A developer working locally from a laptop at home

Your hardware,in the loop.Claude Code and Codex can delegate complex tasks to an agent running on your own computer.

Public Alpha · 0.14.0DownloadmacOS · Windows ARM64 · Windows x64No public signing identity yet

The outcome

Your agent gains a local execution path.

Claude Code or Codex can delegate bounded work to your own computer.

What Forjal does

Forjal makes local execution ready to use.

It detects your hardware, prepares a compatible model, and publishes what that model does well—and where its limits are.

Who decides

Your agent stays in charge.

It chooses the model and the task, while keeping planning, tools, permissions, and final review.

The handoff

Local work returns to the same conversation.

The selected model runs on this computer. The result comes back to Claude Code or Codex for review.

Watch the handoff

Your agent chooses.Your machine runs it.The result comes back.

Claude Code delegates a codebase investigation. The local agent follows the evidence across code and tests, drafts a fix, and returns it for review.

Preparing the handoff…

A local agent can carry the task,
not just answer a prompt.

The supervising agent sets the mission and boundaries. The local agent carries it forward.

Why local AI stayed separate

It asked you to leave the tool you were already in.

Two developers working together at a laptop

The old path

Too many decisions before useful work.

1Choose a runtime and quantization
2Repoint the whole agent
3Rebuild tools and permissions

Forjal puts local execution inside the workflow instead.

How Forjal works

Four steps. Then you go back to work.

Scroll to follow the handoff.

01Model ready

Forjal prepares.

It detects this machine, prepares a compatible deployment, and verifies the connection to Claude Code or Codex.

02Explicit choice

Your agent chooses.

It reads the model’s best uses, real limits, evidence state, and prompt guidance before deciding what to delegate.

03Running locally

Your machine runs.

The selected model executes the bounded task locally while Forjal shows the observed path and every recorded step.

04Returned

Your agent reviews.

The result returns to the same conversation. Your agent keeps planning, tools, permissions, and final review.

What you can delegate

A local agent can carry the work from first clue to review.

01
Investigate
Trace a bug across code and tests. Return the cause, the evidence, and a proposed fix.
02
Build
Draft an implementation, its tests, fixtures, or documentation from a clear goal.
03
Analyze
Work through files, logs, or documents and return a decision-ready result.

How this is different

Most local AI products run a model. Forjal gives your agent another agent to delegate to.

The local agent can use the context and tools its parent grants.

Read-only workspace tools run locally. External or effectful actions stay under the parent agent's permissions.

Comparison of Forjal with Ollama, LM Studio, Jan, Houtini LM, and Conifer.
Capability	Forjal	Ollama	LM Studio	Jan	Houtini LM	Conifer
Adds a local agent while keeping Claude Code or Codex as the parent	Clear public claim	Not the product's position	Not the product's position	Not the product's position	Clear public claim	Not the product's position
Works with context and tools granted by the parent agentRead-only workspace tools can run locally. External or effectful actions stay under the parent agent's permissions.	Clear public claim	Not the product's position	Not the product's position	Not the product's position	Available with setup or limits	Not the product's position
Detects hardware and prepares the execution path	Clear public claim	Available with setup or limits	Available with setup or limits	Available with setup or limits	Not the product's position	Available with setup or limits
Tells the parent what the local agent is good at	Clear public claim	Available with setup or limits	Available with setup or limits	Available with setup or limits	Clear public claim	Not the product's position
Parent chooses the model, with no automatic routing	Clear public claim	Available with setup or limits	Available with setup or limits	Available with setup or limits	Available with setup or limits	Not the product's position
Shows runtime, provider, and device for each run	Clear public claim	Not the product's position	Available with setup or limits	Not the product's position	Available with setup or limits	Not the product's position
Tracks evidence by hardware cohort	Clear public claim	Not the product's position	Not the product's position	Not the product's position	Not the product's position	Not the product's position

clear public claim
available with setup or limits
not the product's position

Compiled from official public material, August 2026.

Qwen3.8-27B: new agent features and benchmark gains

Engineering notesAugust 14, 20266 min read

Local AI should not replace your agent

Product thinkingAugust 13, 20265 min read

Built for the machine you have

The best path is the one your machine actually proves.

Forjal detects the accelerator available, prepares the matching execution path, and records the runtime, provider, and device used on every run. If that path is unavailable, it returns an error instead of silently falling back.

Apple Silicon
MLX and MLX-LM · Apple GPU via Metal
Snapdragon X
Hexagon NPU via AI Hub GenieX
Intel Core Ultra
OpenVINO · NPU AI Boost and Arc GPU
AMD Ryzen AI
Lemonade · NPU or hybrid
NVIDIA RTX
Foundry Local · TensorRT RTX or CUDA

The bound path may use an NPU, GPU, CPU, or a hybrid of them. It stays explicit and visible in the run.

Your hardware, in the loop.

Give Claude Code and Codex a local agent to delegate to.

Install Forjal, prepare a model for this machine, and run the first delegated mission without leaving the conversation.

Forjal running on a MacBook placed on an orange armchair