An AI agent for the enterprise. Open source. Free.
Maia is the enterprise edition of Hermes Agent, maintained by Ampliia, to operate agents, automations, cron jobs, skills, memories, and integrations in enterprise environments, with hierarchy, sensitive data, auditability, and human approval.
curl -fsSL https://ampliia.com/maia/install.sh | bashWhat Maia is for.
Traditional Hermes is excellent for individual productivity and local agents. In companies, the problem changes: who can edit global instructions, which files each team can access, which automations need authorization, and how to audit what happened.
Role-based administration
Most users talk through the corporate gateway, but only after receiving an explicit Governance role: allowlists, pairing, and “allow all” never grant bot access by themselves. The dashboard remains the protected admin surface.
Layered knowledge
Corporate memories and skills apply to governed conversations. Teams have their own context. Users keep personal preferences without contaminating the organization.
Human-in-the-loop
Shared knowledge and sensitive commands can wait with a diff and approval controls. For protected file edits, Maia changes nothing, names a selected manager or administrator, and continues naturally when that person responds in the shared conversation.
Secure administration in chat
An authorized admin can ask Maia in the messaging channel to admit users, assign roles and teams, or change file access. Identity comes from the authenticated message and Governance is checked again for every action; managers remain limited to delegated roots.
Practical architecture.
The architecture separates identity, knowledge, execution, and audit. The goal is to enable automation without losing administrative control.
Entry and identity
Governed core
Operations and audit
What happens when someone asks Maia to do something.
Every message (a question about a document, an edit, a command) is resolved against the sender's identity and checked against policy before anything happens. Nothing runs on trust alone, and every decision is recorded.
How a team actually works with Maia.
People talk to Maia in the channels they already use. The same assistant answers questions about company documents, edits files inside the boundaries each person is allowed, and routes anything sensitive to the right human. The reasoning below comes from real, governed scenarios.
/company/legal/enterprise-msa.pdf, which your role can read. Anything outside those documents I'd have to ask legal for.@Carla, the selected manager, in this shared thread. Carla can approve, reject, ask a question, or revise the request in natural language; Governance rechecks her authenticated turn before writing.Unchanged · review with @Carla@tech-lead to allow it; you'll see it run the moment they accept, and the whole thing lands in the audit log.Command · awaiting @tech-leadOnce provisioned in Governance, each person gets a capable assistant in their own channel, with personal conversation, memory, and skills isolated by platform identity. What changes per person is what Maia may read, change, and run on their behalf, and who signs off when it matters.
Who can do what.
Roles map to a hierarchy your admins define. Higher roles inherit lower ones; approvals and configuration stay narrow. This is an illustrative default: every cell is configurable per tenant.
| Capability | Viewer | Operator | Manager | Admin |
|---|---|---|---|---|
| Ask about documents they're cleared to see | ||||
| Edit files inside their team's folders | ||||
| Change an approval-gated folder | Review | If selected | ||
| Approve file changes & flagged commands | ||||
| Configure tenant, roles, folders & keys |
Review means the person has conditional write access, but Maia changes nothing until a selected manager or administrator continues the request under their own authenticated identity. Authenticated administrators have global file access.
How it differs from traditional Hermes.
Maia preserves the local agent experience while adding the controls required for multiple people, teams, and responsibilities.
From individual to organizational
Personal preferences, memories, and skills remain available and isolated per user, while corporate and team knowledge goes through scope, approval, and audit.
End-to-end file governance
Governance is always active. Adding a person and role enables bot use but grants no files until an explicit policy matches. File tools enforce that policy, while gateway terminal, code execution, and delegated agents run through Docker or Podman with only granted paths mounted. Without that boundary Maia stays safe in Restricted mode: chat and governed file tools remain available, while command automation is refused with no host fallback. Protected writes can require a named manager or administrator to continue the exact planned edit in the shared conversation.
From local automations to approved routines
Cron jobs and recurring actions can require authorization by role or user before execution.
Maia is maintained by Ampliia as a corporate distribution derived from the MIT-licensed Hermes project. The original Nous Research project is credited in the repository and license. Maia's source code is publicly available through the GitHub project linked on this page.
Model agnostic. The LLM is your choice, not a lock-in.
Maia works with whichever model your company chooses. Switching models is a configuration change, not a rewrite of your workflows. The same conversations, skills, memories, approvals, and policies keep working when the provider changes.
Major cloud APIs
Connect the leading providers (Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, xAI, and aggregators like OpenRouter), plus any endpoint compatible with the OpenAI API. Keys live in the managed credential store, outside prompts and memories.
Local models, on your servers
Run with Ollama, LM Studio, or any compatible inference server inside your own infrastructure. No data leaves the company: maximum privacy for sensitive documents and a natural fit for LGPD/GDPR, including air-gapped networks.
No vendor lock-in
Multi-provider fallback keeps operations running if an API goes down, blocks the account, deprecates a model, or changes pricing. Governance, audit trails, and corporate knowledge are yours, independent of any single AI lab.
Governance, security, and reference standards.
The platform is designed to align with recognized practices, without claiming automatic certification. The value is turning those standards into concrete product controls.
NIST AI RMF
Risk mapping, monitoring, governance, traceability, and improvement cycles for AI systems.
ISO/IEC 42001
Management controls for AI systems, with clear roles, documentation, evaluation, and human review.
ISO 27001 and SOC 2
Separation of duties, access control, logs, retention, review, and secure operations.
OWASP LLM Top 10
Guardrails against tool abuse, data leakage, improper execution, and unauthorized changes.
LGPD and GDPR
Minimization, purpose limitation, access separation, and care for personal data in memory, logs, and integrations.
Operational audit
Important events are recorded for review: denials, approvals, imports, cron jobs, dashboard actions, and knowledge changes.
Knowledge, skills, and teamwork.
The separation between levels prevents a personal preference from becoming a global policy and lets teams work with their own context without compromising the whole company.
Corporate
Memories, rules, and skills applied to every governed conversation. They only enter or change through authorized hierarchy and human approval.
Team
Knowledge and skills for departments, squads, or business units. Scope is limited to members and folders allowed for the team.
User
Individual preferences and shortcuts that do not alter other people's behavior or create new permissions.
Read the documentation before deployment.
The documentation covers administrative onboarding, Hermes export migration, memory and skill governance, file-change approvals, dangerous-command approvals, cron jobs with authorization, observability, and secure operations.