Give OpenAI Codexa map of your codebase.
Install Graphify in Codex from the OpenAI plugin directory, or run the Graphify CLI locally through AGENTS.md.
Install Graphify in Codex from the OpenAI plugin directory. The Graphify plugin connects Codex and ChatGPT to your shared codebase context (a code graph, Wiki, Memory and PR reviews) at https://api.graphify.com/mcp with OAuth sign in: no API key, and nothing to build on your machine. Start free at app.graphify.com.
Open the Graphify listing and select Install plugin, or open /plugins in Codex CLI and search for Graphify. Then select Connect and sign in. The same plugin works in ChatGPT. Prefer to add it by hand?
codex mcp add graphify --url https://api.graphify.com/mcp
codex mcp login graphifyPrefer local? The open-source Graphify CLI (uv tool install graphifyy) wires it into AGENTS.md.
Prefer local
- Install the CLI. If
graphifyis not found afterwards, runuv tool update-shelland open a new terminal. - From the project root:
graphify codex install --project. It writesAGENTS.md, the skill under.codex/skills/, and.codex/hooks.json. For a personal skill instead, rungraphify install --platform codex, thengraphify codex installforAGENTS.md. - Enable parallel extraction in
~/.codex/config.toml:
# ~/.codex/config.toml
[features]
multi_agent = true$ uv tool install graphifyy
$ graphify codex install --projectThen build the graph from inside a Codex session. Note the invocation: Codex uses $graphify instead of /graphify.
$graphify .The build writes graphify-out/ with graph.json (the queryable graph), graph.html (interactive view), and GRAPH_REPORT.md (highlights and suggested questions).
What graphify codex install writes
## graphify
This project has a knowledge graph at graphify-out/ with god nodes,
community structure, and cross-file relationships.
Rules:
- For codebase questions, first run `graphify query "<question>"`
when graphify-out/graph.json exists. Use `graphify path "<A>" "<B>"`
for relationships and `graphify explain "<concept>"` for focused
concepts.
- After modifying code, run `graphify update .` to keep the graph
current (AST-only, no API cost).How Codex uses the graph
With the AGENTS.md rules in place, a structural question makes Codex reach for the graph before the file tree: graphify query "<question>" returns a scoped subgraph under a token budget, graphify path "<A>" "<B>" walks the dependency route between two symbols, and graphify explain "<concept>" lists everything connected to one node. Results carry file:line citations, and every edge is tagged EXTRACTED (parsed from the AST), INFERRED (resolved by Graphify, with or without a model), or AMBIGUOUS (flagged by the model pass). The rules also tell Codex to run graphify update . after modifying code, so the graph tracks its own edits.
Because Codex cannot hard-block a tool call, there is no strict mode here (that exists only on Claude Code, whose hook API supports a deny decision). On Codex, the instruction file is the mechanism, the same approach the OpenCode integration builds on.
Optional: a local MCP server in config.toml
This is for a graph you build on your machine, not the hosted server above. The skill plus the CLI is the complete local setup; nothing about it requires MCP. If you would rather expose that local graph to Codex as native MCP tools (query_graph, shortest_path, god_nodes, and more; see the MCP tools reference), register the server in ~/.codex/config.toml:
# ~/.codex/config.toml
[mcp_servers.graphify]
command = "python"
args = ["-m", "graphify.serve", "graphify-out/graph.json"]The stdio server ships as an extra: uv tool install "graphifyy[mcp]".
Troubleshooting
$graphifydoes nothing: confirm.codex/skills/graphify/SKILL.mdexists andAGENTS.mdcontains the graphify section, then start a new Codex session so both are picked up. And it is$graphify, not/graphify, on Codex.graphify: command not found: the uv or pipx tool bin dir (~/.local/bin) is not on PATH. Runuv tool update-shell(orpipx ensurepath) and open a new terminal.- Extraction feels slow on a big repo: check that
multi_agent = trueis set under[features]in~/.codex/config.toml; that is what allows parallel subagent extraction on Codex. - A partial install (say,
AGENTS.mdwritten but.codex/hooks.jsonmissing): just re-rungraphify codex install --project. It is idempotent and fills in whatever is missing without duplicating the AGENTS.md section. - Codex greps anyway: expected while
graphify-out/graph.jsondoes not exist. The rules only apply once a graph is built, so run$graphify .first.
From the blog
- How to switch AI coding assistants without starting over: How to switch AI coding assistants without starting over: what moves with you, what does not, and how to sign a new client in to your code knowledge graph.
- AGENTS.md and CLAUDE.md best practices: why agents follow stale rules, why reviews miss them, and how to check them against the code: AGENTS.md best practices for stale rules: why agents follow them, what a 2026 study found about reviewing them, and how to check each rule against your code.
- Graphify is now a ChatGPT plugin: The Graphify ChatGPT plugin connects ChatGPT, Codex, and Dots to repositories you have indexed, so they can trace calls and judge what a change reaches.
Codex FAQ
Do I need an API key to connect Codex to Graphify?
No. Install the Graphify plugin from the OpenAI plugin directory, or run codex mcp add and codex mcp login. Codex signs in with OAuth in your browser the first time it connects. There is no key to paste or store.
Why $graphify and not /graphify?
Codex uses $graphify instead of /graphify; that is its convention for invoking this kind of command. Everything the skill does afterwards is identical to the other assistants, and the graph it builds in graphify-out/ is the same one every other tool queries.
Does Codex get hooks like Claude Code does?
Not usefully, and Graphify is honest about that. The installer does register a PreToolUse entry in .codex/hooks.json, but the command it runs (graphify hook-check) is deliberately a no-op, because Codex Desktop rejects the additionalContext payload a nudge would need. The always-on AGENTS.md rules carry the graph-first guidance on Codex instead.
What is multi_agent = true for?
Graph extraction can fan out to parallel subagents, which is much faster on large repos. On Codex that requires multi_agent = true under [features] in ~/.codex/config.toml. Without it, extraction still works, just without parallel dispatch.
Does graphify codex install overwrite my AGENTS.md?
No. It replaces or appends only its own graphify section and leaves the rest of the file alone. Re-running the install refreshes that section idempotently, and graphify codex uninstall removes exactly that section plus the .codex/hooks.json entry.
Should I commit the AGENTS.md section?
Yes. AGENTS.md is a plain project file, so committing the graphify section means one install covers the whole team. Teammates need the graphifyy CLI installed and a built graph (graphify-out/graph.json) for the queries to work.
Does my code leave my machine?
With the open-source engine, no. Source code is parsed locally with tree-sitter (about 40 languages, 33 with tree-sitter grammars); building code edges never requires an API call. Non-code files such as docs and PDFs are read by the model you already use in Codex. The engine sends Graphify no telemetry and needs no account. If you use Graphify Cloud instead, it builds the graph from the repositories you connect.
Next steps
The quickstart covers the full CLI, and the MCP tools reference documents every tool the optional server exposes. The same graph serves every assistant you run: see Graphify + Claude Code, Graphify + Gemini CLI, or browse the full list on the integrations page.