Platform

The platform enterprises use to put AI into production.

Everything your teams need to build agents, workflows and internal apps, with the governance your security and compliance people will ask for. Runs in our EU cloud or inside your own infrastructure.

Three pillars, one platform.

Built to work together from the first agent to production.

Glass orb representing a Timbal AI agent

Workflow

12 steps · 840 ms

Trace

1.2 s · 6 spans

Salesforce

14 objects

Prompt

Summarise the ticket

Intelligence

An open-source Python framework for agents and workflows, with observability built in.

Workflow

12 steps · 840 ms

Trace

Salesforce

14 objects

Prompt

Summarise the ticket

Intelligence

An open-source Python framework for agents and workflows, with observability built in.

Accounts at churn risk

Plan

SQL + vector · 4 ms

SQL

accounts · 2 rows · 38 ms

Vector

12k chunks · 41 ms

Q3 account review.pdf

chunk 4 812

NPS on Helios Bank dropped 18 points in Q3. Flag the account for renewal.

Knowledge

Proprietary knowledge bases where vector search and SQL work as one. Documents and structured data, queried together.

Timbal Interfaces pillar: abstract gradient background

Teams

Slack

Voice

Claims desk

Built in Composer

Status of claim 4812?

via WhatsApp · 09:41

Approved this morning. Payout is scheduled for Thursday.

policy.pdf · §4.2

Claim 4812

Status

Approved

Amount

€2,400

Payout

Thu 25

Salesforce · live

Listening

Interfaces

Teams, Telegram, voice and custom GUIs. Channels your team already uses.

Built to pass the security review.

Choose where Timbal runs, keep control of model keys, and give security teams the evidence they need.

Cloud, VPC, or on-prem

Run Timbal on AWS, Azure, GCP, inside your VPC, or fully on-premises. The same API and governance model follow every deployment.

Governance and compliance

SOC 2 Type II in progress, ISO 27001, GDPR and EU AI Act. Encryption at rest and in transit, audit logs, and evidence ready for review.

EU data hosting

Data stays in EU regions with EU-based processing, residency guarantees, and the contractual controls regulated teams ask for.

Bring your own keys

Model calls run on your own OpenAI, Anthropic, or Google accounts. Keys stay in your vault; Timbal never stores them.

Idea to production, governed.

Built once, reused by every use case that follows. Permissions, integrations, evals and audit are part of the platform, so the tenth agent launches faster than the first.

Timbal Platform

Intelligence

Knowledge

Interfaces

Observability

Deployments

Intelligence

Write agents in the open-source Python framework, or describe them to Composer. Any model provider, typed tools, ACE guardrails and an eval suite on every change. Nothing reaches production until it passes.

Design, deploy, and measure in one place.

From the first prompt to production analytics. Configure, test, and watch every run without gluing tools together.

Design your ideal workforce

Configure models, knowledge bases, system prompts, tools, and workflows, then connect them to the systems your team already uses. All from one place.

Financial Reconciliation›Scaffold financial reconciliation workspace

Interface ×New tab × Preview ⌃ Publish ⌄

Financial Reconciliation

Build prompt: Build a production-quality financial reconciliation workspace for comparing bank transactions against accounting records. The first version must accept two CSV uploads, reconcile them, flag exceptions for human review, and present a clean matched/unmatched report with an audit trail. Preserve the product scope exactly: CSV input, bank-versus-accounting comparison, human review of exceptions, and clear reporting. Do not auto-adjust accounting records.

1. Product summary

Create a standalone reconciliation workspace for finance operators. A user uploads one bank-transactions CSV and one accounting-records CSV for a selected reconciliation period, starts a run, and receives a transparent report showing matched transactions, unmatched bank transactions, unmatched accounting records, duplicates, and amount/date discrepancies. The most important outcome is reducing manual spreadsheet comparison while keeping every exception explainable and reviewable by a human. The core journey is: open workspace → upload the two files → map/confirm columns → run reconciliation → inspect summary and matched/unmatched rows → review or resolve exceptions → export the report.

Use exact, plain-language finance copy. Never claim that an exception is definitively an error; call it an exception or review item. Preserve source values and never silently mutate imported records.

2. Knowledge base and data model

Use the fresh knowledge base for structured reconciliation data. Keep imported rows immutable and store every match decision with the rule that produced it.

10 builds ⌃ 2 files changed · +92 −25

Message Composer...

Maestro ⌄

Search surfaces...

Build

New agent or workflowCreate one from scratch or from a template

reviewerAgent — shape its prompt, tools and model

reconciliationWorkflow — wire up its steps and run it

InterfaceThe web app your users see, and its pages

Knowledge baseThe documents and tables your agents can search

Project

OverviewEverything happening in this project, at a glance

TracesStep through what each run actually did, span by span

AI AnalyticsCost, latency and volume over time

DeploymentsWhat is live on each environment, and what shipped it

LogsRaw output from your project as it runs

InfrastructureRegions, compute and the keys your project runs on

Financial Reconciliation›Scaffold financial reconciliation workspace

Interface ×New tab × Preview ⌃ Publish ⌄

Financial Reconciliation

1. Product summary

Use exact, plain-language finance copy. Never claim that an exception is definitively an error; call it an exception or review item. Preserve source values and never silently mutate imported records.

