Japanese readers: start with README.ja.md or the full Japanese documentation hub.
A macOS/tmux workbench for running multiple local AI-agent CLI sessions in one window.
The first supported target is Codex CLI. The goal is simple: open 4, 6, or 9 agent panes against one shared workspace so you can collect ideas, compare approaches, run parallel research, and keep the resulting files in one place.
Use Local Agent Workbench when one AI chat is too small for the job.
- You want 4, 6, or 9 Codex CLI sessions looking at the same problem from different angles.
- You are researching a technical topic and want separate panes for sources, risks, examples, and synthesis.
- You are maintaining OSS and want structured notes for issue triage, PR review, release prep, or docs cleanup.
- You want reusable Markdown output instead of scattered chat replies.
- You want a local-first setup that is easy to inspect, validate, and share safely.
New to this project? Start with a plain-language explanation:
- English beginner guide
- 日本語のやさしい解説
- Recipes: practical ways to use the workbench.
- Promotion kit: short copy for sharing the project.
Choose the shortest path for your situation:
- Start Here / まずここから: understand the project in the first 3 minutes.
- Use Cases / ユースケース: pick a practical workflow such as issue triage, PR review, release prep, or feature discovery.
curl -fsSL https://raw.githubusercontent.com/goonobu-dot/local-agent-workbench/main/scripts/install.sh | bash
cd "$HOME/AgentWorkbench/local-agent-workbench"
make first-runFull setup notes are in docs/install.md.
git clone https://github.com/goonobu-dot/local-agent-workbench.git
cd local-agent-workbench
./scripts/doctor.sh
make demoThe fastest no-risk preview is make first-run: it checks your machine, creates
a temporary demo workflow, closes it into a handoff summary, and lists available
workflow templates.
Read docs/demo-transcript.md to see the expected output shape before running anything locally.
If you are unsure where to start, run:
./scripts/recommend_workflow.sh issue
./scripts/recommend_workflow.sh security- OSS maintainers who need faster issue triage, pull request review, release preparation, or security screening.
- Developers who want several independent Codex CLI angles on one local folder without building a hosted system.
- Researchers and builders who want reusable markdown outputs instead of scattered chat transcripts.
- a local 4, 6, or 9 pane tmux workbench
- workflow templates for issue triage, PR review, release prep, and feature discovery
- one shared folder containing role prompts, pane notes, and a final handoff summary
- a doctor report and quality gates so setup problems are easier to diagnose
- issue triage result
- pull request review verdict
- release checklist
- feature decision memo
- security triage
- documentation improvement plan
- dependency update review
If the workbench helps you make a real maintenance decision, open a usage report. Short reports are useful: the workflow used, what it produced, and where the setup was confusing.
This project is maintained as a practical Codex-assisted OSS workflow experiment. The updated application notes are in docs/CODEX_FOR_OSS_REAPPLICATION.md.
For a fast external review, use:
AI agents are useful, but one chat window is often too narrow for exploration. This workbench gives you a repeatable local setup for:
- parallel research and idea generation
- comparing multiple solution paths side by side
- keeping all panes pointed at the same project folder
- switching between 4, 6, and 9 panes without rebuilding your terminal layout
- launching the setup as a small macOS app shortcut
- macOS
tmux- Codex CLI available on your
PATH - Python 3 and Pillow only if you regenerate the icon
Install tmux with Homebrew:
brew install tmuxOne-command install:
curl -fsSL https://raw.githubusercontent.com/goonobu-dot/local-agent-workbench/main/scripts/install.sh | bashThen launch:
cd "$HOME/AgentWorkbench/local-agent-workbench"
make demo
./scripts/launch_codex_tmux.shCheck your local setup without starting a workbench session:
