DepTime

An API time machine for Python libraries, built for coding agents.

When an agent hits ImportError, AttributeError or "unexpected keyword argument" in a fast-moving library, one call to DepTime says what changed, in which release, and what to use instead. Answers are computed from static diffs of every indexed release, not from docs or model memory.

Free, open source (MIT), and local: your code and errors never leave your machine.

Every night, GitHub Actions reads each new release's public API with griffe, which parses the source without running it, diffs it against the previous release, and publishes the result as one file, index.sqlite.xz, on this repo's Releases. On your machine, deptime mcp checks that file's sha256 and schema before swapping it in, then answers every question locally in milliseconds. The full design is in docs/04-architecture.md, and the decisions are in docs/adr/.

TypeError: Client.__init__() got an unexpected keyword argument 'proxies'

{
  "symbol": "httpx.Client.__init__(proxies)",
  "issue": "param_removed",
  "installed_version": "0.28.1",
  "last_version_with": "0.27.2",
  "first_version_without": "0.28.0",
  "deprecated_in": "0.26.0",
  "replacements": [
    {"target": "proxy", "confidence": 0.8, "source": "deprecation_message"},
    {"target": "mounts", "confidence": 0.79, "source": "deprecation_message"}
  ],
  "pin_to_restore": "httpx<=0.27.2",
  "confidence": "high"
}

That's the real explain_error answer from the published index, trimmed. The full answer also quotes the deprecation message DepTime found in httpx's own source, and adds a one-sentence explanation for your agent to pass on.

Status: alpha. The prebuilt index is rebuilt twice a day. Known gaps are tracked in BACKLOG.md.

How well it works

Each check's right answers were written down before DepTime first ran on it (bench/). Judge DepTime by the first run: the fixes since then were designed from its misses, so the later numbers are flattering.

Check	Cases	First run	Now
Right package, name and breaking release for a traceback from a GitHub issue, on the 2026-11f holdout	26, after every fix so far	76.9%	76.9%
Code that still works after an upgrade, flagged as broken by check_symbols (lower is better)	2,113 uses from 171 upgrades of real projects, labelled by running them	0.1%	0.1%
The same, by diff_versions(used_symbols=...)	2,113	0.1%	0.1%
Code that breaks, flagged by check_symbols / diff_versions	152 uses	100% / 96.7%	100% / 96.7%
The same four, on the 2026-11e holdout: tracebacks; working code flagged by check_symbols / diff_versions; breaks flagged	28; 2,024 / 2,052 uses; 140	85.7%; 0.05% / 0.2%; 99.3% / 97.9%	96.4%; 0.05% / 0.05%; 100% / 100%
The same four, on the 2026-11d holdout: tracebacks; working code flagged by check_symbols / diff_versions; breaks flagged	30; 2,224 / 2,279 uses; 176	86.7%; 0.9% / 0.7%; 96.6% / 96.0%	93.3%; 0.09% / 0.09%; 99.4% / 99.4%
The same four, on the holdout of 2026-11-09: tracebacks; working code flagged by check_symbols / diff_versions; breaks flagged	35; 1,801 / 1,841 uses; 128	88.6%; 0.5% / 0.3%; 99.2% / 99.2%	97.1%; 0.06% / 0%; 100% / 98.4%
The same four, on the holdout of 2026-11-08: tracebacks; working code flagged by check_symbols / diff_versions; breaks flagged	37; 2,503 / 2,542 uses; 146 (140 once six wrong labels are corrected)	91.9%; 1.5% / 0.2%; 95.9% / 95.9%	100%; 0% / 0%; 100% / 100%
The same four, on the holdout of 2026-11-07: tracebacks; working code flagged by check_symbols / diff_versions; breaks flagged	35; 3,600 / 2,399 uses; 130	88.6%; 3.8% / 0.3%; 97.7% / 76.9%	100%; 0.1% / 0.1%; 98.4% / 98.5%
Right package, name and breaking release for a traceback from a GitHub issue	32 collected 2026-11-05	68.8%	100%
The same, on a holdout fixed in advance from TimeMachine-bench	29	72.4%	96.6%
Error kind and package read from the traceback: scripts, pytest, chained exceptions, ExceptionGroup, IPython, Python 3.10–3.14	736, holdout fixed in advance	93.8%	100%
Replacement hints at confidence ≥ 0.75 that are right	hints for 50 documented migrations	11 of 11	42 of 45
Documented migrations that get a right hint at ≥ 0.75	50	11	42 (32 without hand-checked hints from migration guides)
Names that exist at runtime but read as absent	2,428 sampled at runtime from 10 dynamic packages	5.1%	1.2%

