# diffctx

> Selects the minimum set of code fragments an LLM needs to understand a git
> diff, under a hard token budget. Deterministic, runs locally, makes no model
> calls. Ships a Model Context Protocol (MCP) server so AI coding agents can
> fetch diff context in one call instead of exploring a repository by hand.

diffctx is **diff-seeded**: the input is a change (a commit, a range, or an
uncommitted working tree), and the output is the surrounding code needed to
review it — callers, type definitions, imports, config dependencies — packed to
fit a token budget. This is a different problem from whole-repository packing
and from query-seeded code-graph lookup.

Formerly published on PyPI as `treemapper`; that name is sunset and frozen. The
current package, CLI, and repository are all `diffctx`.

## What it is not

- Not a repository packer. Tools like repomix export a whole codebase; diffctx
  starts from a diff and selects a small subset under a budget.
- Not a query-seeded code-graph server. Persistent whole-repo graph MCP servers
  answer "who calls X". diffctx answers "what do I need to read to understand
  this patch". If you want a persistent repo-wide graph, those tools fit better.
- Not an LLM wrapper. It calls no model and needs no API key; selection is a
  deterministic optimization over a typed dependency graph.

## Install

- `pipx install diffctx` (Python CLI, all output formats, MCP server via
  `pipx install 'diffctx[mcp]'`)
- `uvx diffctx . --diff HEAD~1` (zero-install)
- `cargo install diffctx` / `npx diffctx` / `docker run ghcr.io/nikolay-e/diffctx`
  (native binary: diff mode, YAML/JSON output)

## MCP server

Stdio transport, read-only, published in the official MCP registry as
`io.github.nikolay-e/diffctx`. Zero-install registration in Claude Code:
`claude mcp add diffctx -- uvx --from 'diffctx[mcp]' diffctx-mcp`.
Tool: `diffctx_context` — one call returns the selected context for a diff
(`mode=locate` ranks without bodies, `fragment_ids` fetches them). The pre-v3
tools `get_tree_map` / `get_file_context` are opt-in via
`DIFFCTX_MCP_LEGACY_TOOLS=1`.

## Method and measured results

Diff-aware context selection is formulated as budgeted submodular maximization
over multi-resolution fragments on a typed dependency graph, with lazy-greedy
selection under a partition matroid. Evaluated on 1500 instances across
SWE-bench Verified, PolyBench-500, and ContextBench Verified.

At an 8000-token budget: pooled file recall 0.919 [0.908, 0.929]. Against
same-budget baselines, paired deltas favor diffctx by +0.371 over whole-file
BM25 packing and +0.410 over the Aider repo-map oracle upper bound
(permutation p = 1e-5). Every table value traces to a committed per-instance
CSV artifact.

Paper: https://doi.org/10.5281/zenodo.18824579

## Token counting

Budgets and printed counts use tiktoken `o200k_base` (GPT-4o family). Counts for
Claude, Gemini, Llama and other families differ, so treat `--budget N` as an
upper bound in o200k tokens and leave headroom. `--with-raw-diff` bundles the
literal patch ahead of the selected context; it is not charged to the budget,
and the stderr summary reports the real output size either way.
See docs/product/token-budget.md.

## Links

- Repository: https://github.com/nikolay-e/diffctx
- Documentation: https://diffctx.com/
- Command-line reference (every flag, default and meaning, rendered from
  `diffctx --help`): https://diffctx.com/product/cli.html
- FAQ (heuristic or oracle, tokenizer, monorepo limits, raw diff, secrets):
  https://diffctx.com/product/faq.html
- Benchmarks (every published number and what it was measured on):
  https://github.com/nikolay-e/diffctx/blob/main/BENCHMARKS.md
- PyPI: https://pypi.org/project/diffctx/
- crates.io: https://crates.io/crates/diffctx
- npm: https://www.npmjs.com/package/diffctx
- MCP registry: io.github.nikolay-e/diffctx
- Changelog: https://github.com/nikolay-e/diffctx/blob/main/CHANGELOG.md
- License: Apache-2.0