./scripts/doctor.shRun the local validation suite:
make testCreate a reusable maintainer workflow folder before launching panes:
./scripts/new_workflow.sh issue-triage
AGENT_WORKBENCH_IDEA_DIR="$HOME/AgentWorkbench/Idea" ./scripts/launch_codex_tmux.shStart directly from a public GitHub issue or pull request URL:
./scripts/create_workflow_from_url.sh https://github.com/owner/repo/issues/123
./scripts/create_workflow_from_url.sh https://github.com/owner/repo/pull/123Launch with role-specific prompts generated from pane-roles.md:
AGENT_WORKBENCH_IDEA_DIR="$HOME/AgentWorkbench/Idea" \
AGENT_WORKBENCH_USE_ROLE_PROMPTS=1 \
./scripts/launch_codex_tmux.shAfter the panes write their notes, create a handoff summary:
./scripts/close_workflow.sh "$HOME/AgentWorkbench/Idea"Check workflow status before sharing or closing:
./scripts/workflow_status.sh "$HOME/AgentWorkbench/Idea"Export a workflow folder for sharing:
./scripts/export_workflow.sh "$HOME/AgentWorkbench/Idea"Import a shared workflow archive:
./scripts/import_workflow.sh ./Idea-workflow-export.tar.gzAll panes use one shared idea folder by default:
~/AgentWorkbench/IdeaManual clone instead of the installer:
mkdir -p "$HOME/AgentWorkbench"
git clone https://github.com/goonobu-dot/local-agent-workbench.git "$HOME/AgentWorkbench/local-agent-workbench"
cd "$HOME/AgentWorkbench/local-agent-workbench"
./scripts/doctor.shcontrol-b z: zoom or unzoom the active panecontrol-b 4: switch to 4 panescontrol-b 6: switch to 6 panescontrol-b 9: switch back to 9 panes
The public configuration names are AGENT_WORKBENCH_*. Older CODEX_WORKBENCH_* names are still accepted as compatibility aliases.
AGENT_WORKBENCH_PANE_COUNT=4 ./scripts/launch_codex_tmux.sh
AGENT_WORKBENCH_BASE="$HOME/AgentWorkbench" ./scripts/launch_codex_tmux.sh
AGENT_WORKBENCH_IDEA_DIR="$HOME/AgentWorkbench/Research" ./scripts/launch_codex_tmux.sh
AGENT_WORKBENCH_MODEL="gpt-5.4-mini" ./scripts/launch_codex_tmux.sh
AGENT_WORKBENCH_AUTO_SUBMIT=0 ./scripts/launch_codex_tmux.shDefaults:
| Setting | Default |
|---|---|
AGENT_WORKBENCH_PANE_COUNT |
9 |
AGENT_WORKBENCH_SESSION |
local-agent-9 |
AGENT_WORKBENCH_BASE |
~/AgentWorkbench |
AGENT_WORKBENCH_IDEA_DIR |
~/AgentWorkbench/Idea |
AGENT_WORKBENCH_MODEL |
gpt-5.4-mini |
AGENT_WORKBENCH_AUTO_SUBMIT |
1 |
See docs/workflows.md for practical patterns:
- parallel research
- competing hypotheses
- feature discovery
- small task mode
- public-output safety checks
See also:
- docs/oss-maintainer-use-cases.md
- docs/showcase.md
- docs/why.md
- docs/one-minute-demo.md
- docs/evaluation-guide.md
- docs/install.md
- docs/demo-transcript.md
- docs/adoption-scorecard.md
- docs/continuous-improvement-loops.md
- docs/commands.md
- docs/architecture.md
- docs/quality-gates.md
- docs/workflow-templates.md
- docs/workflow-sharing.md
- docs/troubleshooting.md
- docs/faq.md
- SUPPORT.md
- docs/publication-checklist.md
- docs/openai-codex-for-oss.md
- docs/adoption-plan.md
Example:
- examples/issue-triage-demo
- examples/pr-review-demo
- examples/release-prep-demo
- examples/feature-discovery-demo
- examples/security-triage-demo
- examples/docs-improvement-demo
- examples/dependency-update-demo
Project operations:
If you try it on a real or fictional maintainer task, share the result with the usage report issue template. Reports about confusing setup steps are as useful as reports about successful workflows.
./scripts/build_codex_app.sh
open "$HOME/Applications/Local Agent Workbench.app"The generated app opens Terminal and launches the tmux workbench. The checked-in AppleScript does not contain a personal path. By default, it expects this repository at:
~/AgentWorkbench/local-agent-workbenchThis repository intentionally does not include local logs, .env files, prompt histories, Obsidian vaults, or generated agent output.
Before publishing your own fork, run:
make testThis is an early public release extracted from a real local workflow. The current implementation focuses on Codex CLI because that is the tested path. The naming is intentionally broader so future adapters can support other local agent CLIs without changing the project identity.