NodeFox
AI and agent workflows

Turn an AI answer into completed operational work.

NodeFox composes models, tools, code and human decisions in an executable graph. Coordinate the work around each model call: gather evidence, validate its shape, branch on the result and control downstream actions.

One executable workflow

Illustrative order flow · NodeFox coordinates the dependencies and decisions.

  1. 1. AI interprets

    Read a purchase order or case note. Extract fields and prepare a structured proposal.

  2. 2. Tools retrieve

    Get the account, transaction or policy context through configured APIs and permitted tools.

  3. 3. Code checks

    Apply exact requirements, tolerances and transformations before an action can proceed.

  4. 4. People decide

    Review the evidence and authorize decisions that need judgment. Missing information takes a defined path.

  5. 5. Systems act

    Record the approved result in the destination system and inspect the run behind it.

Illustrative application

One return request. A complete resolution process.

AI reads the request, tools retrieve order context, code checks the policy, and a person can authorize an exception. The permitted action reaches the returns system with its context intact.

Synthetic example · A customer sends an unstructured return request

Return request to authorized resolution

Arrives

Return request R-108

Order
SO-8701: matched
Requested
2 of 10 shipped units
Delivered
8 days ago

Runs

  1. Retrieve the customer message and attached invoice through configured access.

  2. Extract order ID, items, quantities and the stated reason into structured fields.

  3. A scoped API lookup returns shipped quantities, delivery date and prior returns.

  4. Require an order match, remaining returnable quantity and a date inside the configured policy window.

  5. The example policy allows 30 days; both units remain returnable. Routine action authorized by the configured rule.

  6. Draft a response using checked facts; do not invent eligibility or authorize a refund.

  7. Review the source, policy rule and proposed action before granting an exception.

  8. Call the configured returns API with the checked quantities and an idempotency key.

  9. Inspect inputs, checks, decision and destination response.

Synthetic teaching example. Model output is variable; the graph defines the checks and decisions around it. No model or external API is called here.

Result

Written

Return authorization recorded

  • Checked quantities sent to the returns system
  • Response prepared from confirmed facts
  • No refund or other financial action implied

Work coordinated: input gathering, system lookups, repeated checks and approved updates. Select a step to inspect its role.

Human-owned: process rules, exception judgment and authorization. Synthetic records and transitions; no model calls or live system writes.

Mechanism and value

Make the surrounding work explicit.

Different models, one process

Configure extraction and drafting separately. Pass structured findings between them so the second task can use the first result without reconstructing context.

Exact checks in code

Validate required fields and source IDs before allowing the next step. A valid shape is one check; the reviewer still assesses evidence and correctness.

Controlled tool actions

Choose the API or browser MCP tools a step can use. Place consequential writes behind a condition or human decision with the required permissions.

Agents and reusable execution

Let an agent work with a process it can inspect and reuse.

An agent can help build or invoke a defined workflow instead of reconstructing every integration and check in conversation. The execution graph keeps the process explicit.

Models and tools inside the workflow

Conversation steps call configured models. Compatible MCP servers can expose selected tools through the browser environment and its configured gateway. Validate tool results and control consequential actions with the surrounding graph.

External agents working with NodeFox

The supported MCP bridge exposes operations to read, inspect, modify and run networks in the connected workspace. Agent permissions, bridge authentication and the active environment determine which operations are available.

Headless invocation is a separate execution path. Verify supported steps, credentials and operator access for the intended runtime; browser MCP or OAuth configuration does not automatically transfer to a headless machine.

Follow the work · 56 seconds

AI prepares the evidence. The workflow carries the decision.

Operations analysts reconstruct events across orders, shipments and credits to explain a disputed outcome.
Illustrative operational investigation product demonstration
Illustrative workflow · sample data. Product-derived UI; simulated integrations and results.
Transcript and demonstration details

Illustrative execution, not a customer deployment. Narration is AI-generated. Actions and outputs depend on workflow configuration.

