Checking live status…

The company that never sleeps.

Shiva Agents is a native platform where always-on AI agents work as a company. Describe a role in plain words, hire from a library of production personas, wire them into durable visual workflows, and let them run around the clock, with every risky action gated, approved and logged.

Runs on your GPU or any OpenAI-compatible modelZero npm runtime dependenciesWeb + Android

Collections cycle 0 10 * * 1,3,5 · IST
  1. Schedule firedMon, Wed and Fri at 10:00 IST trigger
  2. Priya · Collections SpecialistDrafted polite reminders by age bucket; handed anything past 60 days to the controller agent
  3. Approve remindersNothing is sent until you say so. The run waits, for days if it has to. waiting
  4. Log to company memorySo next week’s cycle knows who already paid memory
Every step checkpointed Survives restarts Fork from any step

Live from this instance

Checking…

Agents on staff
—
Active agents
—
Live workflows
—
Runs today
—
Tokens today
—
Approvals waiting
—

Read from /api/public/stats every 30 seconds while this tab is visible.

Three ways in

Build an agent. Hire a department. Wire a company.

Most tools give you one of these. Shiva Agents gives you all three on one runtime, so the agent you describe in the morning can be on a schedule, in a team and under governance by lunch.

01 · Agent Studio

Describe it. The platform drafts it.

Write what you need in plain language and get a working agent: persona, tools, memory, autonomy and a schedule, ready to edit.

  • Persona with role, goal, backstory and style
  • Autonomy from Observe to Autonomous (levels 0–3)
  • Heartbeats, standing orders (Sankalpa) and a verifier
How agents are built
02 · Agent Library

Hire production personas in one click.

85+ agents across 16 departments, from GST to growth to incident command, each with tools, KPIs and a sensible schedule.

  • 9 department-in-a-box blueprints
  • 20+ runnable workflow patterns
  • English and Hindi content teams
Browse the library
03 · Visual workflows

Wire teams with patterns that already work.

A canvas of triggers, agents, branches, verifiers and approvals, plus a Team node that runs the multi-agent topologies proven in research.

  • Cron in IST, webhooks with idempotency, company events
  • Assembly lines, debates, crews, handoffs, ledgers
  • Human approvals and questions as first-class steps
See the node catalogue
What’s inside

A runtime built for work that doesn’t stop at 6 pm.

Always-on company

Agents with inboxes, heartbeats and standing orders.

Every agent has an inbox (Karya), wakes on its heartbeat, follows its standing orders and delegates to or consults its coworkers. Idle agents sleep in the database and wake on events, so a thousand of them cost nothing while they wait.

Durable execution

Every step checkpointed. Nothing lost on restart.

Runs survive restarts, approvals can wait for days, and you can fork a run from any step to try again. Retries, timeouts, loop guards and token and step budgets come built in.

12 3 · fork4 5 replay with an edit
Memory

Agents remember who you are and what worked.

persona
owner
company
scratchpad

+ episodic · semantic · procedural recall, ranked by relevance × importance × recency; consolidated nightly

Governance

Autonomy you can dial, per agent.

Filled: runs on its own. Empty: waits for your approval. No level covers critical.

Open protocols

Every agent is an API.

  • POST/v1/chat/completions
  • POST/v1/tasks
  • MCPserver + client
  • A2A/.well-known/agent-card.json
  • SSElive run and token stream
Models

Your GPU, any OpenAI-compatible provider, or offline.

Runs on the AskShiva GPU through Ollama, or on any OpenAI-compatible provider you configure. With nothing reachable, a clearly labelled simulator keeps flows executing.

  • Gemma 4 · 26Bgemma4:26b-a4b-it-qat · writing in English and Hindidefault
  • Qwen 3.6 · 35Bqwen3.6:35b · reasoning and visionreasoning
Apps

Design on the web. Follow along on Android.

The web dashboard is where you design agents and workflows. The Android app keeps the same company in your pocket, so briefings and notifications reach you away from your desk.

Web dashboardavailable
Androidavailable
iOSnext
Patterns from every framework

The best multi-agent ideas, as one node.

Research labs and open-source frameworks each proved a way for agents to work together. The Team node runs all of them, so you pick the topology that fits the job instead of the framework you happened to install.

  • Assembly line SOP

    Roles hand typed documents down a line: PRD, design, code, review, test.

    MetaGPT · ChatDev
  • Group chat

    A shared room; a selector picks who speaks next, a chair closes with owners.

    AutoGen / AG2 · AgentScope
  • Debate + judge

    Proposer, skeptic and expert argue; a judge decides and records the dissent.

