agent-hub

March 18, 2026 · View on GitHub

Local-first multitask board for routing and observing code-assistant work across projects.

Quick links: Live Site · Demo Page · 中文 Demo · 中文 README · Repository · First Release

This OSS repo is the portable, public-safe slice of the broader idea:

queue code-assistant tasks in SQLite
route work into repo-local code agents through project-backed commands
model dependencies, retries, blocking, and human handoff
expose a small HTTP surface and a thin browser dashboard for multitask visibility

The codebase stays intentionally small and uses the Python standard library only.

📊 At A Glance

Topic	What `agent-hub` does
Primary role	A multitask board for code assistants
Main user	One operator supervising many coding tasks
Main targets	Local repos plus repo-local agents like Codex or Claude Code
Main value	Queueing, routing, dependencies, retries, and human handoff
Primary interaction	Talk to your coding assistant; let it place work on the board
Current scope	Local-first OSS MVP, not a hosted platform

🌐 Project Site

live site: https://dominic789654.github.io/agent-hub/
landing page: docs/index.html
demo guide: docs/demo.md
demo page: docs/demo.html
recommended agent workflow: docs/agent-driven-usage.md
public launch checklist: docs/public-launch-checklist.md
public release: https://github.com/Dominic789654/agent-hub/releases/tag/v0.1.0
GitHub Pages publish is wired through .github/workflows/pages.yml

🚦 Current Status

Current status: public-safe OSS MVP.

ready to share as a local-first, single-operator code-assistant multitask baseline
suitable for demos, exploration, and extension in local environments
not yet positioned as a mature hosted platform or stable long-term API surface
scope is intentionally constrained to queueing, routing, visibility, and handoff

🗺️ Near-Term Roadmap

improve example projects and end-to-end walkthroughs
add richer screenshots or UI captures for the public site
tighten public API framing and compatibility expectations
expand executor and policy examples without turning the repo into a heavy platform

🤖 Recommended Usage Pattern

Use agent-hub as the multitask board and control plane, not as a replacement for your coding agent.

keep agent-hub responsible for queueing, routing, dependency handling, and visibility
keep repo-local coding agents responsible for actual implementation work
register those agents as project-backed local commands, then launch them through agent-hub

The current OSS slice is a good fit for workflows where you already use tools like Claude Code, Codex, Kimi Code, or Qwen Code in local repos and want one explicit multitask board in front of them.

See docs/agent-driven-usage.md for the recommended setup pattern.

Layer	Responsibility
You	Describe what should happen and review outcomes
Codex / Claude Code	Implement or analyze inside the repo
`agent-hub`	Queue work, route it, sequence it, and surface exceptions

🚫 Not A Traditional Task Board

agent-hub is not meant to be a generic to-do tracker for arbitrary operator commands.

the primary unit is a bounded code-assistant task
the main target is a local repo plus a repo-local coding agent
the main value is routing, dependency control, retry visibility, and human handoff
the dashboard is for supervising assistant work, not for replacing the assistant itself

✨ What It Does

agent-hub currently supports:

task queueing for code-assistant work with lifecycle state
dependency edges between tasks
standalone dispatcher execution
project-backed actions, task templates, and pipelines for repo-local agents
task and pipeline run notes / labels
saved query presets and execution
human inbox aggregation for manual triage
dashboard JSON and a thin browser app at /app

Capability	Why it matters
Multi-assistant board	See Codex, Claude, and other repo-local agents in one place
Dependency-aware routing	Make serial work wait and let unrelated work run in parallel
Human inbox	Stop hard cases from disappearing inside automation
Saved views	Slice the board by assistant, repo, or handoff state
Thin dashboard	Check current state quickly without opening many terminals

🧱 What It Is Not

This repo is not yet:

a multi-user SaaS
a hosted control plane
an auth-enabled production service
a generic remote executor framework

Treat it as a strong local MVP for code-assistant orchestration, not a finished platform.

Not the goal	Why
Generic to-do board	The board is centered on code-assistant tasks, not arbitrary tickets
Hosted SaaS	The current value is local-first control and visibility
Assistant replacement	Codex or Claude still do the implementation work

⚡ Quick Start

Requirements:

Python >= 3.11

python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
pytest

✅ Smoke Test

Use the public demo registry for the first verification path:

agent-hub version
agent-hub --projects-file examples/agent-driven-projects.example.json list-projects
agent-hub --projects-file examples/agent-driven-projects.example.json list-project-task-templates demo-codex

If these commands work, your local install is ready for the public demo flow.