  1. 0–8s · A disputed credit needs evidence

    A customer requests six hundred and twenty dollars for four missing units. The evidence spans three systems.

    Investigation IN-530 · disputed credit · Order SO-301 · $6,200 · Customer reports 4 missing units · Credit request: $620

  2. 8–16s · Bring the records together

    Readers retrieve order, shipment and credit records. AI organizes the evidence for exact checks.

    Order: 40 units at $155 · Shipment: 36 delivered · Credit ledger: no prior credit

  3. 16–27s · Test the explanation

    Forty ordered. Thirty-six delivered. Four missing units at one hundred fifty-five dollars supports the claim. No earlier credit exists.

    40 − 36 = 4 missing units · 4 × $155 = $620 supported · Duplicate-credit check passed

  4. 27–35s · Carry evidence into the decision

    The graph connects source records, arithmetic and review. The recommendation follows an explicit decision.

    Order history · Shipment events · Credit ledger

  5. 35–42s · Authorize the recommendation

    The reviewer approves the credit recommendation. Finance retains posting authority.

    Reviewer sees evidence and arithmetic · Approve credit recommendation · Finance retains posting authority

  6. 42–50s · An investigation report you can explain

    The investigation report carries the amount, sources and decision history. It does not claim a credit was posted.

    IN-530 · investigation report · Supported recommendation: $620 · Sources: order, shipment, credit log · Decision and execution history linked

  7. 50–56s · Inspect the record, do not repeat the action

    Review stored evidence. A new execution can repeat external actions.

    Investigation IN-530 · disputed credit · Order SO-301 · $6,200 · Customer reports 4 missing units · Credit request: $620

The executable platform

Design the graph. Follow the work.

The network defines how information moves, which checks run and when an action is allowed. Watch the actual editor and execution experience, or explore the illustrative stages below.

Illustrative NodeFox interface and synthetic data. This local walkthrough explains the graph; it does not call AI, run the NodeFox engine or update external systems.

Reference
CASE-104
Amount (sample USD)
1200

Reader → model interpretation → Data mapping and Code check → human Decision → activated Writer.

Ready to explore

Inspect the input, then run the checks. The Writer remains behind the Decision step. A required reference must be present before review.

Evaluation

Measure the whole task, including review.

Baseline

Track request preparation time, first-pass completeness, time to authorized resolution and correction effort.

Operating cost

Measure model calls, retries, runtime and reviewer minutes. Compare output quality on the same representative requests.

Reuse

Keep the source contract, validation code and review pattern as reusable network components for another request or document workflow.

Prerequisites

Agree source permissions, allowed tools, output destination and an accountable reviewer. Browser MCP tools and headless runner integration have different support boundaries.

Does the graph guarantee that a model is correct?

No. It defines how data moves, how results are checked and when actions proceed. Model quality needs representative evaluation and appropriate human review.

Can we use an existing agent?

Use its supported API or a compatible tool interface where available. Verify authentication, output shape and action permissions as part of the integration.

Can this run headlessly?

Headless execution is available. Choose supported file, API, model and code steps for that environment; browser MCP and OAuth steps have different availability. The deployment page documents the boundary.
Make the opportunity measurable

Value the handoffs between systems.

Choose one request type and count the gathering, checking and re-entry your team repeats. Keep its volume and completion rules consistent when comparing the baseline with the automated process.

Editable hypothetical example · USD

What could less coordination be worth?

Change four assumptions to estimate annual capacity. These are example inputs, not customer results. Calculations stay in your browser.

Annual capacity potentially released
1,950 hours
Annual labor-capacity equivalent
$87,750
Reduction in hands-on time
60%

13,000 cases / year × 9 minutes removed per case ÷ 60. Includes eligibility, exception effort and rework when entered.

Capacity is not automatically cash savings. Value depends on how the released time is used and what the workflow costs to operate.

Bring the workflow around your models into view.

Discuss the inputs, tools, controls and result you need—or start building the graph yourself.