Discover MCPs & agents
Loading MCPs and agents…
Loading MCPs and agents…
Turn Claude Code into a small, organized software office. 11 agents, 21 commands, Agile workflow, multilingual.
From the repo.
Turn your Claude Code session into a small, organized software office. 7 agents. 10 slash commands. One linear flow. No sprint, no retro, no standup.
🇹🇷 Türkçe: README.tr.md
Status: early preview (v0.2.x). Verified locally. APIs (agent names, command behavior, output formats) may still change before 1.0. Community feedback wanted.
Inspired by Claude Code Game Studios and the broader Agile-agent ecosystem (BMAD-METHOD and others). Deliberately curated and smaller than those.
A single Claude session is powerful but unstructured. Nobody asks "do we really need this?", nobody enforces code review, nobody catches when you skip tests.
Software Office gives Claude a tight, linear team:
You still make every decision — but inside a small team that asks the right questions, knows its boundaries, and doesn't ceremonialize routine work.
researcher → qa (analysis) → tech-lead (tasks) → developer(s) → tech-lead (review) → qa (validation) → done
That's /feature — the default path for any non-trivial change.
| Path | When | Steps |
|---|---|---|
/feature | Default. Any change that isn't a one-liner. | researcher → qa → tech-lead → developer → tech-lead → qa |
/quick-fix | < 50 LOC, no architecture impact. | developer + tech-lead glance |
/bug-fix | A reported bug. | researcher → developer → qa (regression test required) |
On-call roles are invoked by trigger only:
| Role | Trigger |
|---|---|
| cto | Stack pick, architectural change, breaking API, scope conflict, release sign-off |
| security-reviewer | qa flagged risk (auth / PII / payments / files / migration), pre-release audit |
| devops | CI / Docker / deployment / observability work in the task list |
If qa or tech-lead pulls one of these in for routine work, that role returns it. The flow stays lean.
/takeover brings it forwardIf you wanted scrum, sprints, retros, standups — those are gone. This is deliberately a flow framework, not an Agile framework.
Three install paths.
In a Claude Code session:
/plugin marketplace add ilkerprdal/Claude-Software-Office
/plugin install claude-software-office@claude-software-office-marketplace
Auto-updates via /plugin update, scoped to user or project.
macOS / Linux / Git Bash on Windows:
curl -fsSL https://raw.githubusercontent.com/ilkerprdal/Claude-Software-Office/main/install.sh | bash
Windows PowerShell:
irm https://raw.githubusercontent.com/ilkerprdal/Claude-Software-Office/main/install.ps1 | iex
Run from your project root. Installs .claude/, CLAUDE.md, .gitignore,
and a production/ scaffold. Existing CLAUDE.md is backed up as
CLAUDE.legacy.md.
git clone https://github.com/ilkerprdal/Claude-Software-Office.git
cd your-project
bash /path/to/Claude-Software-Office/install.sh # Mac/Linux
.\path\to\Claude-Software-Office\install.ps1 # Windows
Or copy .claude/ and CLAUDE.md from the cloned repo into your project.
claude
/start # or /takeover if you have prior AI context
Agents auto-detect your language (verified: English, Turkish).
On-call (Opus)
└── cto → stack, architecture, breaking change, release sign-off
The flow (Sonnet)
├── researcher → investigates topic, returns evidence (no opinion)
├── qa → spec + AC + test plan, then validates result
├── tech-lead → splits into tasks, reviews code
└── developer → implements end-to-end (backend, frontend, tests)
On-call (Sonnet, security-reviewer Opus)
├── security-reviewer → STRIDE / OWASP audit when qa flags risk
└── devops → CI/CD, deployment, observability when needed
| Command | Purpose | Drives |
|---|---|---|
| Onboarding | ||
/start | Detect stack + state, suggest next | — |
/takeover | Import prior AI context (Cursor / Copilot / Windsurf / Aider) | — |
/help | Smart suggestion + command list | — |
| Build | ||
/research | Standalone investigation — facts only, no build | researcher |
/feature | Default flow for any change | researcher → qa → tech-lead → developer |
/quick-fix | Tiny change, skip the flow | developer + tech-lead |
/bug-fix | Locate, fix, regression-test | researcher → developer → qa |
| Gates | ||
/security-review | On-demand security audit (STRIDE + OWASP) | security-reviewer |
/release-check | Pre-release GO / NO-GO | cto |
| Knowledge | ||
/memory | View / add project learnings | — |
/research is the only command that drives the researcher without kicking
off the build flow. Use it when you want to scope a problem before deciding
whether it deserves a /feature, /quick-fix, or nothing at all.
Building a TODO app's password reset:
/feature add password reset)The flow kicks off. Five steps:
production/qa/spec-password-reset.mdproduction/stories/password-reset.md with named files per taskAPPROVE_WITH_NITS (one nit: log message wording)production/qa/validation-password-reset.mdNo standup. No retro. No sprint roll-over. The feature is done when qa GATE = PASS.
