Launching on Product HuntJoin the launch →

Open source · local-first · Claude Code + Codex

Understand the system. Design the work. Run the agents.

Coding agents write code fast. Handraise helps you decide what work should exist—from an evidence-backed map of your system to safe, dependency-aware agent runs.

Get Handraise See how it works

Runs on your machine
Human-reviewed plans
Your CLI credentials stay put

localhost:4177 / repository local

Understand the system. Design the work. Run the agents.payments-platform

✓Understandcurrent snapshot
2Design4 components · 6 fronts
3Run2 fronts ready

Accepted work modelGoals → components → fronts → runsGraphList

Runtime control2 fronts

Repository planning2 fronts

Repository intelligence2 fronts

Accepted 4 components + 6 fronts Derived critical path: map → architecture → publish

Works with the tools already on your machine

◈ Claude Code⌁ Codex▣ tmux⑂ Git worktrees◇ Graphify optional

The coordination problem

More agents don't fix a bad plan.

Worktrees prevent file collisions. Terminals show activity. Neither tells you whether the system was understood, the boundaries make sense, or the work can be verified.

Folders are not responsibilities

A directory tree is evidence—not a product architecture. Handraise proposes durable components around outcomes, interfaces and ownership.

Parallel isn't automatically useful

Fronts expose dependencies, affected components, acceptance criteria and a critical path before a worktree is allocated.

Activity is not an outcome

Agent claims, process state and reviewed evidence remain separate. Completion requires the accepted contract and observable checks.

repository + product intent→system map→work model→agent runs→verified outcomes↺

The order is the product

Model the work
before agents run.

Speed compounds good architecture—and bad architecture. Handraise puts understanding and design back in front of execution.

One evidence-first loop

From codebase to coordinated work.

Handraise keeps observation, human intent, generated proposals, accepted contracts and runtime evidence visibly distinct.

01 · Understand the system

See what exists—and how sure we are.

Choose an exact, read-only scope. Explore responsibilities, dependencies, entry points, interfaces, tests and data flows with traceable evidence.

Built-in local analyzer always available
Optional Graphify semantic graph
Coverage, provenance and uncertainty visible
Snapshot comparison and drift

RESPONSIBILITYAuthentication boundary18 entities · 26 relations

DEPLOYABLELocal control server31 entities · 48 relations

INTERFACEAgent lifecycle hooks12 entities · 15 relations

Why this grouping exists6 evidence referencesmedium uncertainty

No synthetic certainty

Five truths.
Never silently mixed.

A model suggestion isn't architecture. A running process isn't a completed outcome. Handraise makes every state transition explicit and reviewable.

ObservedFacts and inferences from code, docs, tests and history.evidence

DeclaredUsers, outcomes, constraints and priorities supplied by people.intent

ProposedGenerated maps and plans that remain private drafts.review

AcceptedHuman-readable contracts explicitly published to the repository.contract

ExecutionProcesses, worktrees, checks, discoveries and outcomes at runtime.runtime

Local-first is a trust boundary

Your codebase isn't a prompt by default.

Handraise runs on your machine, reuses the Claude Code and Codex accounts already authenticated there, and shows the exact source boundary before model-assisted planning.

01Read-only before reviewAnalysis is bounded, cancellable and cannot mutate the target repository.

02No copied provider credentialsAuthentication remains owned by the first-party agent CLIs.

03Explicit remote accessLoopback is implicit; LAN and Internet clients pair with revocable one-time access.

04Human acceptance boundaryGenerated proposals cannot become repository state without exact reviewed publication.

The infrastructure underneath the idea

A serious local control plane.

The planning thesis is useful because the operational layer is already there when the work is ready.

⌁

Fleet command center

See every repository, session, attention request and Git risk without hunting through terminals.

⑂

Safe worktrees

Isolate fronts and surface dirty, ahead, behind, branch mismatch and local-only commit risk.

◎

Permission control

Review typed Claude Code and Codex attention from the browser; open the terminal only when needed.

↗

Remote clients

Pair another browser over private network or an explicit HTTPS tunnel with revocable access.

◇

Honest degradation

Missing Graphify, models or language coverage stays visible. Manual planning always remains available.

↺

Outcome learning

Turn reviewed checks and architecture drift into new proposals—never silent contract mutations.

Open source · install from source

Raise your hand.
Take back the plan.

Requires Node 20+, tmux and Python 3. Your existing Claude Code or Codex login is reused locally.

View on GitHub Support on Product Hunt

Terminal

$ git clone https://github.com/martin882003/handraise.git
$ cd handraise && npm install && npm run build
$ npm link
$ handraise hooks install
$ handraise repo add ~/code/my-project
$ handraise serve

Questions worth asking

Before you hand over the keyboard.

Is Handraise another coding agent?+

No. It orchestrates the Claude Code and Codex CLIs already on your machine. Handraise models the system and the work, then controls agent sessions against accepted plans.

Does it upload my repository?+

No. Repository analysis is local and read-only by default. Model-assisted planning shows the exact selected snippets, byte count, provider and destination before host-only consent.

Do I need Graphify?+

No. The built-in structural analyzer is always available. A compatible existing Graphify installation can enrich the local system map, but Handraise never installs it automatically.

What happens when I close the browser?+

Your controlled agents keep running in tmux. Reopen Handraise or attach from a terminal. Private drafts and job history can be resumed with revision and staleness checks.

Can I use it from another device?+

Yes. Pair a browser or installed PWA over your private network, tailnet or an explicit HTTPS tunnel. Remote access is revocable; direct localhost access remains an implicit non-revocable server-host client.

The agents can wait one second.

Model the work first.

Understand the system. Design the work. Then run the agents with a plan worth executing.

Install Handraise Join the Product Hunt launch

Understand the system. Design the work. Run the agents.

More agents don't fix a bad plan.

Folders are not responsibilities

Parallel isn't automatically useful

Activity is not an outcome

Model the work
before agents run.

From codebase to coordinated work.

See what exists—and how sure we are.

Turn product intent into a reviewable plan.

Start only work that is actually ready.

Five truths.
Never silently mixed.

Your codebase isn't a prompt by default.

A serious local control plane.

Fleet command center

Safe worktrees

Permission control

Remote clients

Honest degradation

Outcome learning

Raise your hand.
Take back the plan.

Before you hand over the keyboard.

Model the work first.

Understand the system. Design the work. Run the agents.

More agents don't fix a bad plan.

Folders are not responsibilities

Parallel isn't automatically useful

Activity is not an outcome

Model the workbefore agents run.

From codebase to coordinated work.

See what exists—and how sure we are.

Turn product intent into a reviewable plan.

Start only work that is actually ready.

Five truths.Never silently mixed.

Your codebase isn't a prompt by default.

A serious local control plane.

Fleet command center

Safe worktrees

Permission control

Remote clients

Honest degradation

Outcome learning

Raise your hand.Take back the plan.

Before you hand over the keyboard.

Model the work first.

Model the work
before agents run.

Five truths.
Never silently mixed.

Raise your hand.
Take back the plan.