Public Beta — Join as a Founding User and lock in 50% off while the offer is active. Register and create AI agents

Product · How KlicForge works

KlicForge Agent Runtime

KlicForge runs AI agents on its own runtime. It connects them to your business knowledge and tools, manages each step they take to finish a task, and works with AI models from more than one provider.

Say a visitor asks the chat on your website for a table for four tomorrow at 7. KlicForge checks your opening hours, looks at your bookings, saves the reservation and replies in the same conversation. The AI model reads and writes. The runtime decides what it sees, which tools it may use, and what reaches your customer.

Last updated: 13 November 2026

At a glance

Our runtime is our own

KlicForge isn’t built on a third-party agent framework or platform. The runtime that loads your agent, assembles its context, runs its tools and applies its guardrails is KlicForge code.

Models are called, not in charge

The runtime calls an AI model to understand requests and write replies. Your agent stays in KlicForge.

Your agent lives in KlicForge

Knowledge, Data Tables, skills, tools, memory and rules stay in KlicForge, not in the model.

Your choice of AI

Pick from models served through OpenRouter and Cloudflare. On Business plans and above, connect your own OpenRouter or OpenAI key. Change models in a few clicks; knowledge, tools and workflows stay as they are. Available models depend on your plan.

Guardrails run in the runtime

By default, personal details are protected and messages are screened for known prompt-injection patterns before the AI sees them.

Limits, stated openly

Model inference, embeddings and vector search are external services. See what isn't KlicForge →

What a wrapper does, and what an agent runtime does

These describe two approaches, not particular products.

A simple AI wrapper
KlicForge
A simple AI wrapper compared with KlicForge
A simple AI wrapperKlicForge
Sends your message to an AI model and shows the replyWorks out the request, picks skills and tools, looks up your knowledge when needed, carries out actions and prepares the reply
Business knowledge pasted into a promptSearches your knowledge when a question needs it, and reads and updates your Data Tables through tools
No memory of the customerRemembers customers across conversations (when memory is enabled)
Rules depend on the model behavingGuardrails run in the runtime: personal details are protected by default and known prompt-injection patterns are screened
Tied to one modelChoose and change models; add optional backup models

Built in-house, not on a third-party agent framework

KlicForge isn't built on a third-party agent platform. These are the parts KlicForge designs, builds and runs:

  • Guardrails & privacy

    By default, personal details are protected and messages are screened for known prompt-injection patterns before the AI sees them. Trust Hub →

  • Understanding intent & skill routing

    Works out what the customer actually needs on every message, and brings in the right skills.

  • Skills

    Reusable expertise (e.g. bookings, lead capture) that brings the right instructions and tools for the job.

  • Tools & actions

    Lets agents do things — look up records, send messages, create bookings — with permissions checked on every action.

  • Business knowledge & Data Tables

    Searches your documents and website content; reads and updates structured records, with sensitive fields encrypted.

  • Customer memory

    Remembers returning customers’ preferences and history (when enabled).

  • Scheduled tasks & specialist agents

    Runs reminders and outbound messages on a schedule. An agent can ask a specialist agent to handle part of a task and use its answer in the same reply.

  • Testing & versioning

    Test agents before publishing; publish versions and roll back if needed.

  • Model gateway

    Connects to the AI model you choose, with optional backup models.

  • Workspace isolation

    Each business’s data is kept separate at the database level. Trust Hub →

Architecture

Everything inside the KlicForge boundary is designed, built and run by KlicForge. AI models and the apps you connect sit outside it, and the runtime calls them when it needs them.

  • Solid: built and run by KlicForge
  • Dashed: external service or platform
  • Select any part to see what it does

Where conversations come from

Connectors are run by KlicForge.

KlicForge Agent Runtime

Built and run by KlicForge
Runtime edge

External services the runtime calls

Outside KlicForge

Select any part of the diagram to see what it does.

Read the diagram as a list

1. Where conversations come from

  • Website chat widget (KlicForge)
  • WhatsApp (Meta platform · external; connector run by KlicForge)
  • Telegram (Telegram platform · external; connector run by KlicForge)
  • Team Inbox (KlicForge)

2. The KlicForge Agent Runtime (built and run by KlicForge)

  • Guardrails & privacy
  • Intent understanding & skill routing
  • Skills
  • Tools & actions (Built-in tools + MCP connector)
  • Knowledge retrieval
  • Data Tables
  • Customer memory
  • Scheduler
  • Specialist agents (Delegation)
  • Testing, versioning & publishing
  • Workspace isolation
  • Model gateway, with optional backup models, on the runtime's edge, connecting to AI model providers

3. External services the runtime calls

  • AI model providers: OpenRouter · Cloudflare · OpenAI (your key)
  • Search infrastructure: Embeddings & vector search
  • Connected apps through MCP: e.g. Google Workspace & Calendar, Calendly, DocuWare
  • Hosting & cloud infrastructure

One website enquiry, step by step

Here's what happens when a visitor asks your website chat for a table. The runtime runs every step. It calls an external AI model to understand the request and to decide what to do, and calls it again after each tool runs.

