repowise vs code-review-graph, the numbers.
code-review-graph has the highest precision and the lowest coverage of anything we measured. It also advertises 30 tools over a 40,904-node graph that the agent, under Claude Code, called zero times in 15 questions. Under Codex, the same setup was called on all 15. Both numbers ship.
A tool with the highest precision in a field can still be the wrong choice if an agent never reaches for it. Coverage, precision, and adoption are three different questions, and we measured all three separately rather than letting one stand in for the others.
Retrieval was graded deterministically against 42 sealed ContextBench instances. The agent loop was measured on both Codex and Claude Code, same servers, same questions, same indexes, different harnesses. Every number below carries its n.
Highest precision, lowest coverage.
Deterministic grading against gold file spans, no LLM judge. code-review-graph sits at both ends of this table at once.
Finding the right files
File coverage on 42 ContextBench instances pinned before any of this work started and kept sealed until the final measurement. Grading is deterministic and no LLM judge is involved.
| Tool | File coverage | Precision | Files served |
|---|---|---|---|
| repowise (get_answer) | 0.876 | 0.087 | 19.2 |
| repowise (search_codebase) | 0.742 | 0.168 | 8.2 |
| CodeGraph | 0.610 | 0.093 | 14.0 |
| Graphify | 0.546 | 0.033 | 34.5 |
| code-review-graph | 0.445 | 0.240 | 5.4 |
Scroll the table sideways to see every column.
n = 42 sealed instances. Head to head against CodeGraph, get_answer wins 19, loses 1, ties 22, sign test p = 0.00004. The pooled 112-instance figure would be 0.835 and is deliberately not quoted: averaging the halves loses the only number that matters, which is how the tool does on instances it has never seen.
code-review-graph returns 5.4 files a question at 0.240 precision, the tightest, most confident result set in the table. It also covers only 0.445 of the gold file spans, the lowest of the five tools measured, which is the direct cost of that tight result set.
Zero calls on Claude Code. All 15 on Codex.
Same server, same 40,904-node graph, same 15 questions. The only variable that changed was the agent harness.
The second proof point, on Claude Code
The same six arms, the same servers and the same indexes, on a different agent harness. The result depends on the harness as much as on the tools.
| Tool | Tools advertised | Schema cost (chars) | Agent used it | Output tokens | vs bare agent | p |
|---|---|---|---|---|---|---|
| repowise | 10 | 17,561 | 15 / 15 | 2,420 | -15.9% | 0.035 |
| CodeGraph | 1 | 1,567 | 13 / 15 | 2,540 | -11.7% | 0.302 |
| Serena | 29 | 29,050 | 4 / 15 | 2,551 | -11.3% | 1.000 |
| code-review-graph | 30 | 28,118 | 0 / 15 | 2,768 | -3.8% | 0.302 |
| Graphify | 10 | 5,482 | 3 / 15 | 2,878 | 0.0% | 1.000 |
| bare agent (control) | 0 | 0 | n/a | 2,877 | baseline | n/a |
Scroll the table sideways to see every column.
Claude Code (claude-sonnet-5), 15 questions. Treat the "agent used it" column as unstable, including our own 15 of 15: rerunning the same setup on later days returned 4 of 15 and then 3 of 15 for us, and 2 of 14 for CodeGraph. It is a property of the pairing of tool and harness on a given day, not of the tool. No quality or cost figure may be compared across the two harnesses, which were graded by different judges.
code-review-graph advertises 30 tools over a built, embedded graph of 40,904 nodes and 380,168 edges. Under Claude Code the agent never called it once across 15 questions. Under Codex, with the same server and the same questions, it was called on all 15. That number is a fact about the pairing of tool and harness on a given day, not a verdict on the tool, and it is why it appears twice in the full benchmark rather than as one figure.
The same server, index and questions under Codex: called on all 15. That number is a fact about the pairing of tool and harness, never about the tool. How we measured · raw data
Three undocumented steps. Each one a silent zero.
Getting code-review-graph measured fairly was most of the work behind this page. It needs three setup steps that are not covered in its README, and each one, missed on its own, produces a clean, plausible-looking zero rather than an error you would notice. A tool scoring zero because it was set up wrong is not a result. We found all three steps, verified the server actually answered with a working probe before any run counted toward a headline number, and only then measured the zero-call result under Claude Code above. The same discipline applies to every arm in every run on this page.
Which one is right for you?
