Potpie takes a graph-first approach to codebase understanding: before any agent runs, it indexes every file, class, function, and their relationships into a Neo4j knowledge graph. When an agent answers a question or generates code, it navigates this graph rather than relying on keyword search or embedding similarity, giving it structural knowledge of how the codebase fits together.
Key capabilities
Knowledge graph indexing — Potpie builds a structural graph of the entire repository — files, classes, functions, imports, call relationships — using tree-sitter parsing. This gives agents precise navigation ability: "find every caller of this function" or "show all classes that implement this interface."
Specialised agent roles — Rather than a single general-purpose agent, Potpie provides purpose-built agents: a Debugging Agent for root cause analysis, a Code Generation Agent for feature implementation, a Test Generation Agent for coverage, and a Code Review Agent for analysis. Each is tuned for its task.
Custom agent builder — Teams can define their own agents with specific instructions, tool access, and output formats, deploying them org-wide for standardised workflows (e.g., an agent that always generates tests in the team's specific framework).
Human-review gate — All agent-proposed changes are surfaced as pull requests for human review before merging, keeping a human in the approval loop.
Autonomy level
Level 3 — Supervised agent. Agents generate and propose code changes autonomously but all modifications require human review and approval before they land in the codebase.
Strengths
- Knowledge graph gives genuinely structural understanding rather than semantic approximation
- Specialised agents outperform general-purpose agents on their specific tasks
- Apache 2.0 licence with active development (v1.1.1 May 2026)
- Self-hostable via Docker; also available as cloud service at app.potpie.ai
- VS Code extension available alongside self-hosted deployment
Limitations
- Setup requires Docker, Neo4j, and Celery — heavier infrastructure than simpler CLI tools
- Knowledge graph indexing time grows with codebase size
- 5.5K GitHub stars indicates a smaller community than leading tools
- Custom agent builder requires familiarity with the Potpie agent configuration format