OpenContextually

Give your coding agent the context it should read before it starts working.

Quickstart · Discord · Discussions · ContextBench · Ecosystem · Contributing · Community

OpenContextually turns a task into a small, ranked, explainable context package from your repository. It is not another coding agent — it is the open context layer before the agent.

Agent failed because it read the wrong context? Bring us the case.

Quickstart

pip install opencontextually

gctx "fix the authentication bug"

No model. No API key. No network. No database. No setup. One dependency. The same repository and task produce byte-identical output every time.

flowchart TB
    T(["your task<br/>fix the authentication bug"]) --> OC
    subgraph OC ["OpenContextually · local · deterministic · no model"]
      direction LR
      S["SELECT<br/>likely files"] --> F["FOLLOW<br/>imports"] --> B["BOUND<br/>repo limits"] --> C["CHECK<br/>gaps + conflicts"] --> E["EXPLAIN<br/>why each file"]
    end
    OC --> R(["task-ready context<br/>every file with a reason"])

What it looks like

gctx "fix the authentication bug"

Run from examples/auth_bug/ — a small fixture with an auth module, a config file, docs, and a test — this is the real, unedited output:

fix the authentication bug
6 relevant · 19 excluded

  src/auth/middleware.py  defines AuthenticationError
  README.md               defines authentication requirements
  tests/test_auth.py      imports middleware.py  ← via middleware.py
  config/auth.yaml        configuration referenced by authentication code
  docs/security.md        defines authentication requirements
  src/users/session.py    imported by middleware.py  ← via middleware.py

  ⚠ session.timeout_minutes: config/auth.yaml:3 declares 60 minutes, but docs/security.md:6 says 30 minutes
  ○ No test references session timeout minutes (config/auth.yaml:3)
  ○ No test references session expired (src/users/session.py:67)

  -v for code excerpts

Excluded: 19 files
  19 files scanned, not relevant enough

Checks run: configuration_discrepancy, test_reference_gap

Three things happened beyond ranking:

• session.py was reached through an import, not through text. It matches none of the task's words. middleware.py imports it, and the ← via marker records that edge.
• A config/doc disagreement was surfaced — 60 minutes against 30.
• Every file carries a reason, and everything excluded is accounted for.

Search finds matches. Context needs relationships.

Search answers "where do these words occur?". That is a different question from "what should be read for this task?".

grep / ripgrep

  task ──────────────────►  keyword matches


OpenContextually

  task ──►  direct matches
                 │
                 ├──►  imported dependencies
                 ├──►  tests that exercise them
                 ├──►  relevant configuration
                 └──►  governing documentation
                              │
                              ▼
                     bounded context package

A task like fix the authentication bug can require a file that never contains the words authentication or bug. OpenContextually starts from direct relevance, follows Python's own import graph, applies repository boundaries, runs two narrow deterministic checks, and explains every inclusion.

Be honest about the split: relationship-following is the part search cannot do at all, but most of the value on a typical run is ranking and compression of files search could have found. Asking sqlfluff about "indentation rule fires on a templated line", a case-insensitive grep for any of the task's words matches 415 files; OpenContextually returns 12, each with a reason. Both numbers are reproducible from the corpus below.

Install

pip install opencontextually

Python 3.10+, one runtime dependency. To work on OpenContextually itself:

git clone https://github.com/gammalex-ai/opencontextually
cd opencontextually
pip install -e ".[dev]"

Using it

CLI

gctx is short for GammaLex Context. The same command is also installed as octx (the original name, kept working) and opencontextually. Flags compose (-v --all); --json is unaffected by either and is always full fidelity.

Write tasks the way you'd describe the bug. Naming a specific behavior or symbol beats a directory-shaped noun.

Python

from opencontextually import get_context

package = get_context("fix the authentication bug")

print(package.render())     # the text above
package.to_dict()           # the same content as JSON

MCP

Agents that speak MCP can call it directly. Requires the optional extra:

pip install "opencontextually[mcp]"

Point your MCP client at the opencontextually-mcp command:

{
  "mcpServers": {
    "opencontextually": {
      "command": "opencontextually-mcp"
    }
  }
}

It exposes exactly one tool — get_context(task, root=".") — returning the same shape as gctx --json.

Tested on real repositories

Scripted tests lock in fixes; they do not find them. Most real defects in this project were found by running it against unfamiliar repositories and reading the output, so a standing corpus is part of the project (benchmarks/, with a runner that also checks determinism and sweeps for leaked secrets).

What it selected, and what it missed

Fourteen repositories, each with a hand-checked answer key: the files that actually implement or test the behaviour the task names, read out of the project at its pinned commit. The keys are committed in benchmarks/answer-keys.json.

Six were used while tuning ranking, so their results are fitted to an unknown degree. Eight were held out — their keys were written and committed before the tool was run against them, and nothing was tuned afterwards.

The two groups disagree by about 16 points, and the held-out figure is the one that predicts what happens on a repository this project has never seen. 79% and 66% are the numbers to argue with.

The default view matters more than the total: the compact output shows eight files, so a key file recovered at rank 15 was found but not delivered.

