Albor Forge / Software engineering agent newNow with Codex: choose your agent provider

Your software engineering agent.
Integrated with your team's harness.

An engineering agent integrated with your team's harness: goals, plans, coding agents and deliveries with clear owners. With Claude Code, Codex and your company's reusable knowledge in Company Brain. Six integrated modules, from the task queue to CI/CD, built on the AI subscriptions your team already pays for. Host it yourself or use our infrastructure: same binary, same API.

Claude Code and Codex Your AI subscription provides the compute Credentials are never proxied
// Team harness new

Your team decides. Its agents execute. Everyone on the same plan.

The team harness is Forge's operating system, independent of any AI provider: architects turn a goal into signed change plans, each developer executes their route with their preferred agent (Claude Code or Codex), and business users see “How are we doing?” without reading a single technical metric.

agents Claude Code Codex + other CLIs through an adapter · each developer's subscription provides the compute
live · people, teams and roles coming soon · goal plan, routes and “How are we doing?”
Architect role · signing

Decides and signs.

  • Turns a goal into change plans with /sd.objective: AI proposes the breakdown; the architect edits it.
  • Only architects sign artifacts and authorize production deployments.
  • Route board: who owns what, what's ready and what awaits a signature.
Developer role · execution

Takes the next ready plan.

  • “My route”: their plans in order, assigned even while some are still being refined.
  • Ready means unblocked and with a clean semantic review. No false ready status.
  • Their agent, Claude Code or Codex, receives the route position as context: it knows what came before and what comes next.
Business role · read access

Sees how we're doing. That's all.

  • Seven flow signals and zero technical metrics: no tokens, pipelines or diffs.
  • No individual rankings. Measures the team, not its people.
  • Allowlisted access: every new endpoint is closed to this role by default.
// The journey

From a goal to production, with a signature at every step.

$ /sd.objective goal › “launch the public API for integrators”
  1. 01 / goal

    Name the outcome.

    If the team hasn't declared its composition, Forge asks for it first.

    signature · architect
  2. 02 / breakdown

    AI breaks it into plans.

    Proposes child change plans. The architect edits and signs.

    signature · architect
  3. 03 / assignment

    Each developer gets a route.

    Ordered sequences of plans, assigned to a person.

    signature · architect
  4. 04 / refinement

    Refine while you build.

    The architect refines each child with /sd.resume. The route doesn't wait for everything to be ready.

    signature per artifact
  5. 05 / execution

    The agent implements.

    The developer takes the next ready plan and their agent, Claude Code or Codex, executes it against the signed scope.

    developer + agent
  6. 06 / production

    Ship with a named owner.

    Only an architect authorizes deployment. Every signature is attributed to its person.

    signature · architect

Forge creates child plans and routes only once, even if an approval is processed twice. A signature covers exact content: if the artifact changes, the signature is invalidated.

// “How are we doing?”

Need more developers or architects? One view answers it.

The question business teams always ask, rarely answered with data. Forge measures team flow and recommends where to add people.

  • Every recommendation cites its value, threshold and supporting case count.
  • Waiting times use Colombian business days, including public holidays.
  • History survives plan archiving.
  • Sensible default thresholds, adjustable per team.
How are we doing? · Platform team example data · week 39
Platform team flow signals, example data from week 39
SignalValueThresholdStatus
Ready plans not started · median wait4.5 business daysthreshold 3▲ above threshold
Ready plans per available developer3.0threshold 2▲ above threshold
Signatures waiting for an architect0.5 business daysthreshold 1✓ within threshold
Pending production approvals0.4 business daysthreshold 1✓ within threshold
Plans in progress6
Production deliveries per week4
From goal to production18 business days
recommendation

Add developers. There are 3.0 ready plans per available developer (threshold 2), waiting 4.5 business days to start (threshold 3). Based on 12 cases.

agents
Claude Code or Codex, per developer
compute
each person's subscription
signatures
architects only
production
authorized by an architect
attribution
each signature and session, linked to its person
invitations
by email · single use · 7 days
teams
per client, inherited by their projects
no team
everything works as it does today

Your team already has agents. It needs the harness.

