A collective. Under your control.

YOUR MACHINES.
YOUR AGENTS.
ONE COLLECTIVE.

A shared memory. The tools to act.
Bring your agents into a system that remembers your work — and runs on infrastructure you own.

Build your collective Explore the source

Independent & open source. BORG code: MIT ↗

A dark industrial cube ship floating in space, with layered metallic plates, recessed channels and green illuminated details. — Your collective. In orbit.

YOUR CONTEXT. YOUR COMPUTE. YOUR CALL.Enter the collective

BRING YOUR OWN ACCOUNTS

Different agents.
Shared context.

Choose your clients. BORG connects them to your local memory and tools.

GPT / Codex

NATIVE CODEX SETUP

Codex gets a conductor, MCP tools, and automatic memory hooks. ChatGPT web access is an optional OAuth connection.

OWNER-CONFIGURED

Connect your own Claude Code client through MCP. Included conductor source requires explicit configuration.

OWNER-CONFIGURED

Use a compatible Grok CLI with your own login and MCP setup. Configure and verify the optional conductor separately.

Provider marks identify integrations, not partnerships or endorsements. See each provider’s setup boundaries ↗

01 / HOW IT FITS TOGETHER

One system.
Four connected layers.

Bring your agents, context, and tools together. Each installation starts with its own identity, private data stores, and provider profile.

[ 01 / RECALL ]

A shared local brain.

Semantic memory finds relevant facts. The temporal graph connects entities, relationships, and time. Pinned Qwen extraction and local embeddings power the default brain; your source records remain the authority.

Inside the memory layer

[ 02 / COORDINATE ]

Agents with a plan.

Conductors run agent work. Inbox carries scoped messages and assignments. Beads records the work, its dependencies, and the evidence of completion.

Explore orchestration

[ 03 / ACT ]

Tools on your terms.

Work with files, processes, durable jobs, and a dedicated browser. Native UI tools use normal OS permissions. Explicitly select your configured remote hosts and discover the tools they expose.

See the native tools

[ 04 / LEARN ]

Small models.
Explicit evidence.

Three trained LoRA adapters ship with the source: two for graph extraction, one for memory capture. All remain inactive until exact-base and real-pipeline validation.

Inspect the model cards

INCLUDED RESEARCH / INACTIVE BY DEFAULT

Learning is a process.
Promotion is a decision.

The default brain runs pinned Qwen and a local embedding model. The shipped adapters are optional research artifacts, not a prerequisite for daily use.

None has passed a promotion canary. The capture adapter’s semantic content quality is still unknown. Validate the exact base model and the real extraction workflow before activation.

Read the evidence, including the failures

02 / CHOOSE YOUR FOUNDATION

A home for
the whole system.

Memory, graph, local models, and native tools run together. Plan for the system you will use, with room for your data and other apps.

EVALUATE ON AN EXISTING MAC

Start small.

16 GB unified memory

Plan for 50 GB free SSD

A cautious starter evaluation with a light workload. Observe memory pressure and model latency before depending on it.

Evaluation guidance, not a measured minimum.

PRACTICAL FULL-SYSTEM PLAN

Build your daily brain.

32+ GB unified memory

Plan for 100 GB free SSD

A practical planning baseline for the default local brain, services, and a modest daily agent workload.

Capacity depends on context, jobs, and other apps.

MORE WORK IN FLIGHT

Leave room to grow.

64+ GB unified memory

Plan for 200 GB free SSD

More headroom for concurrent work, larger corpora, and experiments. Add owner-configured machines as your needs grow.

Headroom is not unlimited concurrency.

Planning tiers, not benchmarks. The installer documents at least 20 GB free disk for initial runtimes, models, and caches. These larger allowances include working space; no RAM minimum has been measured.

Daily inference ≠ training. Running the default brain does not train models. Optional adapter preparation adds base weights and an MLX runtime; training needs a separately measured workload and budget.

Full native installation: Apple Silicon macOS verified.
Linux and Intel Mac acceptance pending. Windows unsupported.

Hardware guide & evidence

03 / INITIALIZE YOUR SYSTEM

Make it yours.

Install your own system first. Connect your own accounts. Add web access when you need it.

START HERE

One install.
Your own BORG.

Use an Apple Silicon Mac with at least 20 GB of free disk, an ordinary user session, and your own provider account.

01
Choose your identity.
Replace yourname with a stable lowercase owner ID.
02
Install the local system.
Setup downloads pinned runtimes and models into your private BORG home.
03
Sign in and verify.
Read your owner-specific onboarding plan, complete the normal Codex login, then run the readiness checks.

TERMINAL / NEW INSTALLATION

git clone https://github.com/h3ro-dev/borg.git
cd borg
./install.sh --owner yourname
"$HOME/.borg/bin/borg" onboard
"$HOME/.borg/bin/borg" auth codex
"$HOME/.borg/bin/borg" doctor

Default home: ~/.borg

These commands install and start local services. Codex recall and capture hooks are configured automatically in the dedicated Codex profile. The read-only onboard command prints your next steps; provider sign-in and live verification remain separate.

Full install & recovery guide

OPTIONAL / AFTER LOCAL SETUP

Your brain.
Available from chat.

Connect an existing, running BORG to ChatGPT through your own Cloudflare Access application and tunnel.

01
Verify your local installation.
Keep the host awake and BORG services running.
02
Configure your web connection.
Create your own tunnel, hostname, and Access policy with Managed OAuth.
03
Connect and test.
Add your MCP URL in your client’s available controls, complete sign-in, and test a reversible action.

LOCAL SYSTEM → YOUR GATEWAY → CHATGPT

A connection is not an installation.

An MCP URL connects to a system you already run. It does not install a computer fleet or give access to anyone else’s BORG.

Client plan and workspace controls determine which actions are available. Local health and a successful web sign-in are separate checks.

Follow the web setup guide

Configuration belongs on your machine. This website never asks for account credentials, tokens, or tunnel secrets.

EXISTING INSTALLATION

Trust the checks.

Inspect readiness, discover the current tool schemas, and read your system’s status before asking agents to act.

Health checks distinguish responding services, authenticated memory, trusted hooks, and a pinned provider login. A running process alone does not prove a completed operation.

Operation & recovery

TERMINAL / INSPECT

"$HOME/.borg/bin/borg" onboard
"$HOME/.borg/bin/borg" doctor
"$HOME/.borg/bin/borg" tools browser_
"$HOME/.borg/bin/borg" call borg_status
"$HOME/.borg/bin/bd" list --json

Examples use the default home. If you chose a different location, use that home’s bin/ directory. Inspect operation receipts before repeating a write after a lost connection.

PLATFORM STATUSApple Silicon macOS VerifiedLinux & Intel Mac Native acceptance pendingWindows Unsupported

04 / OPEN SOURCE. OPEN TO INSPECTION.

Your system.
All the way down.

Read the code. Inspect the model cards. Build your own collective from the public repository.

Get the source on GitHub Start your installation

BORG code + shipped adapters

MIT licensed. The public source and included adapter weights are yours to inspect and use under the license.

Read the MIT license

Base models + dependencies

Upstream components retain their own licenses, including FalkorDB’s SSPL. Model weights and provider services have their own terms.

Third-party notices

Your installation + your data

Fresh identities and empty data stores. Your provider accounts, machines, and permissions. No automatic connection to another owner’s fleet.

YOUR MACHINES.YOUR AGENTS.ONE COLLECTIVE.

Different agents.Shared context.