Your autonomous
software factory.

Turn Claude Code and Codex into a software factory you control. Delegate development, run independent checks, and deliver from your own environment.

Download AgentD Explore AgentD

Free local orchestration. Developer beta. Bring your coding agents.

$ curl -fsSL https://console.libraos.com/agentd/install.sh | sh

installed ~/.local/bin/agentd

next: cd your-repo && agentd init

$ export AGENTD_USER=you && cd your-repo && agentd init

activationcode from console.libraos.com/agentd ok

eulaAgentD End User Agreement (agentd eula, or libraos.com/eula) accepted

ledger~/.local/share/agentd/ledger.db

servicesagentd-notify, agentd-daily, agentd-watchdog enabled

claudeMCP server and hooks registered

codexMCP server and hooks registered

$ agentd run "Add a test for the parser" --budget 30m --check "go test ./..."

worker: claude in your-repo, branch agentd/hof-3f9a1c2b7e

budget: 30m

started; Ctrl-C detaches and the work continues

result

handoffhof-3f9a1c2b7e done

workerclaude claude-opus-5-5

branchagentd/hof-3f9a1c2b7e 8c41e2d

diff2 files changed, 84 insertions(+), 2 deletions(-)

checkgo test ./...: exit 0

minutes11.6

dollars0.94

Measured,
not promised.

The same 8 screens and the same hidden tests, built twice: once by a single Claude Code session, once by AgentD running 4 Claude and 4 Codex workers from one terminal, with every result checked.

Faster: 2.5×9.5 min to 3.8 min, wall clock to the last task done
Terminals: 18 agents at once, each in its own worktree
Context per agent: −45%137K to 76K tokens, and 8 of 8 results passed their check

Measured on 2026-11-03 with Claude Code and Codex workers. Full video and method

From issue to
verified delivery.

Give each task a goal, acceptance criteria and a check. AgentD keeps the handoffs and evidence together.

Research
Capture context and define what a successful result looks like.
Development
A coding agent works on the task in its own worktree.
Test
Run the project’s check command and keep its output with the change.
Validation
A separate validator checks the result. Questions and exceptions come back to you.

Checks support delivery decisions. Publishing and merging remain separate authorized actions, governed by your repository permissions.

Read the delivery workflow

Every result
has a receipt.

Follow the goal, the change and the checks. See what passed, what still needs attention, and where a person needs to make a decision.

Evidence stays with the task, so your next agent can pick up where the last one stopped.

Acceptance record

Goal: Make dialogs keyboard accessible.
Acceptance: Focus stays in the dialog. Escape closes it.
Verification: Project checks passed. Independent validation recorded.
Delivery: Changes ready. Publication requires authorization.

Local first.
Cloud optional.

AgentD Local orchestration is free. Run it in your own development environment, using your coding agents and model accounts.

Understand the costs

AgentD Local: Free orchestration. Beta invitation, activation and EULA acceptance required.
Models and tools: Coding-tool subscriptions and inference usage are billed separately by your providers.
Cloud Workers: Optional hosted execution, in beta. $0.04 per active minute, plus the job’s model usage from your LibraOS credits.