HARBA / technical

A governed operating layer for AI work across your existing systems.

Harba brings together orchestration, business knowledge, AI models, integrations and human control within one managed platform. It works with the CRM, ERP, communications and operational tools a business already uses rather than attempting to replace them.

ARCHITECTURE AND EXECUTION

Central orchestration and controlled execution.

Harba’s components sit around a central orchestration layer, connecting outward through APIs, webhooks, MCP, scheduled tasks and configured integrations.

Architecture

Customer information, Pipeline definitions, outputs and useful run data remain persistent. Temporary workers are created only when computational work is required, and removed once the job finishes.

Central orchestration
Persistent customer data
Temporary execution workers
Workspace-specific access

Execution

Pipelines arrange AI actions, deterministic code, integrations and human approval points into a controlled sequence. Each step receives structured input, performs its action and passes the result onwards.

Branching, looping, retries and time limits are handled by Harba’s orchestration rather than left to the AI model. A failed step stops the Pipeline unless a configured retry succeeds or an authorised user restarts the work.

Structured Pipeline inputs and outputs
Deterministic branches and loops
Configurable attempts and time limits
Human approval steps
Recorded run history and errors

DATA AND BIG BRAIN

Searchable memory for each customer workspace.

Big Brain stores customer-provided JSON records, searchable through semantic meaning, keywords and structured filters. Pipelines, Teammates, MCP clients and connected applications share the same search capability.

Each customer’s Big Brain information is held in a separate database with unique credentials, and dedicated database infrastructure is available where physical separation is required.

PostgreSQL and vector search
Semantic and structured queries
Customer-defined records and filters
Stable identifiers for updates and deletion
Logical or physical isolation options

MODELS AND PROVIDERS

Model choice at the point of use.

Harba provides access to supported models through OpenRouter. Pipeline builders can select an appropriate model for each AI step and configure the prompt and its parameters.

Model context limits and probabilistic outputs still apply, so larger workloads must be filtered, divided or processed in stages rather than passed to a model as one unrestricted request.

Model selection by Pipeline step
Configurable prompts and parameters
Usage and token measurement
Bring-your-own OpenRouter key
Provider terms considered during implementation

INTEGRATIONS, SECURITY AND CUSTOMER ENVIRONMENTS

Connections, protections and shared responsibilities.

Beyond the Pipeline mechanics above, this covers how Harba connects outward, how access and credentials are protected, and how implementation work is shared with your technical team.

Integrations and extensibility

Harba can receive and return work through REST APIs, MCP, webhooks, scheduled tasks and configured integrations. Pipelines can also use scripts and reusable custom steps where standard connections are not sufficient.

REST API
MCP
Inbound and outbound webhooks
Configured integrations
Custom code and Pipeline steps

Security and data

Application access is checked against the user’s workspace and role. Temporary workers use short-lived access tokens, and long-lived integration credentials stay within the central platform rather than passing into individual jobs.

External communications use encrypted connections. Hosting, model providers, data handling and retention requirements are reviewed as part of the implementation.

Workspace-based access
Short-lived worker credentials
Encrypted external connections
Implementation-specific data review

Working with technical teams

We work with business and technical teams to understand the process, define the data flow, configure the operation and connect the systems involved.

Customer developers can own APIs, webhooks, data contracts, custom code, test data and customer-side deployment. Harba provides the platform knowledge, configuration support and operational context to bring it together.

Collaborative technical discovery
Agreed implementation responsibilities
Customer-owned integrations
Configuration and ongoing support
Separate test workspaces where required

RELIABILITY, OBSERVABILITY AND LIMITATIONS

How Pipeline work is tracked, and where AI still needs care.

Pipeline runs are recorded and inspectable, with checks in place to catch stalled or failing work. The same discipline matters because AI output itself cannot be trusted blindly.

Reliability and observability

Pipeline runs retain step activity, outputs, usage and failure information, and progression checks help prevent work from stalling undetected.

Customers can inspect Pipeline-level information directly. Harba retains additional operational logs to support investigation when a problem originates within the platform or its infrastructure.

Pipeline run status
Step outputs and errors
Usage and cost visibility
Additional support diagnostics

Known limitations

Generative AI output is probabilistic and can be inconsistent or incorrect, and every model has a limit on how much information it can process at once.

Harba addresses this through Pipeline design, deterministic validation, structured outputs, retries and approval steps. Decisions with significant legal, financial, safety, regulatory or reputational consequences should retain accountable human judgement.

AI output is not deterministic
Model context windows still apply
Large workloads require staged processing
Validation should surround AI steps
High-consequence decisions retain human judgement

GET STARTED

Discuss your technical requirements.

Bring your architecture, integration, hosting and governance questions to a working session. We will explain how Harba could be configured around your systems and identify anything that requires further technical validation.

Discuss your requirements