godspeed

a free, open-source ai tutor for your classes, taught by the ai you already pay for.

it's a desktop app. open this page on your computer to install.

works with claude code, codex, or any mcp app scroll no api keys · no subscription · open source

the same problem,
ten days apart.

these are my actual messages from cse 21, studying with claude code. i'd already written a tutoring skill so it would give hints instead of answers, and the replies were short and mostly fine. the problem was everything the chat didn't keep track of.

claude codecse 21 · hw2cse 21 · test 1 review +10 daysoct 12
I'm still lost on what to do here. I get that every room holds 0, 1, or 2 prints, and the total is 9, but I'm just confused about how to go use that information from there.

Let's take one small step at a time. Fill in this table. Each row asks: if this many rooms get 2 prints, how many rooms must get 1 print so the total is 9?

Rooms with 2Prints usedPrints leftRooms with 1
0???
1???
2???
wait so we are not doing subtracting the complement anymore?
new chat

a similar prints-in-rooms problem shows up while i'm reviewing for the test.

yeah i get the 1 room case but the 2 room one im lost by ur explanation

same exact spot as last time. the chat had no clue.

›
"could you stop using different examples? you're just confusing it more"
it kept switching explanations instead of finding the one step i was missing.
"p(a) is the 100% win rate and p(b) is 50%"
on a bayes problem it swapped my letters for its own, then admitted it later.
"you're asking me to write it out again even though we've already had this conversation before… i already did all the hard parts."
it didn't remember what i'd already worked out.

what godspeed keeps track of

where you're stuck

it checks what you know first, so it teaches the step you're actually missing.

one plan

the plan stays on screen and only changes when your answers say it should.

what you learned

every fact and every miss goes into your graph, and the next lesson starts from there.

how sure you were

you rate 1 to 5 on every answer, and topics you're overconfident on get flagged.

how it works.

while godspeed is open it runs an mcp server, and an agent you already use becomes the teacher. that same agent also does the background stuff, like research, diagrams, reading pdfs and building your knowledge graph.

no api keys or logins. there's nowhere to put one.

  • claude codeinstall the plugin, then /godspeed:teach
  • codexpaste one config snippet into config.toml
  • any mcp appconnect over streamable http with an address and a token

while you think, the claude code plugin just waits in a background script, so your agent spends zero tokens until you answer.

figures out where your understanding stops, from your answers, how sure you said you were, and what you typed while thinking.

claude code~/classes
>
>
>
0 / 14 Gradient as an expectation
✓ RightIs ĝ a random variable?

Gradient as an expectation

Freeze the current parameters x\mathbf{x}. The full gradient is

g=∇f(x)=1N∑i=1N∇ℓi(x).\mathbf{g} = \nabla f(\mathbf{x}) = \frac{1}{N}\sum_{i=1}^{N} \nabla \ell_i(\mathbf{x}).

Choose one example II uniformly and use its gradient as the estimate, g^=∇ℓI(x)\hat{\mathbf{g}} = \nabla \ell_I(\mathbf{x}). Each example has probability 1N\tfrac{1}{N}, so

E[g^]=∑i=1N1N∇ℓi(x)=g.\mathbb{E}[\hat{\mathbf{g}}] = \sum_{i=1}^{N} \frac{1}{N}\nabla \ell_i(\mathbf{x}) = \mathbf{g}.
Key idea

Unbiased means the estimate is right on average. It does not mean any single estimate is accurate.

11Minibatch noise

a map of
what you know.

the ai sorts what you learn into topics, colors each one by how often you get it right, and draws an arrow for every "learn this first."

mostly wrongmostly right
wrong → right
  • overconfidenttopics you get wrong after saying you were sure get flagged.
  • look up one ideahover a topic to see what it needs first and what it leads to.

built around
your classes.

every class gets its own lessons, knowledge graph, files and exams. lessons without a class just live in "no class."

files and foldersadd pdfs, notes, transcripts and recordings (transcribed on your computer), or link a folder and it rescans every minute.
examsadd exam dates and which files each one covers. the graph counts down the days and picks what to work on next.
overviewsee how much of each file you've covered, how each lesson went, and whether your confidence actually matches your results.
bayshorescomputer science at uc san diego, class of 2028

i read a lot of ai explanations.
it wasn’t really working.

i'm taking cse 21 (discrete math) and cse 29 (systems programming in c) this fall. i kept notes in obsidian and asked claude or chatgpt whenever something didn't click.

i even wrote a tutoring skill so they'd give hints instead of answers. the replies got better, but every chat started from zero and the same confusion kept coming back days later. my obsidian notes had a confidence field on every concept. i never filled in a single one.

learnharness by amos blomqvist already taught the way i wanted to learn. it checks what you know, plans a path, teaches one step at a time, then makes you practice. it connects with api keys or a chatgpt sign-in. but some agents can't be signed into at all. claude code closes that off, and it's the one i use most.

so i forked it and dropped keys and sign-ins completely. godspeed is an mcp server, so any agent that speaks mcp can teach, and every class gets its own files, graph and exams.

amosblomqvist/learnharnessthe teaching method. api keys or a chatgpt login
bayshores/godspeedforked oct 26, 2026. no logins, just mcp.

built in the open.

star on github 4
latest releaseOct 27, 2026

signed and notarized for mac. the app checks for updates once a day and installs them when you quit.

    ⎇ mainupdated today
      see everything on github
      aug 21779 commits since august, counting the learnharness work it's built on. the white bars are godspeed.today

      questions.

      picktap a star. the ones you've read turn green.

      0 of 7 read

      learn at godspeed.

      download for mac windows linux

      apple silicon, signed and notarized · windows setup.exe · linux appimage or .deb