Skip to main content
At a high level, Qodo operates through three interconnected layers: ingestion, knowledge storage, and agentic research — all surfaced through an MCP interface that integrates with your existing tools and agents. Qodo is built around a persistent, structured understanding of your codebase — not just the diff in front of it. The platform connects three layers. Ingest agents parse your repositories into structured knowledge. A rigid knowledge layer stores relationships, history, and embeddings. Deep research agents reason across that knowledge to answer queries and power code review. Everything is surfaced through an MCP interface, making Qodo’s capabilities composable with your existing agents, CI pipelines, and developer tools. Qodo platform architecture

​
Core components

​
Ingest agents

Before Qodo can reason about your code, it needs to understand it. Ingest agents crawl connected repositories and process them through two complementary lenses:
  • Parsing agents: Analyze source files at the syntactic level, extracting structure, interfaces, and relationships between components.
  • Abstract Syntax Tree (AST) chunking: Breaks code into semantically meaningful units using ASTs rather than arbitrary token windows. This preserves logical boundaries across functions, classes, and modules.
Supporting tooling handles auxiliary artifacts — configs, CI definitions, documentation — to give the system a full picture of how each repo fits together.

​
Rigid knowledge layer

The output of ingestion flows into a structured, multi-modal knowledge layer:
StoreWhat it captures
Graph DBRelationships between components — call graphs, dependency chains, cross-repo references
Vector DBSemantic embeddings for similarity search and fuzzy concept matching
PR & Commit historyHistorical context on how and why the codebase evolved
Auto-generated .mdStructured summaries of modules, APIs, and architectural patterns
This layer is what makes Qodo different from approaches that analyze diffs in isolation. Rather than evaluating a change against itself, Qodo evaluates it against the full structure and history of the system. Embeddings are generated using Qodo-7b (Qodo’s code-specialized model) and text-embedding-3-large, enabling both deep code comprehension and broad semantic retrieval.

​
Deep research agents

When a query arrives — whether from a developer, a code agent, or a CI pipeline — it is routed through an MCP server to a set of specialized research tools:
  • deep-research: Performs multi-step reasoning across the knowledge graph to answer questions that require understanding system-wide context. Used for architectural questions, impact analysis, and root cause investigation.
  • find-similar: Locates semantically or structurally similar patterns elsewhere in the codebase. Useful for detecting duplicated logic or finding prior art for a proposed approach.
  • deep-issue: Investigates specific bugs, regressions, or quality concerns with full traceability across PRs, commits, and dependent modules.
  • ask: A general-purpose interface for targeted questions about the codebase.
These agents share access to a Context Ranking / Judge layer that scores and filters retrieved evidence before it reaches the final response. This is what keeps output relevant and precise rather than noisy.

​
Multi-agent review system

On top of the research layer, Qodo runs a suite of specialized review agents, each responsible for a specific quality dimension:
  • Critical Issue Agent: Detects bugs, logic errors, and edge-case failures
  • Breaking Changes Agent : Identifies whether a change breaks upstream or downstream consumers
  • Ticket Compliance Agent: Validates that code matches requirements and acceptance criteria
  • Duplicated Logic Agent: Flags repeated code across repositories
  • Rules Agent : Enforces organizational coding standards and