
  40. 2:33 But this got a little bit fun because it was okay with like eight tools, but then what happens if you added like a few more?
  41. 2:41 Or a few more?
  42. 2:43 Or a few more?
  43. 2:45 Or a few more?
  44. 2:47 Or a few more.
  45. 2:48 And now you're like, I wanna give an agent access to our whole API surface.
  46. 2:53 And, well, that ain't gonna happen.
  47. 2:56 Why is it not gonna happen?
  48. 2:57 You've exploded a context window of the agent.
  49. 2:59 You've completely annihilated it.
  50. 3:01 This is one point something million tokens.

Open at this second