Invite your team by email, define architects, developers and business users, and work on the same change plans. During beta, team slots open by invitation.

I want the agent for my team →
// How it works

Three steps. Ten minutes. Your subscription connected.

All six modules follow the same pattern: the backend orchestrates; your local agent executes. We never modify your agent's binary (claude or codex) or your credentials: the agent reads them where the CLI stored them, on your disk.

01 / Authentication

Connect your subscription.

Sign in with Google or a magic link. Each developer connects their own subscription: Claude for Claude Code, or ChatGPT for Codex. The agent reads the CLI's OAuth session on your disk (~/.claude/.credentials.json or ~/.codex/auth.json) without proxying it. We never ask for your API key.

02 / Installation

Install the local agent.

A native binary for Windows, macOS or Linux. Register it with a token; it connects over Tailscale or the public internet.

# macOS / Linux
$curl -fsSL forge.sh/install | sh
$forge agent register $TOKEN
03 / Dispatch

Queue it and carry on.

Dispatch from web, CLI or iOS. Live output, rate limits, cost and generated commits. Close the tab — we'll notify you with push.

$forge push "refactor scheduler.ts into a class"
# → enqueued · task t_4f8a2c
// The platform

Forge is an operating system. Your AI subscription provides the compute.

Each Forge module is an OS layer: the queue, secrets, network, CI/CD, skills and scheduler with advisor. All available in two forms — run your own instance or use ours. Same binary, same API.

the system

It's not a tool catalog. It's operating system where every layer assumes the others exist.

Your secrets live in Forge Secrets — injected when the agent starts. Deployments run through Forge CI/CD — and Advisor knows how much usage remains in your subscription window. Your team connects through Forge VPN — with resources behind it. Skills learns from the agent's work — and Projects turns it into backlog. One connected system running on your subscription.

what it isOS for teams and agents
modules6 integrated layers
deploymentself-hosted + managed
credentialsnever proxied
computeyour AI subscription
vendor lock-innone
// The six modules

Six layers. One connected system.

Each module has its own domain and assumes the others exist. Same binary in the cloud and on your infrastructure. No features exclusive to our infrastructure.

Forge CI/CD

v0.4 ● live

CI/CD at the speed of your AI, without paying for minutes like a thousand-person company.

Pipelines that start as fast as the agent commits. EC2 Spot build agents, sccache shared on S3, automatic shutdown when idle. Eight-minute Rust builds drop to 2m 40s. Host your own runners or use ours.

build→ test→ deploy
✓spot agent · 12s cold start ✓sccache hit · 78% ⠋deploying · 2m 41s
self-hosted managed
docs ↗

Forge Secrets

v0.3 ● live

Credentials your AI can always access. Never pasted into prompts.

Encrypted multi-tenant vault AES-GCM, RLS per project. Forge injects secrets into the agent runtime — the agent uses them but doesn't see them in context. No more copying API keys into chat.

●STRIPE_SK sk_live_•••• ●OPENAI_API_KEY sk-proj-•••• ●DB_PASSWORD •••••••• ●AWS_ACCESS_KEY AKIA•••• — injected at startup · invisible to the prompt
self-hosted managed
docs ↗

Forge VPN

v0.5 ● live

Private networking out of the box. Your development stays off the public internet from day one.

Your repository, database and staging environments — all behind Forge's private network. Share it with your team and control who can access what with role-based ACLs. No accidental public exposure.

team3 peers
↔
vpnforge-mesh
↔
resources12 nodes
acl · by role exposed · 0
self-hosted managed
docs ↗

Forge Skills

v0.2 ● beta

Skills that write themselves. Your AI improves with every execution.

When Forge learns something new in a task — a refactoring pattern, a repeated command or a sequence of fixes — it saves it as a reusable skill. The next task that needs it has it loaded.

.forge/skills/
├ refactor-billing-service.md12 runs
├ run-cargo-check-and-fix.md7 runs
├ deploy-spot-and-warm.md3 runs
└ extract-rls-policy.mdnew
+ 14 more · all auto-generated
self-hosted managed
docs ↗

Forge Projects + Advisor

v0.6 ● live

Your AI subscription working autonomously, aligned with the goals you set.

