Govern MCP as a controlled access surface
Define who enters from Claude, ChatGPT, or Gemini, which context they operate on, and under which access, secret, and traceability rules.
IAToolkit
One platform to connect, execute, and monitor AI.
Connect AI through MCP, embed agents in your systems through API, and operate every execution with context, security, monitoring, and analytics.
AI to work with your business. IAToolkit Runtime to automate it.
Operational capabilities
What IAToolkit lets you govern in production
On top of that common layer, IAToolkit governs access, document knowledge, data, automation, and sustained operations.
Operational monitoring for agents and executions
Follow runs, errors, activity, latency, and states from the console so teams can operate and support real processes.
Operate multiple domains or clients
Scale the operating model across multiple Companies, business units, or client environments without reinventing the system.
Operate document knowledge and RAG
Move from basic retrieval to collections, ingestion flows, document inventory, and validation tooling over real knowledge.
Manage changes without touching files
Adjust prompts, agents, collections, and configuration from a GUI instead of depending only on scattered files.
Automate work with agents and channels
Use asynchronous agents to execute prompts, ingest documents, and operate channels such as WhatsApp inside the same operating model.
Run repeatable processes with control
Define pipelines over data and documents to normalize, validate, and publish results without losing operational traceability.
Usage and cost visibility
Measure tokens and USD cost with visibility by user or automation task so teams can operate with tighter control.
Connecting systems
Reaching what lives inside your network, without opening it to the internet.
Your databases and internal APIs stay exactly where they are. IAToolkit reaches them through Bridge, a container that runs inside your network and opens an outbound connection to the platform.
Queries over your databases
You model your tables and their metadata once, and the assistant queries with that context instead of guessing the schema. When the domain changes, you adjust the model rather than every prompt.
Calls to your internal APIs
You define an HTTP tool declaratively — path, parameters, and the shape of the response — and it becomes available to your agents like any other. No integration code per system.
The credential never leaves your perimeter
Your internal API token lives in the Bridge configuration, inside your network. The platform never sees it. Which destinations are reachable, with what timeout and up to what response size, is declared on your side.
Multi-model
Changing model means editing a list, not rebuilding the product.
The model is a configuration decision, not a design assumption. Each Company declares its catalogue of available models and which one it uses by default, and that catalogue can mix providers.
A catalogue per Company
You publish the models your operation is allowed to use, each described by what it is good for, and set which one leads.
Anthropic
OpenAI
Gemini
DeepSeek
OpenRouter
Any OpenAI-compatible endpoint
Embeddings are a separate decision
The text and image embedding providers are independent from the conversation model. Switching model does not force you to reindex your knowledge, and you can run your own embedding models.
The core is open source
The IAToolkit core is open source under an MIT licence. The way to verify there is no lock-in is to read the code that connects the models, not to take our word for it here.
Execution
Heavy work does not compete with someone waiting for an answer.
Everything an agent executes runs outside the request, on queues separated by type of work. That separation is what holds the operation together once the load stops being a pilot: when volume rises, execution capacity rises with it and the model of agents, prompts, and sources you already built stays as it is.
One lane per type of work
Agents, document ingestion, and pipelines run on separate queues, and interactive work takes priority over deferred work. Loading thousands of documents does not leave someone who just asked a question waiting.
A process that falls comes back on its own
Execution processes restart with growing back-off after repeated failures, and start in a staggered way so restarting the platform does not hammer your systems.
No task hangs silently
Pending tasks are recovered at start-up, and a periodic sweep catches the ones left half-finished. An agent that failed shows up in monitoring as a failure, not as a task that never ended.
Operations Console
See the interface in depth
Prompts, document knowledge, SQL, integrations, monitoring, and configuration from one surface. The console page walks every screen with the concrete tasks it solves.
This is the surface where the product is operated day to day: publishing an agent, checking why a run failed, adjusting a prompt, or wiring a new source without touching files.
Next Step
See whether IAToolkit solves your operation.
Talk with us if you already have a use case under evaluation, or check what it costs to run it.