Shiba MascotShiba
Self-host
Open source · Self-hosted

AI employees for the work behind your company.

From code and design to marketing, sales, operations, finance, and revenue work. Give Shiba a task, review the plan, and keep the final decision.

Self host Shiba

Open source · AGPL-3.0-only · Your infrastructure

Free software to self-host. Hosting and model usage cost extra.

Works with:
OpenCode
Claude Code
Codex
Grok
Cursor
Slack
Telegram
Discord
Linear
GitHub
shiba.workspaceExample workspace
Shiba workspace showing AI employee role templates and the Agents dashboard.

Practical outcomes

A chatbot suggests.
Shiba delivers.

Examples of files Shiba can prepare. These are illustrative, not customer results or a live run.

Standard chatbot

Suggests theory

Here's a 7-step marketing strategy: 1. Identify target personas. 2. Draft value proposition. 3. Publish blog posts. 4. Run paid ads. 5. Measure KPIs. Would you like me to elaborate on step 1?

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/launch-briefReady for your approval
  • marketing/brief.md
  • marketing/campaign.md

Drafted campaign brief and announcement copy into repository. Waits for your review before merging.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

You could fix this by checking if the session exists: `if (!session) return false;`. Paste this into your auth handler and restart your server to see if it resolves the issue.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/checkout-fixTests passing · Ready for review
  • src/auth/session.ts
  • test/auth-session.test.ts

Reproduced bug in isolated sandbox, wrote regression test, applied fix, verified build cleanly.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

I suggest a clean mobile navigation with high contrast, a sticky header, and 44px touch targets. Be sure to test keyboard focus indicators and ARIA roles for screen readers.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/mobile-nav-specSpecification drafted
  • design/spec.md
  • design/review-checklist.md

Outlined full mobile navigation flow, state transitions, keyboard shortcuts, and WCAG checklist.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

Common sign-in edge cases include expired tokens, network dropouts, rate limiting, and SQL injection in input fields. Remember to write tests for each of these scenarios.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/auth-qa-runCheck completed in sandbox
  • qa/report.md
  • qa/reproduction-log.txt

Executed sign-in checks in container, captured exact HTTP responses, and logged reproduction steps.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

Start by researching the account, then personalize the outreach: mention a recent trigger, keep it under 120 words, and follow up in three days.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/account-outreachDrafts ready for review
  • sales/account-brief.md
  • sales/draft-outreach.md

Prepared an account brief and outreach drafts from the context you supplied. Nothing is sent until you approve.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

There are several approaches worth considering. Look at recent papers, vendor docs and community threads, then compare the tradeoffs yourself.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/market-questionReport drafted
  • research/report.md
  • research/sources.md

Answered the focused question with sources, and listed open questions and gaps in the evidence.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

A good product brief covers the problem, the audience, the scope and the success metrics. You should interview customers and write it up.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/onboarding-briefBrief drafted
  • product/brief.md
  • product/acceptance-checks.md

Turned customer evidence into a product brief with scope and acceptance checks your team can discuss.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge
Standard chatbot

Suggests theory

Document the process, assign owners, define escalation paths, and review the checklist every quarter to keep it current.

Result: You do all the actual work
Shiba AI employee

Prepares the deliverable

task/release-runbookRunbook drafted
  • ops/runbook.md
  • ops/recovery-steps.md

Turned a recurring process into a runbook with readiness and recovery steps someone can follow.

  1. Plan
  2. Draft
  3. Your review
Result: You review, approve, and merge

Nine dedicated roles

AI employees for
every company department.

Give each employee clear instructions and the files you want back. Four built-in templates, plus your own custom roles.

Built-in template

Marketing

Turn a product brief into campaign drafts for your review.

Example outputs

  • Launch brief
  • Campaign copy
Custom role example

Sales

Prepare account research and outreach drafts from supplied context.

Example outputs

  • Account brief
  • Draft outreach
Built-in template

Research

Answer a focused question with sources and gaps made clear.

Example outputs

  • Research report
  • Sources and open questions
Custom role example

Product

Turn customer evidence into a product brief your team can discuss.

Example outputs

  • Product brief
  • Scope and acceptance checks
Custom role example

Design

Map a user flow and draft interface copy before you build.

Example outputs

  • Design specification
  • States and accessibility checks
Built-in template

Engineering

Make a focused code change and report the checks run.

Example outputs

  • Code changes and tests
  • Review handoff
Custom role example

QA & Testing

Check a feature and document what passed, failed, or needs another look.

Example outputs

  • Test report
  • Reproduction steps
Built-in template

Operations

Turn a recurring process into a checklist someone can follow.