Register projects with roadmap goals. Advisor generates backlog aligned with those goals — executes automatically or asks for approval, as configured. Quick replies from iOS or web when the agent needs a human decision.

project · billing-svcQ2
→ migrate to RLSauto
→ extract invoice domainauto
→ deprecate /v0/authapprove
claude ›Should I delete the legacy endpoints /v0/auth?
No Yes · proceed later
self-hosted managed
docs ↗
// The product

Full visibility, from anywhere.

Web, CLI, iOS — real parity across interfaces. The hero kanban is just the start.

Usage · 5h and 7d windows

/usage
Sonnet · 5h 1.24M / 3.0M (41%)
Opus · 5h 0.68M / 2.0M (34%)
Sonnet · 7d 18.3M / 24M (76%)
Opus · 7d 10.1M / 16M (63%)
Next reset · 2h 47m Plan · Max 20x

Task detail · live output

/tasks/t_4f8a2c
14:02:11→ Starting task t_4f8a2c · sonnet
14:02:13✓ Session restored · session_4b1
14:02:18 Read scheduler.ts · 240 LOC
14:02:22 Edit scheduler.ts · +84 −31
14:02:29 Bash npm test scheduler
14:02:41✓ 14 tests passed · 0 failed
14:02:43✓ Auto-commit · feat(scheduler): backoff

Deployment detail · ci-deploy

/deploys/283a121…
Running 283a121 stage 3 / 11 · cargo-clippy
2/11 · 1.4s
STAGES
✓checkout294ms
✓cargo-fmt1.1s
▶cargo-clippy…
cargo-test
frontend-build
deploy
cargo-clippy ● streaming
48 Checking rand v0.9.4
49Compiling tokio-stream v0.1.18
50Compiling hashlink v0.9.1
51Compiling unicode-normalization v0.1.25
52 Checking hmac v0.13.0
53Compiling crossbeam-queue v0.3.12
agent · forge-agent-c6a-amd64 webhook · HMAC verified · 2m ago

Native iOS · 1:1 parity with web

TestFlight · com.alborlabs.forge

Queue tasks, review output, approve auto-commits and respond to tasks needing input — all from your phone, with native APNs push even when the app is closed.

Full kanbanFilter by project, model and status.
Approve commitsInline diff, approve with a tap.
Native APNs push24 notification categories you can toggle.
Drag & dropPDFs, images and CSVs from Files.
task.completed task.needs_input ci.failed rate_limit.warn advisor.new +19 more
9:41● ● ●
Forge
⚡RUNNING
2
◷QUEUED
3
✓DONE
480
$SESSION
$31.89
13k in · 365k out
◷AVG
50s
769 tasks
⌥Pipelines(8)
Waiting d28cad0e
CI/Deploy triggered by push to main
2/10 · cargo-clippy 1d ago
// Pricing

Pay per person, based on their role. Each person brings their own AI.

Forge doesn't resell tokens: each developer uses their Claude or ChatGPT subscription, paid directly to the provider. We charge for the operating system. USD/month; save 20% with annual billing.

Free

Try Forge with your own agent.

$0/ month
  • ✓1 person
  • ✓Claude Code or Codex
  • ✓Task queue + local agent
  • ✓5 h build time / month
  • ✓Community support

Pro

One developer with all six modules.

$29/ month
  • ✓1 person
  • ✓All six modules
  • ✓30 h build time / month
  • ✓Dedicated agent: +$39 / month
  • ✓Email support

Enterprise

Multiple teams, compliance and self-hosting.

from$1,500/ month
  • ✓Everything in Team
  • ✓SSO / SAML
  • ✓Complete self-hosting with manifests
  • ✓Dedicated agents and build hours
  • ✓SLA + guided support
  • —Annual contract

The OS saves time; each developer pays for AI directly with their provider. No lock-in, no surprise token costs. Prefer self-hosting? All six modules include Docker / Compose / Kubernetes manifests.

// Comparison

Forge vs. the rest.

Claude Code or Codex already works for you today. Forge handles what a single agent can't, and what happens when five people and twenty agents work on the same product.

