Open source · Apache-2.0 · v0.6.0

Give AI less to rediscover.
Give engineering more evidence.

FORGE is a credit-aware control layer for deliberate, governed AI-assisted software development. It prepares repository facts locally, bounds context, classifies risk, plans work, preserves approvals, and records verification evidence before the model spends deeper reasoning.

Install FORGE View release
36/36 release tests SHA-256 verified assets Sigstore provenance Node 20+
FORGE terminal-style overview: prepare context, govern execution, verify with evidence
WHY FORGE

Models should reason about the change — not repeatedly reconstruct the repository.

FORGE moves deterministic preparation out of the model loop. The coding agent still writes and reasons; FORGE gives it a cleaner operating boundary.

01

Bounded context

Inspect repository facts and select only relevant context instead of feeding the entire project into every task.

02

Risk-aware planning

Classify work, activate proportional engineering capabilities, and surface approval points for sensitive changes.

03

Reusable work packets

Cache deterministic preparation so repeated work can avoid unnecessary repository rediscovery when context has not changed.

04

Evidence, not vibes

Track run state, validation, verification and privacy-preserving efficiency telemetry instead of treating a chat transcript as the only execution record.

INSTALL

From release to forge --version.

The v0.6.0 GitHub release ships checksum-verifying installers for Linux/macOS and Windows PowerShell. npm registry publication is separate and is not claimed here.

1 · Download, verify and install

curl -fsSLO https://github.com/gODtECH-Ctl-Create/gODtECH-FORGE/releases/download/v0.6.0/install.sh
sh install.sh 0.6.0
forge --version

2 · Initialize inside your project

forge init --dry-run
forge init
forge doctor
forge validate
REAL INSTALL VALIDATION

The install path was tested outside the FORGE repository.

Condensed terminal captures from the first real Windows and Codespaces/Linux installations. These are session excerpts, styled as terminal windows — not stock images.

Windows PowerShell · verified
PS C:\Users\User> Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
PS C:\Users\User> .\install.ps1 -Version 0.6.0
Downloading FORGE 0.6.0...
Checksum verified. Installing FORGE 0.6.0...

added 5 packages in 18s
FORGE 0.6.0 installed successfully.
PS C:\Users\User> forge --version
0.6.0
PS C:\Users\User> node --version
v24.18.0
GitHub Codespaces · verified
$ curl -fsSLO .../v0.6.0/install.sh
$ sh install.sh 0.6.0
Downloading FORGE 0.6.0...
Checksum verified. Installing FORGE 0.6.0...

added 5 packages in 2s
FORGE 0.6.0 installed successfully.
$ forge --version
0.6.0
$ forge doctor
✓ Node.js 24.20.0 is supported.
✓ Git repository detected.
OPERATING MODEL

Intent → context → reasoning → evidence → ship.

01

Prepare

forge prepare --task "..."

Repository facts, bounded context, risk, discovered commands and remaining model work.

02

Plan

forge plan --task "..."

Proportional capabilities, approvals, workflow and ordered engineering steps.

03

Run

forge run start --task "..."

Persist resumable execution state instead of losing the workflow inside one conversation.

04

Measure

forge metrics

Local preparation telemetry: cache reuse, latency, deterministic work and context-reduction estimates.

AI CLIENT INTEGRATION

Provider-neutral over MCP.

FORGE exposes deterministic preparation and run state through a local stdio MCP server. The AI coding client remains responsible for model execution.

forge mcp serve
EVIDENCE LAB

We test the claim before we market the claim.

FORGE telemetry is deliberately separated from provider billing. The first DeployGuard experiment is published as a pilot with its limitations visible; the next milestone turns that learning into controlled multi-task benchmarks.

PILOT · DG-BM-001

DeployGuard incident CRUD

Plain AI control vs FORGE-assisted implementation of PostgreSQL-backed incident management.

Read benchmark
28.6%fewer unique files inspected7 → 5
0FORGE implementation retriescontrol: several
3/3tests passedboth runs
74.7%estimated context reductionFORGE internal metric
What this does not prove: DG-BM-001 was a pilot, not a clean controlled trial. Starting commits differed, the control could read FORGE-generated project files, Docker verification was not equal, and provider token/cost data was unavailable. The 74.7% figure is a FORGE context-selection estimate — not a claim of 74.7% credit savings.
MILESTONE 2

Controlled multi-task benchmarks

Same baseline commit, same task brief, same acceptance criteria and explicit separation between measured behavior, FORGE estimates and provider billing data.

View protocol
DG-BM-002authentication / securityplanned
DG-BM-003database migrationplanned
DG-BM-004GitHub Actions CI/CDplanned
DG-BM-005Kubernetes / probesplanned
SAFETY BOUNDARIES

More capable does not mean less controlled.

Secret-like paths excluded from generated work packets.
Higher-risk work can require explicit human checkpoints.
Discovered commands are reported, not silently executed.
Application technology choices remain with the consuming project.
Completion should be backed by repository evidence.
Metrics avoid task text, source code and repository identity.
START WITH YOUR REPOSITORY

Prepare locally. Reason proportionally. Ship with evidence.

Install v0.6.0Explore GitHub