Choose repowise if
- You want retrieval that finds most of the right files, not the fewest, most confident ones
- You want a tool the agent reliably reaches for without extra prompt engineering, in our measured runs
- You want a wiki, code health, dead code, git intelligence, and decisions in the same index
- You want an open-source, self-hostable index with agent-neutral MCP tools
Choose code-review-graph if
- You want the single most precise result set in the field when it does return something, at the cost of coverage
- You are building specifically for a PR-review workflow that can use a large embedded graph directly
- You are running under an agent harness, like Codex in our measurements, that reliably discovers and calls the tool
repowise vs code-review-graph
| Capability | repowise | code-review-graph |
|---|---|---|
| Precision on sealed retrieval instancescode-review-graph: 0.240, highest measured. repowise (search_codebase): 0.168 | Partial support | Included |
| File coverage on sealed retrieval instances0.876 (get_answer) to code-review-graph's 0.445, lowest measured, n=42 | Included | Not included |
| Called reliably across agent harnessesrepowise: 15 of 15 on Claude Code that day, unstable across reruns. code-review-graph: 0 of 15 on Claude Code, 15 of 15 on Codex | Partial support | Not included |
| Output-token reduction in a real agent loop-31.6% to code-review-graph's -6.0%, both on Codex, code-review-graph's p = 0.046 | Included | Partial support |
| Indexing speed on django/djangocode-review-graph: 44.8s. repowise (--no-prose): 366.8s | Not included | Included |
| Auto-generated wiki and Q&A | Included | Not included |
| Defect-validated code-health score | Included | Not included |
| Dead code detection | Included | Not included |
| Git intelligence (hotspots, ownership, bus factor) | Included | Not included |
| Architectural decision records | Included | Not included |
| MCP tool surface advertisedrepowise: 10 flagship tools. code-review-graph: 30 tools, 28,118 characters of schema | Included | Partial support |
Self-assessed against publicly documented features as of August 2026. A dash means partial or limited support. Vendor capabilities change, so please verify against code-review-graph's current docs before deciding.
What repowise costs.
The repowise core is open source under AGPL-3.0 and free to self-host, with every heuristic public. Hosted tiers: Free for public repos, Pro at $15 per month, Teams at $60 per month, and custom enterprise licensing when you need it. code-review-graph's pricing changes, so verify it on their site, then compare it with a tier you can read line by line.
Questions, answered
Is code-review-graph a good alternative to repowise?
It has the highest precision of any tool we measured, 0.240, meaning the files it does return are more likely to be right. It also has the lowest coverage, 0.445, meaning it misses more of the gold file set than anything else in the field. And under Claude Code, across 15 questions, the agent never called it once, despite it advertising 30 tools over a built graph of 40,904 nodes.
Did you set code-review-graph up correctly? Zero calls sounds like a broken setup.
That was our first suspicion too, and it is why this took real work to confirm. code-review-graph needs three setup steps that are not documented in its README, and each one produces a clean, plausible zero when missed rather than an error. We found all three, confirmed the server answers correctly with a working probe before any run counted, and then still measured zero calls under Claude Code on 15 questions. Under Codex, with nothing else changed, it was called on all 15. That is a fact about the pairing of tool and harness, not a broken setup.
What version of code-review-graph did you test?
code-review-graph 2.3.7, measured in August 2026 against repowise at commit 081a59fa. This is a dated comparison; verify against current docs before deciding.
Can I reproduce these numbers?
Yes. The retrieval matrix, both agent-loop runs, and the indexing-time measurement are all published with raw data in repowise-bench, including the zero-call rows, linked below.
Why is code-review-graph's precision the highest in the field?
It returns fewer files per question, 5.4 against repowise's 8.2 to 19.2, and is more conservative about what it surfaces. That is a legitimate design choice: fewer, more confident results at the cost of missing more of the full gold set, which is exactly the trade-off its 0.240 precision and 0.445 coverage numbers show side by side.
Does the dollar-cost comparison favor a tool that was never called?
It can, and we ran the control that caught it. Under Claude Code, code-review-graph never called its server across all 15 questions, so it is behaviourally identical to a bare agent, yet came out 43% cheaper on dollars because of prompt-cache position effects unrelated to the tool. That is exactly why this page, like the rest of the benchmark, reports output tokens rather than dollars.
Which tool should I actually use?
If you need a large, purpose-built graph for PR review workflows specifically and are running under an agent harness that reliably discovers and calls MCP tools, like Codex in our run, code-review-graph's precision may be valuable there, which we did not measure directly. If you are running under Claude Code and need the agent to actually reach for a tool without you engineering the surrounding prompt, our measured adoption was more consistent.
Is repowise open source like code-review-graph?
Yes. repowise's core is open source under AGPL-3.0 (OSS 0.39.0) and self-hostable, and ships ten flagship MCP tools for any agent.
Check every number yourself.
Every figure here, including code-review-graph's precision lead and its zero-call result, links back to repowise-bench for the raw run data, and to the full benchmarks page for methodology and sample sizes.
The strongest competitor we measured: second on retrieval and the agent loop, and 22x faster to index. We publish that loss too.
0.546 retrieval coverage from 34.5 files a question, the least precise tool we measured, against repowise's 8.2.
Indexes on demand with no build step, a real advantage, but makes 42% more tool calls than a bare agent while writing less.
Validated code health you can reproduce, not a raw count of issues found, plus a durable index and decisions.