check_symbols answers unknown, with the reason, where the index can't tell (a path through an instance such as pytest.mark, or into a distribution it doesn't cover): 66 working uses and 5 breaks of the 2026-11f holdout (28 working uses of 2026-11e's, 55 working uses and 2 breaks of 2026-11d's, 40 working uses and 4 breaks of 2026-11-09's, 43 working uses of 2026-11-08's, 79 working uses and 4 breaks of 2026-11-07's since that answer was added), left out of its rates above. Its misses are new items in the backlog, to be scored on the next set. Both tools' two false alarms on 2026-11f are one scikit-learn import from a compiled module, which the index can't read; most of its traceback misses are errors as a web page pastes them (a keyword in backticks, __init__ printed as init). The fixes for the 2026-11d holdout's misses are in the index of 2026-11-10 (SNAPSHOT_FORMAT 26): 2 of its 20 and 17 false alarms are left per tool, chromadb's embedding functions, which its __init__ writes into globals(), and the dspy placeholder release that gave 15 of each is now read from the distribution it installs. Since 2026-11-10 the holdout's rule reads a parameter diff_versions reports as required at the older release going down (polars 0.19.6's int_range(n)), and one passed by position (StateGraph(State)), so its "now" figures for diff_versions count both right.

Across all 393 real tracebacks in the benchmark, including the ones fixes were designed from, DepTime names the right package, name and release for 97.2%, and a right replacement is among its first three for 74.6%. CI re-scores them and the holdouts on every code change and every published index, and fails on a drop of more than 2 points (or a false-positive rate up by more than 1).

Install

Requires uv.

# Claude Code. Cursor, Codex, VS Code and Claude Desktop take the same
# command (uvx deptime mcp) in their MCP config.
claude mcp add deptime -- uvx deptime mcp

On first launch DepTime downloads its prebuilt index (one file from this repo's GitHub Releases) in the background, then checks for a newer one at most once a day. Set DEPTIME_OFFLINE=1 to never touch the network. Each index carries a signed record of the CI run that built it; deptime index pull --verify checks it (SECURITY.md).

Claude Code users can also install the plugin in integrations/claude-code with /plugin marketplace add cschmertz/deptime, then /plugin install deptime@deptime. It adds a skill and a local hook that suggests DepTime when a Python version-change error appears in a command's output.

First launch timing out? uvx installs DepTime the first time a client starts it. On macOS with an x86_64 uv (an Intel Mac, or Intel Homebrew on Apple silicon), one dependency, cryptography, has no prebuilt wheel and is compiled, which can take longer than a client waits for a server to start. Run uvx deptime --help once beforehand.

Client compatibility

Checked 2026-11-05 with deptime 0.1.4 (details):

Client	Result
Claude Code 2.1.318	✅
Codex CLI 0.166.0	✅
MCP Python SDK clients 1.0–1.30 and TypeScript SDK clients 1.0–1.32 (protocols 2024-11-05 to 2025-11-25), over stdio and against deptime serve	✅
Cursor, VS Code, Claude Desktop	not checked yet; they take the same uvx deptime mcp command

Tools

Tool	Use it when
explain_error(traceback, installed_versions?, pip_freeze?)	An import or call into a library just failed
symbol_history(package, path)	Before trusting an API you remember
check_symbols(package, version, symbols[{path, kwargs}])	Pre-flight check of the symbols and kwargs you're about to use
diff_versions(package, from_version | date, to_version?, used_symbols?)	Upgrading, or "what changed since my training cutoff?"
coverage()	What's indexed (53 fast-moving packages: AI/agent stack + libraries with known breaking changes; list)

Every answer carries evidence and a confidence level. If a package isn't covered, DepTime says so instead of guessing. The contracts are in docs/03-product-spec.md.

CLI

uvx deptime explain --installed httpx==0.28.1 < traceback.txt
uvx deptime history httpx Client.__init__
uvx deptime check httpx 0.28.1 'httpx.Client:proxies,proxy'
uvx deptime diff httpx 0.27.2 0.28.1
uvx deptime diff pydantic 2024-06-01              # everything since a date
uvx deptime index status | pull

Hosting (optional)

Local mode covers coding agents. For web clients that only accept a URL, run deptime serve (MCP at /mcp, JSON at /v1/*) on any container host. The repo includes a Dockerfile and fly.toml, and the server applies anonymous rate limits. See docs/05-self-hosting.md.

Prior art and credits

• griffe does the static analysis.
• since-cutoff diffs your pinned dependencies against a model's training cutoff, locally and on demand. DepTime differs by shipping a prebuilt cross-version index and starting from a traceback.
• Context7 and Ref serve version-specific docs. DepTime answers symbol-level "when and what replaced it" questions.

Contributing

The most useful contributions are wrong-answer reports, which become benchmark cases. See CONTRIBUTING.md. Development setup:

uv sync && uv run pytest
# build a small local index from real PyPI wheels
uv run deptime ingest httpx --versions 0.27.2 0.28.1

Docs

Product spec · Architecture · Self-hosting · Risks & spikes · Privacy & security · Changelog

License

MIT