Compared with other tools
CodeGraph · GitNexus · Graphify · Serena · potpie.
ctx-optimize is a Go CLI plus an agent skill: no server, no model in the gather, no MCP.
What each one is
Section titled “What each one is”| ctx-optimize | CodeGraph | GitNexus | Graphify | Serena | potpie | |
|---|---|---|---|---|---|---|
| Shape | Go CLI + skill | SQLite + MCP | MCP, 16 tools | Python skill | LSP over MCP | Neo4j + agents |
| Model in the gather | no | no | no | labeling | no | LLM in the loop |
| MCP | no — folder + CLI | 42 tools | 16 tools | no | yes | API |
| License | MIT | MIT | noncommercial | open | MIT | open-core |
| Outer surface (env, hosts, binaries, routes) | boundaries |
no | no | no | no | no |
| Add a language yourself | grammar URL | no | no | no | via LSP | no |
| Lives outside your repo | yes | yes | yes | no — graphify-out/ |
yes | server |
MCP tool counts and licences are as published by each project; we have not audited them.
Where the numbers are
Section titled “Where the numbers are”Serena and potpie have no rows below — we have not run them, and an empty cell stays empty.
Linux kernel · v6.9 · 144,011 files ᴬ
Section titled “Linux kernel · v6.9 · 144,011 files ᴬ”| gather | query | |
|---|---|---|
| ctx-optimize | 118.2 s — 2.85M nodes, 5.54M edges | 4.11 s |
| CodeGraph | 289.9 s | 0.98 s |
| Graphify | 527.7 s | 22.8 s — only after raising its 512 MB cap |
| GitNexus | did not finish within 45 min | — |
On that run we built the graph 2.5× faster, and we are the only tool that produces a complete kernel graph. CodeGraph answers ~4.2× faster than we do at this scale: it seeks in SQLite, we deserialize the whole graph per invocation. (An earlier version of this page said 0.79 s and 7.5×. That came from a single-word query against everyone else’s full phrase; re-measured fairly on 2026-08-16 it is 0.98 s and 4.2×. The original figure is left in the result file so the error stays visible.)
Small corpora · 253–1,474 files ᴮ
Section titled “Small corpora · 253–1,474 files ᴮ”ctx-optimize leads all four tools on cold gather, warm re-gather, query and disk.
Big repos · 723–10,142 files ᴬ
Section titled “Big repos · 723–10,142 files ᴬ”Cold gather: we lead all six. Warm re-gather: we lose all six to CodeGraph, 10–31× — we re-gather where its sync is a true incremental. Full tables on benchmarks.
Graded agent · gorilla/mux · gpt-4o-mini ᶜ
Section titled “Graded agent · gorilla/mux · gpt-4o-mini ᶜ”| ctx-optimize | Graphify | |
|---|---|---|
| Correct | 67% | 40% |
| Tool calls per session | 15.0 | 26.0 |
ᴬ Unpinned. Kernel and big-repo runs predate our pinning harness and were measured on ctx-optimize v0.8.0–v0.12.0 (HEAD is v0.15.0); single run, not best-of-3. Being re-run — the drift runs against us, since v0.14 roughly halved gather. Our column in that run included wiki generation, which left the default path in v0.12 — so it measures work the tool no longer does. Re-measured on HEAD, the same kernel gather is 61.2 s median (60.13 / 61.17 / 63.24, byte-identical output). Competitors have not been re-run, so we state no new ratio until they are.
ᴮ Pin-verified 2026-08-15 (CodeGraph 572d22bf, Graphify 2fa6cd3d, GitNexus 91b22676), but recorded at load average 8.77.
ᶜ 12 questions × 3 runs, n = 36, no LLM judge, one small repo. Transcripts.
Who to pick
Section titled “Who to pick”| If you want | Take |
|---|---|
| MCP on every host today | CodeGraph |
| Fastest queries on a very large repo | CodeGraph |
| Deepest Claude MCP (noncommercial) | GitNexus |
| Type-exact rename | Serena — LSP is more precise than any static graph, ours included |
| A funded Neo4j platform | potpie |
| An exact string | ripgrep — and your agent should still use it |
One binary, no server, no model, cited file:line, and what your system talks to |
ctx-optimize |
What we will not claim: limits.