A team of coding agents
Ruflo helps developers divide work among agents, share useful context and coordinate a coding workflow.
Ruflo vs Orchestrator Zero
oz0 records agent jobs running on your infrastructure. Ruflo starts with coordinating agent teams in developer workflows. oz0 starts with operating agent jobs across a fleet, with a shared server and Edge gateways checking supported actions.
What Ruflo is
Ruflo, formerly Claude Flow, coordinates agent teams around tools such as Claude Code and Codex. It combines shared memory, model routing, workflows and communication between installations.
github.com- Runs
- Self-hosted or managed
- Licence
- MIT
- Checked
- 29 October 2026
Reviewed 29 October 2026 at release v3.63.9 (7f6fb33). Source and documentation review; runtime behavior and benchmarks were not independently tested. Ratings distinguish the core, host integrations and optional services.
What are you trying to run?
Agent jobs across your systems
oz0 helps teams run jobs across their machines, with shared access rules, approvals, model budgets and job history.
Where rules take effect
Both have real controls with coverage limits. Ruflo combines host checks and MCP policy. oz0 checks supported calls through its Edge gateway; some external coding tools act outside those checks.
Read the sourceWhat happens after an interruption
Ruflo persists workflow progress and can pause for human review. oz0 uses Temporal to retain completed steps; an interrupted step still follows its agent’s recovery rules.
Read the sourceSide by side.
Ruflo's ratings come from its own pages. A column's name opens its definition and how Orchestrator Zero does it. How we rated
Where it runs
Any machine you own
Yes
DocsYes: Your own hardware
Run the CLI, MCP server and agent hosts on your hardware; federation connects installations.
Linux, macOS and Windows
Yes: Native CLI installation
Native CLI setup is documented for Linux, macOS and Windows. The newer terminal console has a narrower tested platform list.
Self-hosted control plane
Yes
DocsYes: Local core and federation
The core and federation run locally. Optional hosted AI Team and managed-agent integrations have separate deployment boundaries.
No inbound ports
Yes
DocsConditional: Local stdio; network peers
Local stdio needs no inbound port. Network federation exposes a peer listener; outbound-only workers are not a platform-wide guarantee.
Fleet operations
Yes
DocsConditional: Federation peer management
Federation supports peer joining, removal and trust management. These sources do not establish the full machine-label, work-routing and host-upgrade lifecycle.
Durability
Durable runs
Conditional: Persisted workflow progress
MCP workflows persist progress and support resume. Autopilot parks interrupted steps for a person; recovery differs between execution paths.
Governance
Tool policy outside the model
Conditional: Host and transport coverage
Claude Code mods enforce tool checks; MCP policy gates cover stdio. The audit documentation explicitly excludes HTTP/WebSocket dispatchers from that enforcement.
Approvals in chat
Conditional: Human review gates
Workflows have human review gates and hosts can request approval. Durable approval of risky tool calls through a separate chat app is not established.
Tool snapshots
Yes
DocsNot documented
Plugin trust checks are documented, but automatic blocking and reapproval when a tool description or schema changes is not established.
Tenants built in
Yes
DocsConditional: Separate AI Team service
AI Team derives tenants from verified tokens and separates stored records. This is a separate service, not tenancy across every Ruflo execution path.
Cost
Budgets that stop a run
Conditional: New subagents can be stopped
A configured hard stop prevents new subagents at the session cost limit. It does not establish cancellation of every running action or uniform spend enforcement.
Security
Provider keys off the workers
Yes
DocsNo: Keys in runtime config
The provider manager accepts API keys in the runtime configuration. It does not require all provider credentials to remain in a separate gateway.
Sandbox with an egress allowlist
Conditional: Depends on execution path
Local WASM agents have no host filesystem or network access. Claude Code mods run with account permissions; isolation is not uniform across paths.
Audit and compliance
Tamper-evident audit trail
Conditional: Selected receipt chains
Autopilot adaptation checks receipt hash chains. MCP audit storage explicitly lacks tamper evidence, so coverage depends on the action and record type.
Erasure by key
Not documented
Memory deletion and service privacy controls do not establish per-person payload-key erasure with handling for restored backups.
Retention with legal hold
Conditional: Bounded local audit logs
MCP audit logs have configurable size and file-count limits. Legal holds and age-based retention are not documented for those records.
Openness
Model provider choice
Yes
DocsYes: Multiple vendors and Ollama
The provider system supports multiple vendors and local Ollama models, with routing and failover.
MCP and A2A
Yes
DocsConditional: MCP; A2A discovery only
MCP is supported. The A2A adapter exchanges discovery cards; its source explicitly excludes A2A tasks and messaging.
Where Ruflo is stronger.
Coordinate a coding team
Workflows divide development tasks among agents and combine their results, with both persisted MCP workflows and native host scripts.
github.comReuse useful context
A shared memory system combines semantic and keyword search, persistent indexes and learning from previous work.
github.comConnect separate installations
Federation provides peer discovery, trust management and communication between agent installations.
github.com
Differences to examine.
Compare the scope, conditions and editions in the source notes.
Find your first task.
Explore an example or read the public Docs preview. Software access is coming soon.