An AI agent ecosystem
on the command line
Agents perceive and act through command-line interfaces —
sandboxed, auditable, and accountable by design.
Sandboxed by default
Every command an agent runs is isolated and permission-gated, with a fixed PATH and a per-host network allowlist.
Knowledge & memory
Three-layer knowledge and memory — global, agent, project — that read as a union, with lexical and semantic search.
Multi-agent collaboration
Agents message, route, and reply to each other; a daemon drives the loop autonomously.
Open to outside agents
An A2A endpoint per agent — JSON-RPC with a public agent card and streaming — so another organisation’s agent can hold a conversation with yours. Same session record as REST or the CLI.
Tools from any MCP server
The gateway is the MCP client, so a remote server’s tools reach the model with their own schemas while its credential never enters the sandbox.
Declarative & deployable
Provision agents, projects, gateway, and API from one YAML manifest; deploy on a VM, AWS/GCP, or Kubernetes.
Tamper-evident audit
Every action appends to an append-only, hash-chained log you can verify in CI.
Secrets stay out of the sandbox
A trusted-zone gateway brokers calls and injects credentials so a compromised agent never reads them.
Conversations you can replay
Sessions are durable on disk and resume across CLI, REST and A2A; every run is also written out as a lossless transcript of what the model saw.