An open-source daemon that governs, executes, and observes AI agents on your hardware. Explicit permissions. Process isolation. Full audit trail. No cloud.
| Agent | Status | Permissions | CPU | Mem |
|---|---|---|---|---|
| code-reviewer | running | fs:r mcp:git net | 2.1% | 48 MB |
| doc-writer | running | fs:rw mcp:md shell | 1.4% | 32 MB |
| data-analyst | idle | fs:r net:api mcp:sql | 0.0% | 24 MB |
| security-scan | running | fs:r net:local shell | 3.8% | 64 MB |
| dep-updater | stopped | fs:rw shell net | — | — |
OpenAI launched Frontier on Feb 5, 2026 — an enterprise platform managing AI agents as "AI coworkers." 57% of organizations now have agents in production (LangChain State of AI Agents, 2026). Every major vendor wants to be your agent runtime.
They share one assumption: your agents run on their cloud. Your data flows through their systems. Their policies govern your agents. OpenSentience inverts this — a supervised, local-first daemon where you define the rules.
Palo Alto Networks, Token Security, and AccuKnox all confirm: agents need runtime identity, least privilege, and egress controls. "Governance has to operate at the same speed as the agents themselves." OpenSentience enforces at runtime.
Launched to ensure AGI remains open-source and decentralized. OpenSentience embodies this at the infrastructure layer.
An OTP supervision tree that manages agent lifecycles, permissions, and inter-agent communication. Every agent is a supervised BEAM process.
:rest_for_one
strategy. One agent crash doesn't bring down the
system.
Elixir 1.17+
OTP 27
~/.opensentience/
— portable, inspectable, yours.
Fine-grained, per-agent capabilities. Filesystem, network, shell, MCP tools — nothing auto-granted. Users approve before activation.
Every agent runs as a supervised BEAM process. Memory-isolated, fault-tolerant. One rogue agent can't crash others or read their state.
Every tool call, file access, and network request logged. Full traceability, queryable via MCP or admin UI.
Agents expose and consume MCP tools. Connect to Claude Desktop, Cursor, VS Code. The protocol IS the API.
Install agents with one command. Permission review before activation. Community trust scores and audit history.
Agents connect to Graphonomous for continual learning. Each agent gets its own knowledge graph that grows with use.
The landscape is full of agent runtimes. None combine local-first execution, explicit permissions, and MCP-native governance.
| OpenSentience | OpenAI Frontier | Salesforce AgentForce | Goose (Block) | CrewAI | |
|---|---|---|---|---|---|
| Runtime | Local-first | OpenAI cloud | Salesforce cloud | Local | Cloud / local |
| Permissions | Explicit, fine-grained | IAM-based | Role-based | None | None |
| Process isolation | BEAM VM | Sandbox | Shared | Single process | Python threads |
| Protocol | MCP native | OpenAI API | Proprietary | MCP | Custom |
| Audit | Built-in, queryable | Logs | Salesforce audit | None | None |
| Lock-in | None (MIT) | OpenAI | Salesforce | MIT | Minimal |
| Continual learning | Graphonomous | Feedback loops | None | None | None |
| Price | Free | Enterprise $$ | Enterprise $$ | Free | Free / Paid |
OpenSentience is the runtime layer of the [&] Ampersand Box portfolio.
Discover, install, and distribute agents.
Self-evolving knowledge graphs for agents.
Spec-driven development. SPEC.md files.
Enterprise builder. Test → Deploy → Govern.