2. Knowledge base and data model

Use the fresh knowledge base for structured reconciliation data. Keep imported rows immutable and store every match decision with the rule that produced it.

10 builds ⌃ 2 files changed · +92 −25

Message Composer...

Maestro ⌄

Search surfaces...

Build

New agent or workflowCreate one from scratch or from a template

reviewerAgent — shape its prompt, tools and model

reconciliationWorkflow — wire up its steps and run it

InterfaceThe web app your users see, and its pages

Knowledge baseThe documents and tables your agents can search

Project

OverviewEverything happening in this project, at a glance

TracesStep through what each run actually did, span by span

AI AnalyticsCost, latency and volume over time

DeploymentsWhat is live on each environment, and what shipped it

LogsRaw output from your project as it runs

InfrastructureRegions, compute and the keys your project runs on

Financial Reconciliation›Scaffold financial reconciliation workspace

Interface ×New tab × Preview ⌃ Publish ⌄

Financial Reconciliation

1. Product summary

Use exact, plain-language finance copy. Never claim that an exception is definitively an error; call it an exception or review item. Preserve source values and never silently mutate imported records.

2. Knowledge base and data model

Use the fresh knowledge base for structured reconciliation data. Keep imported rows immutable and store every match decision with the rule that produced it.

10 builds ⌃ 2 files changed · +92 −25

Message Composer...

Maestro ⌄

Search surfaces...

Build

New agent or workflowCreate one from scratch or from a template

reviewerAgent — shape its prompt, tools and model

reconciliationWorkflow — wire up its steps and run it

InterfaceThe web app your users see, and its pages

Knowledge baseThe documents and tables your agents can search

Project

OverviewEverything happening in this project, at a glance

TracesStep through what each run actually did, span by span

AI AnalyticsCost, latency and volume over time

DeploymentsWhat is live on each environment, and what shipped it

LogsRaw output from your project as it runs

InfrastructureRegions, compute and the keys your project runs on

Meet customers where they are

Timbal keeps the full context. Customers never repeat themselves across voice, chat, and the interfaces you build.

Deploy with confidence

Define guardrails, run simulations, and score every change against your eval suite before anything reaches a customer.

Measure performance and improve outcomes

Track eval scores, latency, model usage and success rates with built-in analytics, or dig into full traces, run history, and debug logs.

Simple to build.

Model, tools, knowledge, and guardrails in one place. No glue code.

Model agnostic

Route each task to the model that fits it: OpenAI, Anthropic, Google, Mistral, Llama, or any OpenAI-compatible endpoint. Change providers without redesigning the app.

Custom instructions

Write persona, tone, and escalation rules in a markdown file. What the operator writes is what the model reads.

support-agent.md

#Support specialist

You handle ACME customer conversations about orders, billing, and returns. Always reply in the customer's language and cite the source document when you quote a policy.

#Escalations

Refunds above $500, legal questions, and anything flagged as fraud go to a human teammate with the full conversation attached.

sales-assistant.md

#Sales qualifier

You pre-qualify inbound leads for the B2B sales team. Ask about company size, use case, and timeline. Score urgency from 1 to 5 before routing.

#Boundaries

Never discuss pricing or contract terms. If asked, share the public pricing page link and book a call with the account executive.

data-analyst.md

#Analytics assistant

You query the data warehouse and return insights in plain English. Always include the SQL used so the analyst can verify. Round percentages to one decimal.

#Guardrails

Never run DELETE or UPDATE statements. If a query would scan more than 10M rows, ask for confirmation before executing.

content-reviewer.md

#Content moderator

Review user-submitted content against the community guidelines. Flag posts that contain hate speech, misinformation, or personal data leaks.

#Tone

When rejecting content, explain the specific guideline violated. Be firm but respectful. Users should understand what to fix, not feel attacked.

Create agents on the web

Upload SOPs, knowledge bases and instructions so your team can build, test and deploy agents from the studio.

Get started

Support Agent

Hey there.

What can I help you with today?

Can you check the status of claim 4812?

Approved this morning. Payout is scheduled for Thursday.

Build with APIs

Integrate Timbal through APIs and SDKs to provision agents, connect systems and run workflows in production.

Explore docs Get API key

from timbal import Agent

async def main():
    agent = Agent(
        name="claims_desk",
        model="anthropic/claude-sonnet-4-6",
    )
    await agent("Status of claim 4812?")

main()

Control with ACE

The Action Control Engine sits in front of any model. Same input, same path — the playbook is the contract.

Learn more

ACE™

Create agents on the web

Upload SOPs, knowledge bases and instructions so your team can build, test and deploy agents from the studio.

Get started

Support Agent

Hey there.

What can I help you with today?

Can you check the status of claim 4812?

Approved this morning. Payout is scheduled for Thursday.

Build with APIs

Integrate Timbal through APIs and SDKs to provision agents, connect systems and run workflows in production.

Explore docs Get API key

from timbal import Agent

async def main():
    agent = Agent(
        name="claims_desk",
        model="anthropic/claude-sonnet-4-6",
    )
    await agent("Status of claim 4812?")

main()