What holds everywhere: zero fixture, vendor, generated or CI files selected, every path inside the configured root, and 0.05%–2.1% of repository bytes delivered. sqlfluff's 5,249 test fixtures and django's 736 documentation files are excluded in full.

The held-out repositories found what the tuned six could not, which is the entire reason for holding them out:

• A bundled previous major version. pydantic ships Pydantic 1 inside Pydantic 2. Six of eighteen slots went to pydantic/v1/* while main.py was missed. Fixed — a v1/ directory inside a v2 package is now damped.
• A repository holding several copies of one document. rich spends five slots on README translations; fastapi repeats one page across docs/en, docs/hi and docs/tr; pytest has 50 release announcements. Not fixed. A family cap was written, measured, and reverted for collapsing genuinely different pages that share a filename.
• Vocabulary collisions, as on django: rich ranks progress.py (ProgressColumn) first for a table-width task, and scrapy misses its own test_dupefilters.py. Tracked as issue #3.

The corpus

Ten public Python projects, each with a plausible task, all reproducible with benchmarks/dogfood.py:

All ten are MIT- or BSD-licensed public projects, unaffiliated with this one, chosen for a spread of size and layout rather than for flattering results. Each was cloned with --depth 1 on 2026-08-30 at the commit above; file counts and timings are specific to those commits.

18,693 files in total. Zero secret-shaped strings reached any package, and every result was byte-identical across repeat runs. Times are best-of-three on an M-series Mac running Python 3.13 with a warm page cache; treat them as orders of magnitude, not a benchmark.

A caveat on the file counts, because the honest number is smaller than the flattering one: these are all files in a clone. What actually gets read is what survives your ignore rules, and on a repository with heavy build output that is a small fraction. A 42,000-file checkout completing in two seconds sounds impressive and mostly means 41,000 files were gitignored and never opened. Of the 18,693 files above, 16,331 are actually scanned; django's real figure is 5,580 files in 7.5 seconds.

Speed is listed last on purpose. It is a property worth keeping, not the claim — a tool that walks a repository quickly and hands an agent the wrong eight files has not helped anyone.

For gctx "option prompt does not hide the input" against click, the top three are core.py (defines Option), decorators.py (defines option), and termui.py (defines _mask_hidden_input) — the third being a private helper whose name no part of the task literally matches.

Ecosystem & Community

Works with today

Verified by the test suite and by hand against a clean install — nothing here is aspirational.

Community integrations

Nothing here yet — this project is new and we would rather show an empty table than a fictional one. The --json output and the MCP server are both stable, documented interfaces, so anything below is buildable today by anyone:

• an editor or IDE extension that runs gctx on the current task
• a wrapper for an agent harness — Cursor, Continue, OpenCode, Aider, or your own
• a GitHub Action that posts the context package for an issue onto its PR
• a shell or tmux integration, a TUI, an alternative renderer
• language support beyond Python's import graph (see GOOD_FIRST_CONTEXT.md)

Built something with gctx? Open an issue or a PR and we may feature it here. Community projects are not maintained by us, and we will say so next to each one.

The question this project is trying to answer

What should an agent know before it acts, and how do we prove it got the right context?

The second half is the hard half, and it is why ContextBench exists: every claim in the section above is checkable against committed answer keys, and the honest number — the held-out one — is the one quoted.

Come argue with the numbers, bring a case where the wrong files were selected, or add a benchmark case from a repository we have never seen: COMMUNITY.md is the map of every way in.

What it deliberately does not do

• Follow imports outside Python. Expansion uses the stdlib ast module; other languages get lexical matching only.
• Understand your code. Ranking is lexical scoring plus import expansion. It is weakest when your task's words are also the repo's naming convention — asking about "the context agent" where many files are named *context* — because filename matches then dominate.
• Find problems for you. The two checks are narrow, named rules that expect to stay quiet: they flag detectable gaps and conflicts — a config value contradicting a documented one, a config key or symbol no test references — not arbitrary missing context. Across the six answer-key corpus tasks they produced zero findings, which is the honest scope: they fire on the patterns they name, and examples/ is where you can watch them do it. Across eleven real repositories, one produced a false positive (since fixed) — the honest measure of how much "high precision" has actually been tested. Silence is the normal outcome; a footer always names which checks ran.
• Guarantee secrets stay out of excerpts. Redaction masks secret-shaped keys and high-entropy strings, but it is best-effort pattern matching, not a secrets scanner. See SECURITY.md.

Scope, determinism, and safety

Discovery reads everything under --root minus what git already ignores — honoring nested .gitignore files, .git/info/exclude, and the global core.excludesFile, plus an optional .opencontextuallyignore. All resolved without shelling out to git, so it works in directories that aren't repositories at all.

Runs are deterministic: the same task and repository produce byte-identical output, which is asserted in the test suite and re-checked by the corpus runner. Nothing is written anywhere, and no network call is ever made.

Contributing

Bug reports, context failures, ContextBench cases, integrations, language support and documentation fixes are all welcome — CONTRIBUTING.md covers the workflow and the scope boundaries, GOOD_FIRST_CONTEXT.md lists concrete places to start, and COMMUNITY.md is where to find people.

License

MIT