Code-graph MCP servers, October 2026
Coding agents waste most of their budget rediscovering structure: grep, open a file, grep again. A wave of tools now hands them a pre-built map instead — a graph of files, symbols, imports and calls, exposed over MCP. This page is the field as of 26 Oct 2026, collected while rebuilding solograph.
The convergence
Almost every tool that is growing landed on the same design:
- tree-sitter for parsing, often dozens of grammars.
- SQLite with FTS5 for storage and search, BM25 ranking. A few keep a graph DB (FalkorDB, Kuzu forks, Neo4j as an option) for Cypher.
- Embeddings optional or absent. The pure-vector entrant (claude-context, Milvus + OpenAI keys) stopped being pushed to in July.
- Freshness as a feature. A file watcher or a per-query sync, so the agent never reads a stale map.
Solograph reached the same place from the other side: its vector search measured worse than plain FTS5 on every axis it was tested on (see the 0.7 section of project-solograph).
The table
Stars measured with gh api on 26 Oct 2026. GitHub no longer serves star histories, so “trending” is fork velocity: days the last 100 forks took, as forks per day.
| Repo | ★ | Store | Embeddings | MCP | What sets it apart | Forks/day |
|---|---|---|---|---|---|---|
| Graphify-Labs/graphify | 136k | graph.json |
none, by design | yes + skill | Leiden communities, edges tagged EXTRACTED/INFERRED, git hooks | 100 |
| Egonex-AI/Understand-Anything | 88k | JSON | none, but an LLM builds the graph | skills | dashboard; the graph costs tokens to build | 25 |
| colbymchenry/codegraph | 75k | SQLite + FTS5 | none | yes | FSEvents sync under 1 s, Swift↔ObjC and RN-bridge edges, ~35 languages | 33 |
| abhigyanpatwari/GitNexus | 48k | Kuzu fork | optional | yes | PolyForm Noncommercial; growth has cooled | 2.5 |
| DeusData/codebase-memory-mcp | 48k | SQLite + FTS5, Cypher-like | optional, compiled into the binary | yes, 17 tools | one static C binary, hybrid LSP, camelCase-aware tokenizer | 12.5 |
| tirth8205/code-review-graph | 32k | SQLite | optional | yes | minimal file set for a diff | 11 |
| oraios/serena | 31k | none, live LSP | none | yes | exact references and renames via language servers | 14 |
| zilliztech/claude-context | 13k | Milvus | required | yes | pure vector; last push July | 0.8 |
| trailhq/Trail-Code | 11k | markdown + JSON | none | yes + hooks | every query tops the graph up (~50 ms) | 6.7 |
| repowise-dev/repowise | 8k | graph + vectors | yes | yes | AGPL; measured its own call-edge precision (85.7%) | 8.3 |
| vitali87/code-graph-rag | 5k | SQLite (Memgraph optional) | optional | yes | moved off Memgraph to embedded SQLite | 1.5 |
| CodeGraphContext | 4k | FalkorDB Lite / Kuzu / Neo4j | not documented | yes | closest architectural twin of solograph; SCIP import | 1.1 |
| jgravelle/jcodemunch-mcp | 3k | local index | not documented | yes | symbol retrieval by byte offset; paid commercial licence, telemetry on by default | 0.8 |
| probelabs/probe | 0.7k | none | none | yes | BM25 + AST on the fly, “agents don’t need embedding search” | 0.1 |
| ast-grep/ast-grep-mcp | 0.5k | none | none | yes, 4 tools | structural pattern search | 0.1 |
Three worth a look
- colbymchenry/codegraph — the cleanest version of the converged design: SQLite + FTS5, no vectors, a watcher that keeps it current, and real Swift/Kotlin/ObjC support. For one repo it does what solograph’s graph does, better.
- codebase-memory-mcp — keeps Cypher-style queries on top of SQLite in a single binary. The answer to “I want a graph query language without running a graph DB”.
- Graphify — fastest-growing, no embeddings at all. What to borrow from it is already in graphify-vs-solograph: edge provenance, a PreToolUse nudge towards the graph, rebuild on git events.
Serena is a complement rather than a rival: LSP precision where tree-sitter guesses.
Measured here: codegraph against plain grep
On 27 Oct codegraph 1.6.2 was run against a real repository with its own README’s method: claude -p, --strict-mcp-config, WITH = only codegraph’s MCP server, WITHOUT = no MCP at all, its CLI blocked in both arms. The repository was a 1,105-file Rust + Swift app; codegraph indexed it in 5 s (20k nodes, 47k edges, a 76 MB index). Two questions, two runs per arm, every answer checked against call sites found with rg:
| Question | Cost, median (with / without) | Tool calls | Correct |
|---|---|---|---|
Trace how plan_montage is reached, from the UI down, across a Swift↔Rust FFI boundary |
$1.73 / $2.07 (−16%) | 16.5 / 25 | 4/4 both |
What breaks if resolve_montage_profile’s signature changes |
$0.69 / $0.65 | 6 / 5.5 | 4/4 both |
Both arms found the same chain through the same files, including the UniFFI hop. The graph saved a sixth of the cost where the agent had to explore, and nothing where it could grep its way there in five calls — which is what codegraph’s own README says (“near-even where it got there in 7”). For that, it wants a global agent install, a per-repo index and a background daemon (one was still running after the runs). Not adopted here; two runs per cell is a small sample, and a repo where questions need 30–40 lookups could come out differently.
What none of them do
Not one covers several projects at once, or the agent’s own history. Solograph’s distinct part is the portfolio: symbol and package search over 77 repositories, and full-text search over every Claude Code session, including the ones Claude Code has deleted. That is the part it kept. Its own structural graph went in 0.8: agents asked it nothing structural in four months, and inside one repo Grep does the job, as the measurement above shows.
Caveats
- Star counts look inflated. In July three sources gave Graphify 76k, 97k and 98k on the same day; it shows 136k now. Fork velocity is a better signal and can be gamed too.
- Every benchmark is the vendor’s own. codegraph claims 62% fewer tokens, codebase-memory 10× (its own preprint), jcodemunch 96%. Only codegraph’s was rerun here, above, and came out far smaller. codegraph’s own README adds the honest counterpoint: answers stay resident, so its sessions end with about 80% more retrieval context in the window.
- “No embeddings” for CodeGraphContext and jcodemunch means their READMEs do not mention any.
Related
- project-solograph — the rebuild this survey was for
- graphify-vs-solograph — the one-to-one comparison
- fff-agent-file-search — file search for agents, the layer below
- agent-toolkit-landscape — the wider tool landscape