For releases: /release-check — cto walks the gate and signs GO/NO-GO.
your-project/
├── CLAUDE.md # Auto-loaded each session
├── .claude/
│ ├── settings.json # Permissions (allow/deny)
│ ├── agents/ # 7 agent definitions
│ ├── commands/ # 9 slash commands
│ ├── memory/ # Accumulated learnings
│ └── docs/ # Coordination, collaboration, standards
├── src/ # Source code
├── tests/ # Tests
├── docs/
│ ├── architecture/ # cto's architecture docs
│ ├── adr/ # Architecture Decision Records
│ ├── api/ # OpenAPI / GraphQL SDL / .proto
│ └── security/ # security-reviewer threat models
└── production/
├── qa/
│ ├── spec-*.md # qa Mode A specs
│ ├── validation-*.md # qa Mode B validations
│ ├── bugs/ # bug reports
│ └── security-review-*.md
├── stories/ # tech-lead task breakdowns
├── releases/ # cto GO/NO-GO decisions
└── session-state/
└── active.md # session context
No sprints/, no retros/, no standup-log.md — those are gone.
This project leverages Claude's existing multilingual ability rather than shipping localized templates.
What it actually does:
Verified locally: English, Turkish. Designed for: any language Claude itself supports — but unverified by us. If you've used it in another language, share a screenshot in Discussions.
This is a template, not a locked framework.
.md file in .claude/agents/.md file in .claude/commands/.md.claude/docs/coding-standards.md.claude/settings.jsonAfter adding a file, restart your Claude session.
Q: What happened to sprints, retros, standups, story points? A: Gone in v0.2. The previous setup had 13 agents and 24 commands modeling a full Scrum process. For solo / small teams that was overhead, not value. This version is a flow framework, not an Agile framework. If you need sprint ceremonies, use a heavier tool — see "Wrong tool" above.
Q: Does this work with Cursor / Copilot / Windsurf?
A: The framework is built for Claude Code (@anthropic-ai/claude-code).
Other AI tools may pick up CLAUDE.md partially, but agent delegation, slash
commands, and Task calls are Claude Code features. /takeover imports
their context files into our memory layer.
Q: My change is too small for /feature.
A: Use /quick-fix (< 50 LOC, no architecture impact). It skips researcher
and qa-analysis — goes straight to developer with a tech-lead glance.
Q: My language isn't English. Will it still work? A: Yes. Each agent has a Language Protocol — it detects your language and responds in it. Code stays English (industry convention), but docs, comments, and chat all follow your language. Tested with English and Turkish.
Q: Can I add my own agents or commands?
A: Yes. Drop new .md files in .claude/agents/ or .claude/commands/.
Restart your Claude session to load them. See
CONTRIBUTING.md for the format.
Q: Will it conflict with my company's existing process? A: Likely yes for any org with established Jira / sprint / retro process — this doesn't try to model those. For solo / small teams, this is the process.
Q: Does it call out to external services? A: No. Everything is local files + your existing Claude Code session. No telemetry, no analytics. The only network calls are git operations you control.
Q: How do I update to a newer version?
A: Re-run the installer (curl|bash or irm|iex). Your CLAUDE.md is
backed up as CLAUDE.legacy.md; .gitignore is merged. Custom files in
.claude/agents/ you added stay; standard files get refreshed.
Q: Can multiple developers share .claude/memory/?
A: Yes — it's just markdown files. Treat them like any other project doc:
commit, review, merge in PRs. Conflicts are markdown line conflicts.
You typed it before installing, or in a different folder than where you ran
the installer. Check that .claude/commands/start.md exists in the project
root, then restart Claude Code.
The Language Protocol detects from recent messages. Send 1–2 messages in your
language, the agent will switch. If it doesn't, check that the
### Language Protocol section exists in the agent's .md file.
chmod +x install.sh
bash install.sh
Or run directly via bash install.sh without making it executable.
powershell -ExecutionPolicy Bypass -File install.ps1
The installer doesn't change system state.
Check for CLAUDE.legacy.md in your project root — that's your previous
version, automatically backed up. Re-merge what you need into CLAUDE.md.
python scripts/validate.py
It tells you which file and what's missing. Common fixes: missing
description: field, wrong model value (must be opus/sonnet/haiku/inherit).
Open an issue with the Bug report template.
MIT
Replace {MCP_ENDPOINT_URL} with this MCP’s endpoint URL (from its repo or docs above). No API key — you connect directly.
Tool
OS
Config file: ~/.cursor/mcp.json
{
"mcpServers": {
"mcp-server": {
"url": "{MCP_ENDPOINT_URL}"
}
}
}Paste into mcpServers in the config file. Restart Cursor after saving.
If this MCP is also published on mcpchannel.ai, you can subscribe from Browse and use the gateway config there instead.