Example outputs

  • Operating runbook
  • Readiness and recovery steps
Custom role

Your custom role

Give an employee instructions for the work your team repeats.

You define

  • Your working brief
  • The files you need

Engineering, Marketing, Research, and Operations have built-in templates. The other roles shown here are examples you can create with custom instructions.

Roles use information you supply to prepare drafts. A role alone does not connect accounts, make payments, update external services, or grant permission to publish.

Create your first employee

Permission is the point.

Capable by design.Never in charge.

See the proposed task before it runs. Approve the exact request, reject it, or change direction. Unattended automations require explicit, scoped permission.

Inside the approval gate

Keep the work yours.

Host Shiba in your own account and choose the connections it uses. Your model provider still receives the context sent with its requests.

Review before you rely on it.

Read the returned work and checks, then decide what to keep. Teams evaluating Shiba can start with the readiness guide.

Example approvalWaiting for you
Proposed taskPrepare a draft budget from supplied figures
Repositoryexample/team-workspaceBranchtask/draft-budget
Human approval required
Sandbox not started

Review the request before work begins.

Illustrative boundary, not a live run. Planning may call a model before approval.

Governance & safety

Autonomous,
not unsupervised.

Real guardrails built into the architectural foundation. No hidden external calls or unvetted merges.

Your Cloudflare account
Runs on Workers and Containers in your cloud infrastructure. Stored data stays yours.
Sandbox per task
Each task executes in an ephemeral container with strict memory and CPU limits.
Repository required
Non-coding and coding tasks operate inside a designated GitHub repository.
Deployment-owned credentials
Model keys stay inside Cloudflare AI Gateway; containers never receive raw secrets.
Reviewable diffs
Inspect line-by-line diffs, generated drafts, and test logs before accepting work.

A task. A plan. A result to review.

From “what if”
to ready for review.

Start with a piece of work you can describe. Shiba helps turn it into files and changes your team can review.

Self host in your account

Deploy to Cloudflare Workers with Docker sandbox container in minutes.

Connect a repository

Select a GitHub repo to give your AI employees context and output files.

Assign your first task

Describe the work, review the proposed plan, and approve the run.

bash — shiba local setup
$ git clone https://github.com/princepal9120/shiba.git
$ cd shiba && pnpm install
$ pnpm shiba dev
› worker listening at http://127.0.0.1:8787
› sandbox container image built: standard-4
✓ ready for task assignment

Know what you’re trying.

Start small.
Check the result.

Shiba is an early prototype. Explore the source and the verification notes before giving it work your company depends on.

Read the readiness notes

Yours to inspect.

The source is public under AGPL-3.0-only. Read how work is planned, approved, and returned before running it in your own account.

Explore the source

What still needs proving.

No successful cloud task with a live model and GitHub is established in the readiness record.

In the repository
Automated checks for task handling, approval, and execution adapters.
Needs live proof
A complete task using your hosting, model connection, and GitHub setup.

Know the boundaries.

Connections have different access and data-handling requirements. Review them before adding private company information or enabling an experimental agent.

Connection setup and limits

A few things worth asking.

Good
questions.

The short version here.
The details are in the docs.

Open the docs
What is Shiba?

Shiba is an open-source workspace for AI employees. Give an employee a task and the context it needs, review the proposed work, and inspect the files or changes it returns.

What kind of work can I give it?

Draft campaign copy, research an account using supplied context, outline a design, fix a bug, prepare an operating checklist, or organize supplied figures into a draft budget or pipeline summary. Set clear instructions and review the result before using it.

Do I need a repository for non-coding tasks?

Yes. Every task currently needs a GitHub repository, including non-code work. It gives Shiba context and a place to return draft files. The task does not have to change code.

What happens before I approve?

Shiba may use a model to draft a plan. For normal runs, no sandbox starts until you approve the exact task. Reject it and nothing runs.

Where does my information go?

Your installation stores Shiba data in your Cloudflare account. Your AI provider receives the context sent with each request. Check access before sharing confidential data.

Which agents and connections are supported?

OpenCode, Claude Code, Codex, and other adapters are available when configured. Each integration needs setup; listed support does not mean a live task is verified.

What does Shiba cost?

Shiba software is free under AGPL-3.0-only. You pay your Cloudflare and model providers; costs vary by usage. See the cost guide.

Is Shiba production-ready?

Shiba is an early prototype. Tests cover core workflows, but no live cloud task with a model and GitHub has been verified. Try a non-critical repo.

How do I get started?

Follow the self-hosting guide to configure Cloudflare and a model provider. Then run a small task in a test repository.

Your next employee runs on your terms.

Give the next task
a clear place to start.

Set up Shiba, choose one piece of work, and review what comes back.

Self host Shiba