AGENT FORGE

Burning Within

A private AI workspace that decides where work should run.
Agent Forge discovers configured models, routes requests using capability and live availability, coordinates agents and subtasks, stores persistent context, and records traces, reviews, approvals, and user ratings in one browser interface.

Request Access → Product Tour →
PRIVATE OWNER-ONLY LAUNCH  ·  CLOUD-FIRST ROUTING  ·  OPTIONAL USER RUNTIMES  ·  PUBLIC SIGNUP DISABLED
SCROLL
Why It Exists

One workspace.
Multiple execution paths.

Different requests need different capabilities, context, tools, and risk controls. Agent Forge keeps those decisions inside one observable workflow instead of forcing the operator to manually move work between disconnected model interfaces.

Without orchestration
  • The operator chooses a provider before the work is understood
  • Provider limits and outages interrupt the workflow
  • Context is scattered across sessions and project tools
  • Subtasks are manually copied between specialist models
  • Quality feedback is disconnected from future routing
  • High-risk actions lack a consistent review boundary
Agent Forge
  • Live discovery across the providers and runtimes you configure
  • Routing considers capability, health, limits, latency, and evidence
  • Three core slots plus a wider pool for agents and subtasks
  • Persistent sessions, projects, checkpoints, and vector retrieval
  • Trimurti review, Rta signals, decision traces, and user ratings
  • Action previews and approvals for operations that need control
  • Optional authenticated endpoints for user-owned GPU runtimes
  • Data Grid workflows for uploaded or connected structured data
  • Fallbacks are attempted only when a compatible backend is healthy
Decision Flow

What happens when
you hit send.

A request moves through a staged pipeline. The exact path varies by task, available providers, enabled storage, tool requirements, and risk level; not every request invokes every subsystem.

01
Understand
Request Intake
The service reads the request, active session, project context, and requested tools to identify task requirements and policy boundaries.
authenticated request · session context
02
Context
Memory Probe
When persistence and vector retrieval are enabled, relevant session, project, checkpoint, and semantic context can be added to the request.
MongoDB Atlas · vector retrieval
03
Route
Candidate Ranking
Compatible candidates are ranked using declared capabilities, live health, current provider limits, latency, stored quality evidence, and bounded exploration.
registry · health · limits · Karma evidence
04
Coordinate
Review + Agents
Complex work can trigger Trimurti review, dynamic agents, and subtasks. Straightforward requests can remain on a shorter path.
review policy · agent dispatch · subtasks
05
Execute
Selected Runtime
The selected backend handles the request. Compatible fallbacks may be tried when the first candidate fails, rate-limits, or becomes unavailable.
provider adapter · streaming · tool policy
06
Observe
Trace + Feedback
Dispatch, latency, errors, review state, and explicit ratings are recorded where configured. Later routing can use that evidence without changing the underlying model.
traces · ratings · routing evidence
Explore Agent Forge

One system.
Four clear ways in.

Start with the architecture, follow a workflow, inspect runtime options, or walk through the interface. Each destination has its own purpose.

Platform

Routing, Karma evidence, Trimurti review, memory, observability, and the request lifecycle.

Open Platform Workflows

Chat, projects, files, agent subtasks, structured data, traces, approvals, and ratings.

Open Workflows Models & Compute

Cloud providers, optional Ollama, logical core slots, wider pools, and user-controlled GPU runtimes.

Open Models Guided Tour

An interactive walkthrough of the Agent Forge application shell with clearly illustrative sample data.

Start Tour
Access

Private preview.
Access is deliberate.

The current deployment is owner-only. Public signup is disabled and preview access is handled manually.

Owner-only launch

Agent Forge is operating as a private production preview. The platform connects to configured cloud providers and optional user-owned runtimes, so availability and usage limits depend on the services attached to each account.

Submitting a request does not create an account or promise approval. It starts a direct conversation about the intended workflow, required providers, data boundaries, and preview readiness.

  • Multi-model routing — live provider discovery, health checks, limits, and evidence-aware selection
  • Agent workflows — dynamic agents, subtasks, projects, tools, and checkpoints
  • Review and observability — Trimurti, traces, ratings, proposals, and approvals
  • Persistent context — sessions, projects, MongoDB Atlas, and vector retrieval when configured
  • Optional runtimes — authenticated endpoints for user-controlled compatible model servers
Request Access →
Manual review · No public signup · Provider usage and limits vary by configuration
Launch updates

Ask to be notified.

Enter your email to prepare a launch-update request in your email client. This static page does not store or submit your address.

No address is collected by this page.