FAQ

FAQ

Is SourceBridge open-source?

Yes. SourceBridge is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). The source is at https://github.com/sourcebridge-ai/sourcebridge.

What languages does it index?

Ten languages via tree-sitter: Go, Python, TypeScript, JavaScript, Java, Rust, Ruby, PHP, C++, and C#. See internal/indexer/languages.go for the full list of extraction queries per language.

Does it require a cloud LLM?

No. Indexing and graph traversal (search, callers, callees, requirements list, impact report) work without any LLM. Generation features (cliff notes, code tours, QA, living wiki) require an LLM, but that can be a locally-run model via Ollama, vLLM, llama.cpp, SGLang, or LM Studio — no internet connection or API key required.

Fully air-gapped: run SurrealDB embedded, use Ollama with a local model, and you have a self-contained instance with no outbound dependencies except for your own repositories. See docs/self-hosted/air-gapped.md.

What's the difference between the knowledge engine and living wiki?

• Knowledge engine produces on-demand field guides (cliff notes, code tours, learning paths, workflow stories, architecture diagrams) when you click Generate in the web UI or call the MCP tool. The artifacts are stored and reused until stale.
• Living wiki runs on a schedule and publishes a maintained wiki to an external sink (Confluence, Notion, or a git branch). It uses block-level reconciliation to preserve human edits across regeneration cycles. Think of it as the difference between "generate when I ask" and "keep this wiki current automatically."

How does living wiki handle human edits?

Block-level reconciliation. Every page is a tree of typed blocks with stable IDs. Blocks you edit (in a PR review, directly in Confluence, or in Notion) get ownership state human-edited. SourceBridge leaves those blocks alone on future regeneration passes. Only blocks owned by SourceBridge (generated) are updated when the source code changes. See Living Wiki for details.

What is the MCP server, and how many tools does it expose?

The MCP server exposes 23 tools for AI agent integration. They cover symbol search, code explanation, QA, call graph traversal, requirements, indexing lifecycle, compound workflows, and subsystem clustering. Clients like Claude Code, Claude Desktop, and Cursor can call these tools directly to get grounded, repository-aware answers. See AI Clients and MCP.

What is subsystem clustering?

Label-propagation clustering (Raghavan et al. 2007) over the call graph. It runs automatically after each index and groups symbols into named architectural subsystems. The clusters drive the Subsystems tab in the web UI, three MCP tools (get_subsystems, get_subsystem_by_id, get_subsystem), and the living wiki's page taxonomy. The sourcebridge setup claude command writes per-cluster sections into .claude/CLAUDE.md so Claude Code understands the codebase's architecture before it starts editing.

How do I connect Claude Code to SourceBridge?

sourcebridge setup claude --repo-id <id>

This writes .mcp.json (MCP server registration) and .claude/CLAUDE.md (per-subsystem skill card) into the repository. Claude Code picks them up automatically. See the AI Clients and MCP page for manual configuration instructions.

Do I need requirements to get value?

No. Requirements make traceability, coverage gap detection, and change-impact analysis much more powerful, but all the generation, search, QA, and clustering features work without them.

Can multiple people use the same SourceBridge instance?

Yes. The data model has a tenant_id column on all major tables. Operator-level access is managed through JWT auth and OIDC SSO. Team management and role assignment are enterprise features.

What's the difference between OSS and enterprise editions?

The OSS edition includes all 23 MCP tools, the living wiki, subsystem clustering, QA, field guides, requirements, and OIDC SSO. Enterprise adds: cross-repo impact analysis, audit log, SAML, team management, org-level settings, notification hooks, per-operation model overrides, and long-form enterprise report generation.

Edition is set via SOURCEBRIDGE_SECURITY_MODE (oss or enterprise). The capability registry at internal/capabilities/registry.go is the definitive list of what each edition offers.

How do I contribute?

Read CONTRIBUTING.md in the repository for the development setup, coding standards, and PR process. First-time contributors must agree to the Contributor License Agreement before their PR can be merged.

Quick development setup:

git clone https://github.com/sourcebridge-ai/sourcebridge.git
cd sourcebridge
make build       # builds Go binary + web
make dev         # starts API server
make dev-web     # starts web UI (separate terminal)
cd workers && uv run python -m workers   # starts Python worker
make test        # run all tests
make lint        # run all linters

Do upgrades preserve data?

The migration runner (42 migrations as of this writing) applies new migrations automatically at startup and skips already-applied ones. Minor version upgrades are generally data-safe. Check CHANGELOG.md before upgrading across multiple releases for any notes on schema changes that require attention.

For development-stage installs where the data is expendable, wiping the SurrealDB volume and re-indexing is a valid recovery path.

Where does telemetry go?

Anonymous install telemetry is sent to https://telemetry.sourcebridge.ai/v1/ping at startup. It includes version, platform, provider kind, and feature flags — no code, file paths, or user data. Opt out with SOURCEBRIDGE_TELEMETRY=off or DO_NOT_TRACK=1. Full disclosure in TELEMETRY.md.