Describe the work.
Name the result you need and choose its repository. For a finance task, supply the figures; for account research, supply the sources you want considered.
From a brief to a result
Different company tasks.
One clear handoff.
A campaign brief, a code change, an account summary, or a draft budget: give an employee the context and the result you need. Choose a working repository, approve the request, and review the files it returns.
Self-host ShibaYOUR INFRASTRUCTURE. YOUR CALL.See the handoff
Coordination, without
giving up control.
Shiba prepares a proposed task for your review. Once you approve its exact request, a selected AI tool works in a separate environment and returns files or changes to inspect.
SHIBA / APPROVED TASK FLOWONE TOOL PER RUN
Inspect the handoff.
Check the task, repository, and any publishing request on the approval card before work begins.
Review the result.
Read the draft, changes, and reported checks. Decide what to use, revise, or publish.
Tools behind the work.
An employee’s role describes its job. A harness is the AI tool that carries out the approved task. Your workspace owner configures the available tools and their connections.
OpenCode
Default harnessThe default tool for carrying out an approved task in its working repository.
Configuration
Claude Code
Harness adapterUse your configured Claude Code connection for an approved task.
Configuration
Codex
Harness adapterChoose Codex for the task while keeping the same approval and review flow.
ConfigurationACP agents
Registry configurationUse a compatible agent that your workspace owner has configured and allowed.
ConfigurationA setup your team can reuse
Save the way an employee role should work.
Save instructions for a campaign writer, sales researcher, designer, operations helper, or finance reviewer. Reuse that brief for the next task instead of starting from scratch. Roles guide the work; they do not grant access to external services.
Create an employee →Bring a task from
where you work.
Start in the dashboard or a supported conversation. Your workspace owner sets up each connection, so check what is available in your installation.
01
Dashboard
Describe a task and review its run.
02
CLI
Work from your terminal.
03
Cron
Queue work on a schedule.
04
GitHub
Use configured repository events.
05
Slack
Start from a mention or /shiba.
06
Telegram
Send a task from a conversation.
07
Discord
Use configured interactions.
08
MCP
Call Shiba from an MCP client.
Bring your own workflow
Request → approve → reviewUse a supported MCP client to request work and read its status.
MCP tools & setup →Keep access separate
from the assignment.
Creating a role does not give it new account access. Shiba’s configured execution path keeps credentials separate from the working task; review the security guide for its boundaries.
Real keys stay outside.
The supported provider-backed path gives the task placeholder model keys. Configured credentials are added outside its working environment.
A separate workspace.
Each delegated run gets a separate working environment. The selected tool and your installation settings determine which services it can reach.
The handoff stops here.
Review the exact proposed task before repository execution. Automated tasks use approval by default too.
Inspect what happened.
Supported tool activity can leave an audit record for review. Records are best-effort and do not guarantee that every action is captured; use the task’s results and checks as well.
Future ideas,
clearly separated.
Direct messaging between agents is not available. For now, keep related work in a thread and use saved employee instructions to make repeat tasks easier.
PLANNED / NOT AVAILABLE
Agent-to-agent handoffs are an idea under consideration. There is no available API or delivery date.
Before you
connect.
Is this a peer-to-peer agent network?
No. Shiba coordinates each task and sends the approved work to one selected AI tool. Direct messaging between agents is only a design preview below.
Do all the agents run on every task?
No. Each run uses one selected AI tool. You can use different tools for different tasks, but they do not form a self-organizing team or share memory automatically.
Who approves the work?
An authorized person reviews and approves the proposed request before standard repository execution. Automated tasks also require approval by default; advanced unattended settings are a separate opt-in.
Can I use this with my team?
You can save employee instructions for work your team repeats, such as campaign drafts, account briefs, operating plans, or budget notes from supplied figures. Engineering, Marketing, Research, and Operations have starting templates; other responsibilities use custom instructions. Check the setup and readiness guides before relying on an installation.
Is every integration proven in a live deployment?
No. A supported connection still needs configuration and a successful test in your installation. The readiness guide records what remains unverified.
Your agents.
Your final say.
Start with one brief and a result you can check. Shiba is free to self-host; hosting and model usage cost extra. Follow the setup and readiness guides before relying on a new installation.