Code intelligence

raph links symbol references across your codebase so an agent can ask “where is this global read or written?” and get an answer from the graph instead of guessing. Resolution comes in tiers — the higher the tier available for a language, the more accurate the USES / MUTATES edges.

The tiers

Tier	Languages	How
Compiler (Go)	Go	go/types — always on, type-accurate.
Compiler (SCIP)	any with an installed indexer	An external compiler-backed indexer (e.g. scip-typescript uses tsc, scip-python uses pyright). Cross-file accurate.
Import-aware (bundled)	JS/TS/JSX/TSX, Python	Pure-Go fallback that resolves references through import bindings. No install, offline.
Within-file (bundled)	every other tree-sitter language	Symbol + reference extraction within a single file.

Every language always gets at least the bundled tiers — installing a resolver only raises precision, it is never required.

How a full index resolves references

After the file walk, Run runs the tiers in order, each only raising precision for the languages it owns:

1. linkGoUsages — go/types for Go.
2. runSCIP — installed compiler-grade indexers; for a covered language the within-file pass is skipped so its edges are authoritative.
3. linkImportAwareUsages — the pure-Go import-aware fallback.
4. The within-file tree-sitter pass handles everything else (run during the walk).

The symbol index these passes resolve against is built once and shared, and each pass writes its edges in a single batched transaction.

See what you have

raph code-intel

Code-intelligence resolvers (compiler-grade tier)
  typescript   scip-typescript  installed (/usr/local/bin/scip-typescript)
  python       scip-python      not installed
               raph code-intel install python   (pip install scip-python)
  ...

After an index, raph init reports the same: which languages got compiler-grade resolution and a one-line raph code-intel install <language> for any that could be upgraded.

Upgrade a language

raph code-intel install python   # installs scip-python
raph init --path .               # re-index to use it

Use --dry-run to print the command without running it. The install shells out to the language’s package manager, so the matching runtime must be present:

Language	Installs via	Requires
typescript / javascript	npm i -g @sourcegraph/scip-typescript	Node.js
python	pip install scip-python (or npm)	Python+pip (or Node.js)
rust	rustup component add rust-analyzer	rustup
ruby	gem install scip-ruby	Ruby+gem
java	coursier (scip-java)	JVM
c / c++	prebuilt scip-clang release	manual download

If a prerequisite is missing, raph code-intel install tells you exactly which tool to install first. raph does not bundle these indexers — they are large native/runtime programs in other languages, and embedding them would bloat the binary and require license review.

For agents

The MCP index_codebase result carries scip_active (languages resolved compiler-grade) and scip_suggestions (languages that could be upgraded, each a SCIPSuggestion with an agent_action command).

Protocol: an agent must ask the user for permission before running an install command. If the user declines, the agent hands them the command to run themselves and continues with the bundled resolver. Agents never install unattended.

Disable it

Set RAPH_NO_SCIP=1 to skip the compiler-grade tier entirely and rely on the bundled resolvers (useful in sandboxes or offline CI).