  1. A visitor types into the chat on your website. KlicForge checks the request comes from your site and is within your plan.

Read the example step by step

Example: a visitor to your website asks “Do you have a table for 4 tomorrow at 7?”

  1. Message arrives. A visitor types into the chat on your website. KlicForge checks the request comes from your site and is within your plan.
  2. Load your agent. The published version of your agent is loaded: its instructions, skills, tools and knowledge sources, plus the conversation so far.
  3. Protect & screen. Email addresses and phone numbers are swapped for placeholders, and the message is screened for known prompt-injection patterns.
  4. Understand the request. A quick AI model call works out what’s needed (a table booking for 4, tomorrow, 7pm) and picks the bookings skill and its tools.
  5. Look up knowledge. The runtime decides a search is needed and fetches the restaurant’s opening hours.
  6. Assemble context. Your agent’s instructions, the bookings skill, the opening hours, what it remembers about this customer (if memory is on) and the allowed tools are put together within a size limit.
  7. AI model decides. Your agent’s AI model reads the context and asks for a tool: check bookings for tomorrow at 7pm.
  8. Tools run. The runtime checks the tool is allowed, runs it and returns the result. The model then asks to save the booking, and the runtime runs that too. This repeats until the model writes its reply, up to a set limit.
  9. Reply delivered. Placeholders are swapped back for the real details and the reply streams into the chat.
  10. Saved & counted. The conversation is stored encrypted and the interaction counts towards your plan.
  11. After the conversation: the agent learns (memory). Once the conversation ends, the agent can remember this customer for next time (when memory is on).

When an AI provider has a problem

Built so one problem doesn't become every problem. KlicForge's runtime, your data and your team inbox don't depend on any single AI company.

  • If

    Your main model doesn’t respond

    Then if you’ve added backup models, the runtime tries them in order, and they can come from a different provider.

  • If

    The AI is unavailable

    Then your team can still see every conversation and reply to customers themselves from the KlicForge Inbox.

  • If

    A connected app, such as your calendar, is unavailable

    Then your agent carries on without that one tool.

Backup models are optional and set per agent.

Technical details

For developers and technical evaluators.

In agent terms, the KlicForge Agent Runtime is the agent harness: the software that equips a model with context, knowledge, tools and guardrails, and runs the loop in which the model requests a tool and the runtime executes it. Orchestration is how the runtime sequences those steps for each message: which checks run first, what context is assembled, which tools are offered, and when the loop stops.

Turn lifecycle

  1. Resolve the request: allowed origin, plan allowance, contact, the agent’s published configuration and the conversation history
  2. If a team member has taken over the conversation, stop: the AI doesn’t run
  3. Input guardrails: contact details tokenised, card and ID numbers redacted, rule-based prompt-injection screening (on by default, set per agent)
  4. Intent classification and skill routing, as a separate model call
  5. Tool selection, limited to the tools the agent has been granted
  6. Rule-based retrieval decision, then knowledge search checked against each source’s access rules
  7. Context assembly within a token budget
  8. Model call through the model gateway
  9. Tool loop: permissions re-checked when each tool runs, a timeout per tool, a bounded number of iterations and detection of repeated calls
  10. Delivery: protected values restored for display, plus response validation when it is enabled
  11. Persistence and usage accounting

After the conversation ends, memory and summaries are extracted in batch.

Provider abstraction

  • One gateway over OpenAI-compatible chat-completions endpoints, routing each request by the chosen model’s provider.
  • Per-agent model plus an optional ordered list of backup models, which may use a different provider and are tried in order when a call fails.
  • Tool calls that some models emit as text are recovered.

Delegation

  • An agent can call an allowed specialist agent as a tool. The specialist runs within the same request, and its answer returns to the calling agent, one level deep by default.

Extensibility

  • Skills declare the tools they need.
  • Tools are built-in or connected through MCP (HTTP transport).
  • Tool permissions are enforced on every call, and each tool has its own circuit breaker.

Knowledge and data

  • Hybrid semantic and keyword search, falling back to keyword search.
  • Data Tables with field-level encryption and per-agent permissions.

Isolation

  • Database row-level security per workspace.

Quality

  • Agent versioning with publish and rollback.
  • Evaluations with an optional publish gate.

What isn't KlicForge

Stated limits. The runtime calls these; it doesn't replace them.

  • Model inference, embeddings and vector search are external services.
  • Platform models are served through OpenRouter and Cloudflare Workers AI; OpenAI is available with your own key. There is no direct Anthropic or Google integration.
  • Backup models are opt-in per agent. The runtime has no provider-outage detection: a backup is tried when a call fails.
  • MCP connections use HTTP transport.
Free plan available — no credit card needed

Your AI assistant is ready.
Are you?

Join businesses already saving hours every week with an AI that handles customer conversations around the clock.

Setup takes 10 minutes. Works on WordPress, Shopify, and any website.