Agent Infrastructure
Production-grade tools for orchestrating, automating, and securing AI agent workloads. From multi-agent coordination to safe database access.
Coordinate Claude Code, Aider, Codex, and other AI coding tools working in parallel on shared codebases. Built on an append-only event log for crash-safe, deterministic state management.
- Crash-safe append-only event log (Grite)
- Real-time web dashboard at localhost:5173
- Merge management with configurable policies
- Mayor AI orchestrator for codebase analysis
- Convoy grouping for sprints and epics
- Witness process for agent monitoring
$ cargo install bratGo-based workflow automation platform delivering 5-10x faster execution than n8n with 70% less memory. Runs existing n8n workflows without modification.
- ~100ms avg execution (vs n8n ~500ms)
- ~150MB memory (vs n8n ~512MB)
- 35+ node types: HTTP, DB, AI/LLM, cloud
- Full n8n expression syntax support
- CLI agent orchestration with sandboxing
- SQLite, PostgreSQL, or in-memory storage
$ curl -fsSL https://raw.githubusercontent.com/neul-labs/m9m/main/install.sh | bashOne command-line interface to manage multiple AI coding agents. Switch between tools, share context, and maintain consistency across your agent fleet.
- Unified interface for all coding agents
- Context sharing between agents
- Session management and history
$ cargo install ringletORM-native capability runtime that transforms existing SQLAlchemy, Tortoise, and Peewee models into secure, typed tool surfaces for AI agents.
- Policy engine with field-level access control
- Masking, hashing, and redaction built-in
- Automatic tenant isolation
- Query budget enforcement
- Immutable audit logging
- Native support for SQLAlchemy, Tortoise, Peewee, Django, SQLModel
$ pip install ormai[sqlalchemy]How the four pieces fit together
These are not four views of one product. They sit at different points in an agent deployment and are meant to be adopted separately. brat owns coordination: it decides which coding agent gets which task and keeps the record of what happened. m9m owns scheduled and triggered work that is not conversational at all — the webhook, the nightly job, the pipeline with thirty nodes. ringlet owns the credential and cost boundary around whichever agent CLIs a team already runs. ormai owns the data boundary, which is the only one of the four that is a security control rather than an operational one.
The dependency worth knowing about is that brat stores its state in the grite append-only event log rather than in a database it owns. That is why a crashed run can be replayed rather than reconstructed, and it is also why brat has no server to operate. If you adopt brat you are implicitly adopting that log, which lives in the repository alongside the code.
Choosing between brat, m9m and ringlet
The three are confused for one another often enough that the shortest distinction is worth stating directly: brat runs agents against a codebase, m9m runs workflows against systems, and ringlet runs the agent binaries themselves.
| If your problem is | Start with | Because |
|---|---|---|
| Several coding agents stepping on each other in one repo | brat | Task assignment, merge policy and crash-safe state are the whole product |
| A workflow engine that is too heavy for the box it runs on | m9m | Single Go binary, n8n-compatible nodes, no external infrastructure |
| Four agent CLIs, three providers and no idea what they cost | ringlet | Isolated profiles plus one token-level ledger across all of them |
| An agent that needs to read production data | ormai | Field redaction, tenant scoping and query budgets are enforced, not prompted |
| All of the above, eventually | ormai first | The data boundary is the one that is expensive to retrofit |
If you are still deciding what belongs in an agent platform at all, the reference architecture in the agentic infrastructure stack walks through the layers in order, and database access patterns for AI agents covers the four ways teams usually wire the data boundary.
Adoption notes and limitations
Each project installs independently and none of them requires the others. That is deliberate — a team should be able to take the data boundary without adopting an orchestration model, or run a workflow engine without changing how its coding agents are launched.
Two honest constraints. First, the performance figures quoted on this page were measured on specific workloads and versions described in each project repository; they are not a general claim about your graph, your nodes or your hardware. Second, none of these tools makes an agent correct. They make its failures bounded, recoverable and auditable, which is a different and smaller promise.
Frequently asked questions
Does brat replace LangGraph or CrewAI?
No. brat coordinates coding-agent processes such as Claude Code and Aider; LangGraph and CrewAI are libraries for building an agent inside your own program. The distinction is set out at brat vs LangGraph.
Can m9m run my existing n8n workflows?
That is the design goal, and the node coverage is documented in the repository. Verify your specific node set against the current release before migrating anything you depend on; compatibility claims are version-specific.
Is ormai a replacement for database grants?
No, and treating it as one is a mistake. It is a second layer above least-privilege credentials, not a substitute for them.
What does this cost?
The projects are open source and free to run. The recurring cost of an agent platform is model spend, which is a separate discipline covered in cost control for production AI agents.