Skip to content
Thaliq
ES

Thaliq · Framework for AI agents

Agents that work. People who decide.

Thaliq is the framework to build, evaluate and observe AI agents. They do the work; what is delicate waits for a person.

Request access See pricing
Watch an agent from start to finish. At the end, you decide.
supportv15 · candidate
agents/support.tq.yaml · saved
apiVersion: thaliq.com/v1alpha1kind: Agentmetadata:  name: supportspec:  model: { provider: bedrock, name: global.anthropic.claude-haiku-4-5-20251001-v1:0 }  system: Help customers with their orders. Never promise a refund without approval.  tools:    - ref: Tool/orders-lookup    - ref: Tool/issue-refund  memory:    longTerm: { store: pgvector, collection: refund-policies }  policy: { ref: Policy/refunds-need-approval }
Checked when you apply it.✓ valid

What Thaliq understands

Model
Claude Haiku 4.5 · Bedrock
Instructions
Never promises a refund without approval
Tools
orders-lookup, issue-refund
Knowledge
refund-policies
Rule
Refunds wait for a person

It is a file: keep it in Git and review it like any code.

Declare. A file describes the agent: its model, instructions, tools, knowledge and rules.

A scripted example, not a live agent.Step 1 of 4

Works with

  • Amazon Bedrock
  • Anthropic
  • OpenAI
  • Ollama
  • vLLM
  • MCP
  • OpenTelemetry
  • A2A
  • pgvector

The hosted service answers with Claude Haiku 4.5 on Amazon Bedrock. The other model providers are for dedicated deployments.

Three products. One framework underneath.

01 · Products

  • A meta-agent proposes changes; you publish them
  • Organizations, roles and versions
  • Lens in every plan
  • Evals, guardrails and the inbox with Enterprise
Request access

What you just saw, from the inside.

02 · Framework

Four layers and Lens watching all of them. Pick a piece to see what it does.

Runtime

where each conversation runs

Runs each conversation with state: if something fails, it resumes from the last step.

  • A checkpoint at every step
  • Pauses for a person to approve
  • Sub-agents and the A2A protocol
policies/refunds-need-approval.tq.yaml
apiVersion: thaliq.com/v1alpha1kind: Policymetadata:  name: refunds-need-approvalspec:  hitl:    money:      approvers: [role:supervisor]      timeout: 30m

Evals, guardrails and the conversation inbox come with Enterprise. See pricing

You decide where each piece lives.

03 · Where it runs

The hosted service is what the plans give you. For your own cloud, or a network with no internet, a dedicated deployment is arranged with us. Pick a mode and see what stays inside your network.

Hosted by Thaliq

We run the framework. Your channels and your app connect through the API and the widget, and your CRM receives signed webhooks. Agents answer with Claude Haiku 4.5 on Amazon Bedrock.

index.html
<script src="https://widget.thaliq.com/widget.js" async></script><thaliq-chat publishable-key="tq_pk_…"></thaliq-chat>
See pricing
Pieces inside your network
x

Thaliq Cloud

  • Consolecontrol
  • Gatewaychannels · API
  • Runtimestate
  • Lenstraces
API · widget · signed webhooks

Your network

your app · your CRM

Only your app and your systems live here

    Model provider

    • ModelBedrock

    How we protect your data and your agents

    Made for people who write code.

    04 · Developers

    Your agents are files: they live in your Git, get reviewed in a pull request and are applied through the API.

    Declarative
    Every agent, tool, policy and eval is a resource with its own apiVersion.
    Checked when you apply
    The API validates every manifest and points to what is wrong.
    Evals as a gate
    A new version of an agent with an Eval goes out when it passes.
    A plain REST API
    Open a conversation and stream the answer with curl, or embed the widget in two lines.
    agents/support.tq.yaml
    apiVersion: thaliq.com/v1alpha1kind: Agentmetadata:  name: supportspec:  model: { provider: bedrock, name: global.anthropic.claude-haiku-4-5-20251001-v1:0 }  system: Help customers with their orders. Never promise a refund without approval.  tools:    - ref: Tool/orders-lookup    - ref: Tool/issue-refund  memory:    longTerm: { store: pgvector, collection: refund-policies }  policy: { ref: Policy/refunds-need-approval }
    Review it in a pull request, like the rest of your code.thaliq.com/v1alpha1

    Three rules we do not negotiate.

    05 · What we believe

    1. 01 · Lens

      If you can't see it, you can't trust it.

      Every step is traced, by person and by channel.

    2. 02 · Runtime

      If you don't control it, it isn't yours.

      What is delicate waits for a person, and a dedicated deployment can run the whole stack on your own servers.

    3. 03 · Evals

      If it isn't measured, it doesn't improve.

      A new version of an agent with an Eval goes out when it passes. If it gets worse, the current one keeps answering.

    The latest we wrote.

    06 · From the blog

    1. GuideA version that gets worse doesn’t reach production
    2. EngineeringHow your agent knows things it was never trained on
    3. EngineeringAnatomy of an agentic loop
    See all posts

    Your first agent is one file away.

    Request access and start on the Free plan, or talk to us about running Thaliq on your own servers.

    Request access Talk to us