Should you trial Gortex for agent context?
Gortex is worth evaluating if your coding agents need to follow relationships across repositories. The evidence here supports a successful source build and a reason to check workspace behavior carefully. It does not establish how accurately Gortex indexes your code or how much context it saves an agent.
The README describes a persistent SQLite graph containing functions, classes, call chains, HTTP routes and cross-service contracts. A daemon maintains the graph, and the CLI, MCP server, HTTP API and web interface expose it. This gives a team a concrete trial question: can graph queries recover the relationships its agents currently have to find by reading files?
The README advertises 257 languages or grammars and token reductions of up to 50 times per response. Those are project claims. We did not measure language coverage, token savings, indexing throughput or query latency. The README also gives differing agent-integration counts, so this review does not present one as a verified total.
What our October 8 run established
Our recorded run used commit e12965558e80c153805521d174efcbda2b0654d5, dated October 5, 2026. Dependency setup completed in 206 seconds, with 417 packages in the dependency summary. The build step succeeded in 445 seconds. These observations describe the source checkout and command steps; they do not verify a release installation, an installed CLI on PATH or a working MCP connection.
The recorded sandbox had 3 CPUs, 8 GB RAM and image golang:1.24-bookworm. The setup tail says go: downloading go1.27.0 (linux/amd64). Calling this a successful Go 1.24 build would therefore misrepresent the evidence. The checkout scan recorded 4,930 files, approximately 1,176,896 lines of source and 48.9 MB. Those are repository measurements, not release-binary size or indexing-workload results.
The test command ran for 852 seconds, exited with code 1 and did not hit the harness timeout. The parser summary recorded 206 successful and 8 failed package results, with 43 additional packages reporting no tests. Its total of 214 refers to the successful and failed package results. It is not a count of individual tests, and no complete-log reconciliation established unique, exhaustive coverage.
The supplied tail names github.com/zzet/gortex/internal/mcp as a failed package after 360.091 seconds. It does not show a failing assertion or explain the other reported failures. A nonzero test-command exit is a reason to investigate this commit in the intended environment, not proof that the CLI or graph engine is broken.
Installation has two distinct paths
The installation guide describes self-contained release binaries for Linux, macOS and Windows. Source builds require Go 1.26+ and a C toolchain because the tree-sitter bindings use CGO. The static Linux binary cannot load custom tree-sitter shared objects; the guide directs users who need those grammars to build from source.
Installing a binary is only the first check. Agent configuration, daemon startup, repository tracking and confirmation that the assistant actually calls graph tools are separate steps. Our recorded build does not demonstrate their completion.
Separate historical reports from current blockers
Search-truncation issue 672 reported that search_text silently capped results and their reported count. In the original report, a query returned 1,000 matches while git grep found 8,963. The supplied issue history now shows it closed as completed through pull request 845 on October 3, 2026. It should no longer be described as an open, unchanged defect. Closure does not establish that every released version contains the change or that all indexed searches match raw source search.
Workspace-scoping issue 673 remains open in the supplied extract. The reporter describes already-indexed nodes keeping their old workspace ID after a repository joins a shared workspace. Known text then returns count: 0. This is a reported failure on v0.63.7+0988412, not a behavior reproduced in our October run or established for every current build.
PDF-memory issue 676 is also historical rather than an unchanged open blocker. Its supplied history shows closure as completed through pull request 926 on October 19, followed by a referenced PDF extraction memory-and-work fix on October 20. Our October 8 run predates those events and did not index a PDF workload. Neither the old report nor its closure measures memory behavior in a current release.
A useful acceptance check
Choose a representative repository set and write down a few known symbols, callers and cross-repository relationships before indexing. Check those answers through the interface your agent will use. Then move an already-indexed repository into a shared workspace and repeat the checks. This directly addresses the remaining workspace report without assuming it occurs everywhere.
Keep ordinary source search available when completeness matters. Compare indexed and raw-search results with their exclusions in mind: a difference may reflect indexing policy rather than a query defect. Measure memory, query latency and agent context use separately if they determine whether adoption is worthwhile. None was measured in our run.
How to read the ratings
These are editorial evidence ratings on a 1-to-5 scale: 1 means little usable support in the supplied evidence; 2 means substantial unresolved limits; 3 means enough support for a controlled trial; 4 means detailed supporting evidence with limited gaps; 5 requires broad verification in the relevant category. Unknown adoption, compatibility or reliability cannot earn a high rating. The scores are not benchmark results or an overall production-readiness certificate.
Setup receives 3 because the source build succeeded and installation paths are documented, while installed-binary and agent checks remain unverified. Documentation receives 3 because the installation extract answers practical setup questions, but the README has inconsistent integration counts and most linked documentation was not supplied in full. Community receives 2 because issue discussion provides concrete reports and maintainer responses, but adoption and support responsiveness were not measured. Maturity receives 2 because workspace scoping remains an open report and the recorded test-command failures are unexplained, despite documented progress on other issues.
For adoption, start with the known-query checks and a separately verified installation. Expand the trial only after Gortex answers the relationships your agents need.

