Add agent teams and memory to Claude Code and Codex with an execution layer
Running a single command in your project folder turns AI coding tools into managers that form teams, distribute tasks, and verify results. Ruflo is a harness that wraps tools like Claude Code and Codex to handle inter-agent coordination, memory management, and security checks. Once installed, you can observe agents forming swarms to identify risky modules and write tests when you assign real work.
How the agent team operates
Ruflo includes over 100 specialized agents that work together in specific domains such as coding, testing, security, and documentation. When a user assigns a task, a router selects the appropriate agents and organizes them into a swarm to distribute the work. The experience generated during this process is stored in memory to improve the accuracy of subsequent tasks. It also provides integration features for secure communication with agents on other machines or in other organizations, verifying identity via mTLS and ed25519 authentication, automatically removing personally identifiable information (PII) before transmission, and evaluating trustworthiness based on behavior-based scores.
Two installation paths
Ruflo offers two installation methods that can be selected based on the intended use.
- Memory system It uses AgentDB with HNSW indexing to perform vector search. Search speeds were measured to be approximately 1.9 times faster than brute force under N=20k conditions and approximately 3.2 to 4.7 times faster under N=5k conditions.
- Learning features It learns from past successful cases through SONA neural patterns and ReasoningBank. The results of this learning vary depending on the quality of the configured memory and feedback.
Performance comparison and verification
Ruflo publishes benchmark results compared with LangGraph, AutoGen, and CrewAI. Tests of version v3.8.0 in darwin-arm64 and linux-x64 environments showed advantages ranging from 1.3 to 1953 times in cold start, single-turn processing, and RSS (physical memory usage). Additionally, the ruflo verify command can cryptographically prove that installed bytes match a signed witness, a procedure to confirm safety from supply chain attacks.
Pre-application checks
Ruflo is released under the MIT license and is written in TypeScript. The Claude Code plugin method requires Claude Code version 2.1.287 or higher, and version 2.1.292 or higher when installing terminal-based plugins. Since the plugin runs with the user's account privileges without a sandbox, code review is essential before installation. In particular, in-process modes like ruflo-mods perform prompt routing and policy enforcement, so security settings must be carefully inspected. The scope of operation for learning and integration features varies depending on the connection status with external LLM providers (such as Claude, GPT, and Gemini), so it is recommended to first check compatibility with your infrastructure environment.