    Shiva war room · AgentScope
  • Manager + crew

    A manager plans, delegates to specialists in parallel and synthesises one answer.

    CrewAI hierarchical
  • Role-play pair

    An instructor and an assistant persona work step by step until the task is done.

    CAMEL
  • Handoff network

    Triage answers or hands the conversation to the specialist who finishes it.

    OpenAI Swarm · Agents SDK
  • Orchestrator ledgers

    Task and progress ledgers; the orchestrator re-plans when progress stalls.

    Magentic-One

See every runnable pattern

How it works

Describe. Hire. Wire. Run always-on.

  1. 01

    Describe

    Say what the role is for. Agent Studio drafts the persona, tools, memory and autonomy for you to edit.

    you write“Someone who chases overdue invoices politely, in Hindi and English, three mornings a week.”
  2. 02

    Hire

    Or skip the drafting: hire a ready persona or a whole department-in-a-box from the library.

    finance office
  3. 03

    Wire

    Connect triggers, agents, verifiers and approvals on the canvas, or start from a ready pattern.

    workflowSchedule → Agent → Approval → Memory
  4. 04

    Run always-on

    The scheduler wakes agents on time, runs survive restarts, and anything risky waits for your yes.

    statusAlways on · briefings on your phone
Governance

Autonomy is earned, per agent, per action.

Every tool carries a risk class. Every agent has an autonomy level. Anything above that level waits for a human, and everything is written to a tamper-evident log.

  • Kill switchOne switch stops every model call and tool call. Running work pauses.
  • Hash-chained auditEach entry commits to the one before; edits and deletions show.
  • Data-residency guardIndian personal data and children’s data stay on models in allowed regions.
  • AI disclosureAgents that speak to outsiders say they are AI, in line with EU AI Act Art. 50.

Read about security

Levelreadfetchwriteexternalcritical
0 · Observerunsasksasksasksasks
1 · Assistrunsrunsasksasksasks
2 · Actrunsrunsrunsasksasks
3 · Autonomousrunsrunsrunsrunsasks
“asks” = the step pauses for approval, unless you’ve set an “always allow” rule for that tool.
Where it fits

Not a chatbot. Not just a workflow tool.

Shiva Agents sits between three familiar categories and borrows the best of each.

CapabilityShiva AgentsWorkflow automation toolsAgent frameworksChat assistants
Agents that keep working while you’re awayInboxes, heartbeats, standing ordersRuns when triggeredYou build the runtimePer conversation
Visual workflow canvasYes, with agent and team nodesYesMostly codeNo
Multi-agent team patternsSeven topologies in one nodeRareUsually one style eachNo
Durable runs and long approvalsCheckpoints, resume, time-travel forkVariesVariesNo
Per-agent long-term memoryCore blocks + ranked recallNoAdd-onLimited
Autonomy levels and tamper-evident auditBuilt inPermissionsYou build itLimited
Run on your own GPUYes, or any OpenAI-compatible modelVariesYesRarely
Category columns describe typical tools in each category, not any particular product.
Questions

Straight answers.

Something missing? Ask us when you request access.

Is this another chatbot?

No. A chatbot waits for you to type. Shiva Agents runs a company of agents that wake on schedules and events, work through their inboxes, hand work to each other and come to you only when a decision is yours.

Which models does it run on?

By default, open models on the AskShiva GPU through Ollama: Gemma 4 (26B) for writing in English and Hindi and Qwen 3.6 (35B) for reasoning. You can add any OpenAI-compatible provider. If no model is reachable, a clearly labelled simulator keeps flows running so you can build offline.

What stops an agent from doing something I didn’t want?

Each tool has a risk class (read, fetch, write, external, critical) and each agent an autonomy level. Anything above the agent’s level pauses for your approval, and no level covers critical actions. A kill switch stops all model and tool calls at once, and every action lands in a hash-chained audit log.

Where does my data go?

The runtime is one Node.js process with a built-in SQLite database, on a machine you choose. Data is classed; with the residency guard on, Indian personal data and children’s data can only be processed by models in India or on your own machine, and the call is refused otherwise.

Can other tools use my agents?

Yes. Any OpenAI-compatible client can call an agent as a model (agent:<key>). Agents are also exposed over MCP and described by an A2A agent card, and runs stream over Server-Sent Events. See Developers.

Is there a mobile app?

Yes, an Android app alongside the web dashboard. iOS is next.

Start tonight

Hire your first agent before you sleep. It won’t.

Open the app, pick a blueprint, run one workflow with the example input and approve what comes back.