Skip to content

Spiron · Always-on AI agents

Work moves on.Your agent stays.

A standing crew of agents with their own session, memory, and tools. Hand off follow-ups from Telegram, the web app, or the spiron terminal—and get back a clear outcome you can check.

DoneExample outcome · not a recorded run

Compare a schedule change against saved plans

Input
Updated travel notice · saved calendar notes
Conflict
14:00 arrival overlaps Thursday pickup
Prepared
Reschedule note drafted (held for approval)

✓ doneYour call before anything is sent

Illustrative task structure. Actual tools, channels, and approvals follow your configuration.

01 / How it works · Continuity

A task ends.
The agent doesn’t.

A Spiron agent is an ongoing participant, not a fresh process wrapped around each task. It carries its own conversation and memory, and can collaborate with other agents.

Always-on does not mean always computing. Idle compute can be released while durable context remains. Continuity depends on the configured storage and runtime—not a model promising to remember.

Understand agents and sessions
One agent runtimeArchitecture · not live activity
Entry pointsTerminal · Web · ChannelsMessages and ongoing work
Living agentSession + memoryContext belongs to the agent
Granted capabilitiesTools for the workAccess checked outside the model
Shared Rust capability core
SkillsMemoryBrowserComputerDeviceWorkspaceWeb
Local adapter

MCP / SQLite / Host

Cloud adapter

Multi-tenant sandbox / Postgres

One shared core, two thin adapters. Read the architecture and check implementation status. Availability varies by adapter and configuration.

02 / Memory, with depth

Keep the essentials.
Recall the original.

Three layers separate a small working index from deeper knowledge and the conversation it came from.

  1. L1

    MEMORY.md

    A compact index. What matters, and where to look next.

    Orient
  2. L2

    Topic files

    Focused knowledge, loaded when the work calls for it.

    Understand
  3. L3

    Raw transcripts

    Return to the original conversation instead of trusting a summary.

    Verify context

Memory helps an agent find context. It is not evidence that an external action succeeded. Read the memory model.

03 / Try something inspectable

Start small.
Check the result.

Use an installed Spiron CLI in a disposable project checkout. Ask for a result you can compare with the files in front of you.

Follow the quickstart
TerminalExample input · not a recorded run
# Inspect the installed command first
spiron --help

# Ask a focused question about your project
spiron -p "Read this project. Identify its
main entry points and test commands. Cite
the files you used. Do not edit anything."
Then inspect the cited files and your Git diff. A prompt is not a write-permission boundary.

05 / Plans

A clear starting point.

Choose a plan from the published price list. Check included usage and provided resources before you upgrade.

Compare plans and limits

Your next step

Give it a project.
Keep your bearings.