Configuration

Configuration

CodeGraph is configured through three things: build-time feature flags, runtime environment variables, and on-disk config / output files. There is no dotenv loading; environment variables are read directly from the process environment.

Feature flags (build time)

All feature flags are off by default except the language extractors. Enable integrations at build time (see Installation):

Feature	Enables	Used by
pg	Postgres schema introspection	codegraph ingest pg
push	Live Neo4j/FalkorDB export	codegraph export neo4j|falkordb --push
office	Spreadsheet ingest	codegraph ingest office
gws	Google-Workspace ingest	codegraph ingest gws
media	Audio/video transcription and YouTube URL ingest	codegraph ingest media

Language support is controlled by 38 lang-* features, all on by default. To compile a single language (used in CI), build with --no-default-features --features lang-<name>. See Languages and Development.

Environment variables

LLM backend selection (semantic pass)

The semantic pass auto-detects a backend by checking these in order: Gemini, Kimi, Anthropic, OpenAI, DeepSeek, Azure OpenAI, Bedrock, Ollama. Set CODEGRAPH_BACKEND to force one. See Semantic Analysis.

Variable	Purpose
CODEGRAPH_BACKEND	Force a specific backend (the only way to select claude-cli)
CODEGRAPH_LLM_TEMPERATURE	Temperature override; none/omit/default omits the parameter
OPENAI_API_KEY, OPENAI_MODEL, OPENAI_BASE_URL	OpenAI-compatible backend
GEMINI_API_KEY / GOOGLE_API_KEY, GEMINI_MODEL	Gemini
MOONSHOT_API_KEY, MOONSHOT_MODEL	Kimi (Moonshot)
DEEPSEEK_API_KEY, DEEPSEEK_MODEL	DeepSeek
ANTHROPIC_API_KEY, ANTHROPIC_MODEL, ANTHROPIC_BASE_URL	Native Anthropic
AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_DEPLOYMENT, AZURE_OPENAI_API_VERSION	Azure OpenAI
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, AWS_REGION / AWS_DEFAULT_REGION, BEDROCK_MODEL	AWS Bedrock
OLLAMA_BASE_URL, OLLAMA_API_KEY, OLLAMA_MODEL	Ollama (local; presence opts in)
CLAUDE_CLI_MODEL	Model for the claude CLI backend

Server, ingestion, and database push

Variable	Purpose
CODEGRAPH_API_KEY	Bearer token for the HTTP MCP server (fallback for --api-key)
CODEGRAPH_QUERY_LOG	Path to write the server query log
CODEGRAPH_QUERY_LOG_DISABLE	Disable the query log (1/true/yes)
CODEGRAPH_CHANGED	Newline-delimited changed-file list read by update (used by the git hook)
NEO4J_PASSWORD, FALKORDB_PASSWORD	Credentials for export --push
CODEGRAPH_GWS_CMD	Google-Workspace CLI name (default gws)
CODEGRAPH_TRANSCRIBE_CMD	Transcription CLI (default whisper)
CODEGRAPH_WHISPER_MODEL	Whisper model (default base)

Other

Variable	Purpose
HOME / USERPROFILE	Locate the global store ~/.codegraph (falls back to .codegraph in the working directory)
CODEGRAPH_SKIP_HOOK	Skip the installed git hook for one invocation (1)

Config and output files

Path	Read/Written	Role
.codegraphignore	read	Extra ignore rules, layered per directory; takes precedence over .gitignore on conflicts
.gitignore	read	Honored during discovery
codegraph-workspace.toml	read/written	Workspace manifest ([workspace] members, [[repos]]); written by workspace init. See Workspaces and Federation
codegraph-out/	written	All output: graph.json, GRAPH_REPORT.md, visualizations, exports, ingested/, surfaces/
codegraph-out/cache/ast/	written	Per-file AST cache, keyed by content; auto-invalidated when extractor logic or enabled languages change. Clear with codegraph cache clear
codegraph-out/cache/semantic/	written	Semantic-pass response cache
~/.codegraph/	read/written	Global cross-repo store (global-graph.json, global-manifest.json)
.claude/settings.json	read/written	PreToolUse hooks installed by codegraph install (Claude). See Assistant Integration
CLAUDE.md / AGENTS.md / GEMINI.md and per-platform skill files	written	Assistant instruction sections written by install
.codex/config.toml / .codex/hooks.json (+ ~/.codex/config.toml)	read/written	Codex MCP server + SessionStart hook from codegraph install codex (project, or global with --global). See Assistant Integration

Notes

• A code-only extract never reads any of the LLM variables and never makes a network call. They matter only for extract --semantic and ingest.
• The AST cache lives under codegraph-out/, so deleting that directory resets everything.