The Context Budget Index · November 2026 · updated monthly · counted with ctxcap
The Context Budget Index
What the most-starred repositories on GitHub spend on agent instructions. An AGENTS.md or CLAUDE.md is read by a coding agent at the start of every session. Whatever it weighs is spent again each time, before the agent has looked at a line of code. This page counts that weight.
A named index, rebuilt once a month. The address above always shows the newest month; every month has a permanent page. Cite the month, not the front page.
- 25 of 56most-starred repositories have one at the root
- 45%hit rate, checking in star order
- 1,034tokens in the median file, of 30 files
- 6,249tokens in the largest
This is a sample, not a ranking of GitHub: it is the first 25 repositories with such a file, going down from the most stars. A large file is not a bad file. Six thousand tokens of rules an agent needs are well spent, and the table says what each file costs, not whether it is worth it.
The cost is not the money, which comes to cents. It is the room: these tokens sit in the context window for the whole session, taken before the agent has read any code.
Where does your file sit?
Enter the token count of one context file. It is placed against the 30 files in this month's index. The same lookup is built into the tool: npx ctxcap --compare.
This is the permanent page for November 2026. The newest month is here.
30 files in 25 repositories, largest first
| # | Repository and file | Stars | Tokens |
|---|---|---|---|
| 1 | n8n-io/n8nAGENTS.md | 210,567 | 6,249 |
| 2 | deepseek-ai/deepseek-harnessAGENTS.md | 281,464 | 4,335 |
| 3 | openclaw/openclawAGENTS.md | 394,509 | 4,124 |
| 4 | anomalyco/opencodeAGENTS.md | 226,984 | 4,025 |
| 5 | Significant-Gravitas/AutoGPTAGENTS.md | 188,035 | 3,665 |
| 6 | NousResearch/hermes-agentAGENTS.md | 263,552 | 2,891 |
| 7 | affaan-m/ECCAGENTS.md | 303,571 | 2,794 |
| 8 | morluto/reaAGENTS.md | 461,130 | 2,691 |
| 9 | f/prompts.chatAGENTS.md | 177,503 | 2,226 |
| 10 | obra/superpowersAGENTS.md | 317,895 | 2,056 |
| 11 | twbs/bootstrapAGENTS.md | 175,064 | 2,026 |
| 12 | ultraworkers/claw-codeAGENTS.md | 194,770 | 1,643 |
| 13 | huggingface/transformersAGENTS.md → .ai/AGENTS.md | 168,992 | 1,482 |
| 14 | torvalds/linuxAGENTS.md → README | 255,740 | 1,265 |
| 15 | mattpocock/skillsCLAUDE.md | 324,849 | 1,041 |
| 16 | affaan-m/ECCCLAUDE.md | 303,571 | 1,027 |
| 17 | avelino/awesome-goAGENTS.md | 194,037 | 774 |
| 18 | TheAlgorithms/PythonAGENTS.md | 226,129 | 688 |
| 19 | DietrichGebert/ponytailAGENTS.md | 191,216 | 597 |
| 20 | f/prompts.chatCLAUDE.md | 177,503 | 571 |
| 21 | multica-ai/andrej-karpathy-skillsCLAUDE.md | 233,318 | 534 |
| 22 | vinta/awesome-pythonAGENTS.md | 334,906 | 436 |
| 23 | firecrawl/firecrawlCLAUDE.md | 201,951 | 420 |
| 24 | langgenius/difyAGENTS.md | 161,198 | 394 |
| 25 | firecrawl/firecrawlAGENTS.md | 201,951 | 370 |
| 26 | ultraworkers/claw-codeCLAUDE.md | 194,770 | 250 |
| Under 200 tokens: 4 files. In 3 repositories, every context file is this small. Not numbered. | |||
| – | ollama/ollamaAGENTS.md | 187,049 | 91 |
| – | react/reactCLAUDE.md | 252,626 | 87 |
| – | microsoft/vscodeAGENTS.md | 195,399 | 51 |
| – | ollama/ollamaCLAUDE.md | 187,049 | 21 |
A common guess is that the best-maintained projects write the least. The numbers do not support it: size and stars move together weakly, if at all.
Also in this index: the MCP Context Cost Index — what tool definitions cost before the first message — and stale-reference receipts — dead paths in other repositories' context files, checked by hand.
How it was counted
- The 56 most-starred repositories that are not archived were checked in star order, stopping at the 25th with a context file. Stopping on a hit slightly overstates the 45% rate.
- Only files at the repository root are counted. Nested
AGENTS.mdandCLAUDE.mdfiles, files these import, and other rule files are not. - A symlink is counted as the file it points to, because that is what an agent that opens it reads.
AGENTS.md → READMEmeans the agent is given the README. - A file that is a copy of the one beside it, or that only imports it, is counted once. A repository with both files appears twice when they differ.
- Tokens are counted with
gpt-tokenizer@3.0.1(OpenAI's o200k encoding). Other models tokenize differently, so read these as approximate and compare them with each other. - Fetched on 2026-11-15 from each repository's default branch. This is the first month, so there is nothing yet to compare against; changes appear from the next one.
Count your own repository
Every number here was counted with ctxcap, an open-source linter for these files. It runs on your machine and makes no network calls. Two lines, in any repository:
npx ctxcap # tokens in each context file npx ctxcap --stale # references in them that no longer resolve
Data for this month: 2026-11.json. Method and how to rebuild it. Something wrong with your repository's row? Open an issue.