⏱️ Five-Minute Operator Flow

The intended usage is:

one background terminal for the board
one background terminal for the dispatcher
one assistant terminal where you talk to Codex or Claude Code

🖥️ Background terminal A

python -m agent_hub --projects-file examples/agent-driven-projects.example.json serve --port 8080

🖥️ Background terminal B

python -m agent_hub --projects-file examples/agent-driven-projects.example.json dispatch

💬 Assistant terminal

Open Codex or Claude Code in the environment where you want to operate the board, then ask it to submit tasks for you.

Example prompts:

Create a Codex task in demo-codex to investigate why the local build script is flaky.
Create a Claude task in demo-claude to review the proposed fix and summarize risks.
If the review is clean, queue the review-then-implement pipeline in demo-codex for "Add a dry-run mode to the deployment helper".
Show me the current human inbox and tell me if any task needs manual routing.

That is the primary interaction model. The operator mainly talks to the coding assistant, and the assistant uses agent-hub to place and inspect work on the board.

If you want to evaluate the board manually before introducing an assistant, you can run the same public flow directly:

python -m agent_hub --projects-file examples/agent-driven-projects.example.json run-task-template demo-codex delegate-task --input "Investigate why the local build script is flaky"
python -m agent_hub --projects-file examples/agent-driven-projects.example.json run-pipeline demo-codex review-then-implement --input "Add a dry-run mode"
python -m agent_hub --projects-file examples/agent-driven-projects.example.json dashboard

Then open:

http://127.0.0.1:8080/
http://127.0.0.1:8080/app
http://127.0.0.1:8080/dashboard

The canonical public onboarding path uses examples/agent-driven-projects.example.json.

If you want the lower-level CLI walkthrough, use docs/demo.md or docs/demo.html. The runnable assistant-board example registry lives at examples/agent-driven-projects.example.json.

📁 Default Local State

data dir: ./.agent-hub/
database: ./.agent-hub/agent_hub.db
projects registry: ./.agent-hub/projects.json

Overrides:

--data-dir
--projects-file
AGENT_HUB_DATA_DIR
AGENT_HUB_PROJECTS_FILE

🧰 Common Board Operations

These are the operations your assistant should usually perform on your behalf.

🗣️ What You Ask The Assistant

Create a Codex task in demo-codex for this bug report.
Create a Claude review task for the proposed fix.
Queue the review-then-implement pipeline in demo-codex.
Check the human inbox and summarize what needs routing.
Retry the failed Codex task and tell me what changed.

⚙️ What The Assistant Uses Under The Hood

python -m agent_hub --projects-file examples/agent-driven-projects.example.json run-task-template demo-codex delegate-task --input "..."
python -m agent_hub --projects-file examples/agent-driven-projects.example.json run-task-template demo-claude delegate-task --input "..."
python -m agent_hub --projects-file examples/agent-driven-projects.example.json run-pipeline demo-codex review-then-implement --input "..."
python -m agent_hub --projects-file examples/agent-driven-projects.example.json list-human-inbox
python -m agent_hub --projects-file examples/agent-driven-projects.example.json retry-task <task-id>

The CLI remains important, but mostly as the mechanism behind the assistant-facing workflow, not as the primary operator experience.

📚 If You Want The Technical Details

architecture notes: docs/architecture.md
assistant-first workflow: docs/agent-driven-usage.md
CLI quick reference: docs/cli-quick-reference.md
runnable board walkthrough: docs/demo.md
browser-friendly demo: docs/demo.html
release / launch checklist: docs/public-launch-checklist.md
public release notes: docs/release-notes-v0.1.0.md

🧪 Example Project Registries

The repo includes two example registries:

examples/agent-driven-projects.example.json for the recommended code-assistant board pattern
examples/projects.example.json for the lower-level portable sample registry used by tests and scaffold validation

📝 Release Notes

Before cutting a public release, run the checklist in:

docs/release-checklist.md
docs/oss-repo-prep.md
docs/oss-split-plan.md
docs/license-recommendation.md
docs/public-repo-commands.md
docs/release-notes-v0.1.0.md

🤝 Contributing

See:

CONTRIBUTING.md
CODE_OF_CONDUCT.md
SUPPORT.md
SECURITY.md

🔒 Safety

no private infrastructure assumptions
no secrets in examples
local state stays under ignored paths such as ./.agent-hub/
destructive actions remain explicit