Technical integration

Your tools. Your infrastructure.
Your authority.

Forge connects to the systems where your AI does the work. Review the integration boundary, evidence and acceptance path with your technical team.

Explore ForgeBusiness fitTechnical integrationSame Forge. Choose the detail you need.
ArchitectureConnectionsCapabilitiesData & securityEvidenceAcceptance
The operating picture

Put the check
where the work happens.

CUSTOMER-CONTROLLED ENVIRONMENT
01 / PROPOSE

Your AI workflow

Prepares an action and supplies the relevant evidence.

02 / CHECK

Forge Orbital

Checks the evidence and rules. Returns a disposition and decision evidence.

03 / APPLY

Your enforcement point

Permits, holds or refuses the action. People retain authority.

04 / REVIEW

Your reviewers examine the evidence, rule, result and approval. You agree where evidence lives and how long it is retained.

An external AI provider remains a separate data boundary. Hosting Forge locally does not make a separate external AI service local.

ChatGPT, Claude and your own agents

Keep the tools.
Add the checks.

We scope connectors and wrappers around the AI workflows you use. You don’t need to know the technical terms to start.

CONNECTOR

The link to your tools.

Passes the proposed action and relevant information from a connected tool to Forge.

WRAPPER

A checkpoint around the work.

Adds a check before a connected action proceeds, then returns the result to your application.

Your team keeps its familiar tools. We work with you or your integration partner to agree the connection, permissions and review steps.

Where the connection and data boundary sit

ChatGPT and Claude connection options depend on the product, account permissions and workflow. This is a scoped integration—not a claim that Forge monitors every chat or that an app-store plugin is approved. Customer-hosted Forge can run within the agreed on-premises or private-cloud boundary; any external AI provider remains a separate data path. Your application enforces Forge’s result.

Connection scope

A defined checkpoint.
Not blanket monitoring.

ChatGPT, Claude and agent systems

The connection depends on the APIs, account permissions and controls available in the tool you use. This does not automatically monitor every conversation.

Customer applications

The application supplies the proposed action and relevant evidence. Its protected workflow must enforce the agreed controls; a returned result alone does not stop execution.

Confirm the exact product, version and deployment environment during the integration review. This page is not a universal compatibility or qualification claim.

Integration scope

Governance. Evaluation.
Optimization.

01 / GOVERNANCE

Check before AI acts.

Supply the proposed action, evidence, rules and required approvals. Forge returns a disposition; the connected application enforces it.

Review: action binding, enforcement and exception handling.
02 / EVALUATION

Check before you change it.

Compare agreed cases and evidence using representative inputs and expected outcomes. Review changes before relying on a new model or agent version.

Review: test cases, expected outcomes and the acceptance criteria.
03 / OPTIMIZATION

Review quality and cost.

Use supplied usage, prices and evaluation evidence to recommend candidates. Optimization does not automatically switch models or guarantee savings.

Review: price basis, quality requirements and approval before changes.

Scope the capability around the work. The customer keeps authority over execution and changes.

Data and security review

Keep the boundary explicit.

Hosting

Tested on NVIDIA hardware.

Forge runs in customer-controlled environments, on-premises or in a private cloud. Agree infrastructure and access for your deployment.

Data use

Forge does not train product models on your data or learn across customers. External model providers retain their separate data paths.

Evidence and custody

Agree the evidence store, retention period, signing identity and authorized reviewers with your team. Obtain trusted verification material through a separate channel.

Operational ownership

Agree permissions, exception routing and support responsibilities. Connection and check failures require their own escalation path.

Questions to settle in the deployment review
  • Which application controls the action, and can it enforce a hold?
  • What may Forge access, and where will evidence be retained?
  • Who controls keys, rotation, deployment changes and rollback?
  • What is the agreed behavior during outages or timeouts?
  • What operating environment, capacity and support arrangements must be qualified?

These are review questions, not claims that a particular customer environment has already passed qualification.

Report a vulnerability or read the disclosure policy →
Evidence you can inspect

Review the basis.
Verify the integrity.

UNDERSTAND THE RECORD

What was checked?

Inspect the supplied evidence, applicable rules, selected result, uncertainty and human-review boundary.

Read the annotated sample →

The existing sample uses illustrative signature bytes. It is an explanation, not a verifying artifact.

SEPARATE CHECKS

Verification and replay.

Independent verification checks the record’s signature and integrity against separately trusted key material. A reviewer need not rely only on the producing service’s own verification response.

Replay uses the retained inputs, pinned Forge engine and configuration to check reproducibility. It does not promise deterministic output from an external model. Neither check establishes that the source evidence was true.

See the integration acceptance checklist →
For your acceptance review

An AI provider’s explanation is not a substitute for independently checkable evidence. Request an executable verification example, trusted key material and the deployment-specific integration contract. A website illustration is not evidence of a qualified customer integration.

Before relying on it

Agree what must pass.

ExerciseWhat the review must establish
Expected actionThe supplied evidence and rules produce the agreed result.
Missing approvalThe connected application holds the action and routes it to the right person.
Unavailable checkThe workflow follows the agreed failure policy; silence is not treated as proof of safety.
Changed action or evidenceThe result is tied to the exact action being considered, not a different request.
Edited recordThe agreed verifier detects the change using independently trusted verification material.
Model or cost changeRepresentative evaluation evidence and required approvals support the proposed change.

Proposed acceptance checklist to agree with your team. Not a report of completed customer qualification.

Your next step

Forge is ready for your business.

Talk to Forge about deploying in your environment. Bring your application, action boundary and evidence requirements; we’ll work through the integration with your systems owner.