Albor Forge
A single agent
Hosted SaaS agents
Agent provider independentnew
✓ Claude Code and Codex
— One per CLI
~ Depends on the provider
Uses your AI subscription
✓ Claude or ChatGPT
✓ Native
— Tokens billed separately
Compute on your machine
✓ Local agent
✓ Local
— On their infrastructure
Dispatch and carry on
✓ Queue + push
— Blocks the terminal
✓ Async
Rate-limit visibility
✓ Per subscription, live
~ Per terminal
— Not applicable
Multiple machines / operating systems
✓ Win/Mac/Linux
~ One session per terminal
— Fixed server-side execution
Compute vendor lock-in
✓ None
~ Bound to its provider
— Complete
A team on the same plansnew
✓ Goal → plans → routes
— One person, one terminal
~ Depends on the provider
Role-based signature before productionnew
✓ Architects only
— Not applicable
~ Depends on the provider
Business view without technical noisenew
✓ “How are we doing?”
— Not applicable
~ Depends on the provider
Company Brain / reusable brains

Your company's knowledge,
in your engineering agent too.

Prepare a Company Brain with your company's documents, decisions and context. Reuse it across engineering, customer support and sales agents, with each team's scope and permissions. Each company keeps its own information space.

// Software engineering · early access

Work with your team and its agents. Join the software engineering waitlist.

Leave your email and we'll contact you when we can support your team's access.

Received. We saved your interest in the software engineering agent for . We'll use your email to contact you about Albor Forge software engineering. The other agents' lists are independent.
// FAQ

Questions we've been asked.

Does it work with Codex or only Claude?
Both. Each developer connects their own Claude or ChatGPT subscription, and Forge dispatches to Claude Code or Codex. A team can mix them: the plans, signatures and routes stay the same.
How much does a team pay?
$99/month for the platform, $49 per architect and $39 per developer. Business users are free. A team with 1 architect and 4 developers pays $304/month, or $242 with annual billing. Each developer's AI subscription is separate, paid directly to their provider.
Who can sign and deploy to production?
Anyone with the architect role on the project's team. Developers execute plans on their routes; business users have read access only. Projects without a configured team keep working as they do today.
What does a business user see?
Only “How are we doing?”: seven team flow signals and a recommendation on where to add people. No tokens, pipelines, diffs or individual rankings. No technical notifications.
Do I need to register everyone on day one?
No. The team declares its size and architect/developer counts before everyone is registered. Each person joins with a single-use email invitation valid for 7 days, using a magic link or Google.
I work with multiple clients. How are teams organized?
One team per client, inherited by their projects. If a project needs different people, it declares its own team.
Which AI subscription do I need?
One per developer, bought directly from the provider: Claude Max (5x or 20x) for Claude Code, or a ChatGPT plan that includes Codex. The backend doesn't run the agent: your local agent runs claude or codex with each developer's credentials. Without a connected subscription, nothing can execute the tasks.
Does it work with Claude Pro or the free plan?
For Claude, no: we assume at least Max 5x. Usage windows (5h / 7d) and agent availability are designed around that plan; Pro / Free reaches rate limits almost immediately. For ChatGPT you need a plan including Codex; OpenAI sets the usage limits, and smaller plans reach them sooner with a task queue.
Does my code go to Albor Labs?
No. Only metadata: prompts, status, timestamps, cost and output text. Source code and the agent session (Claude Code or Codex) stay on your agent, on your machine.
Can I run it entirely on my infrastructure?
Yes — self-hosting is fully supported. Every Forge module uses the same binary in cloud and self-hosted setups; we provide Docker / Compose / Kubernetes manifests. The managed option at alborforge.com gets you started in minutes, with no features exclusive to our infrastructure.
Which operating systems support the agent?
Windows, macOS and Linux. A native binary on each, installed with one command.
How long does setup take?
Less than 10 minutes: sign up → install the agent → register with a token → first task queued.
Does it support multiple machines?
Yes. One agent can run on your laptop and another on an EC2 instance you start for heavy builds. The dispatcher routes by platform and tags.
What happens when I reach the weekly cap?
The scheduler pauses dispatch before a 429 response. You can still use claude in your terminal — Forge reserves headroom for the person.