Welcome to Braidrun Workflow

If you open Braidrun Workflow for the first time, this page will explain clearly in two minutes: what we are, what we can do for you, who we are suitable for, and who we are not suitable for. Then decide whether to continue reading.

What We Are

Braidrun Workflow is an "Agent-driven AI workflow" construction platform. You use it to turn something that you do repeatedly - such as catching news every day and making a summary, generating ASA reports every week, and submitting PRDs for review by several Agents every time - from "manual script running" to a "scheduled automatic triggering + observability + approval" pipeline.

Where has the industry gone?

As 2026 arrives, the conversation around agentic workflows has shifted from concepts to engineering implementation and return on investment. A few common points of agreement:

1. From single agent dialogue → multi-agent collaboration + system orchestration — Enterprise-level agents with commercial value must possess four core capabilities: intelligent planning, long-term memory, tool invocation, and autonomous action. Simple prompt series is no longer enough. Instead, "complex business requirements are automatically broken down into several sub-tasks, which are completed collaboratively by Agents or modules with different expertise."
2. The last-mile pain points are now fully exposed — Many companies get stuck on three things during real deployment: system integration (ERP / CRM / OA with no modern APIs), knowledge cold start (the high cost of building RAG over unstructured documents and cross-product-line corpora), and enterprise governance (permissions, cost, audit, and isolation, all blank once it hits production).
3. Low-code becomes the standard for Agent orchestration platforms — The people who understand a business process best are usually frontline staff. Visual orchestration plus letting business users define their own processes has become the baseline expectation for platforms like this.
4. Human-in-the-loop — AI steps are orchestrated directly into business processes: people own the review and exception-handling nodes, while AI handles the high-frequency, standardized segments.

What did Braidrun do in response to these consensuses?

• Tool call + system connection — Built-in MCP client/server, Webhook triggers, REST API, and a credential center with provider binding. Agents can genuinely reach your existing CRM / OA / internal APIs.
• knowledge cold start — A built-in RAG tool group indexes documents and directories for semantic retrieval, and the template library spares you the cost of starting a flow from nothing.
• Enterprise level governance — Credential encryption, approval chains, team isolation and plan-based quotas are on by default; audit events are recorded server-side throughout, with the viewer opened on the Enterprise plan.
• Visualization + low code — Canvas drag and drop + YAML two-way synchronization + AI assistant natural language generation, both business personnel and engineers can get started.
• Human-in-the-loop — manual_approval is a first-class step modifier: any step can get an approval gate, with an approval center and email notifications, and approvals can also be handled via API.

Why Naked Agent Is Not Enough

Talk to the Agent directly in the IDE or Chat, and soon you will encounter:

• Not repeatable - the same problem has different results every time, making regression testing impossible
• Unobservable - No one records which tools the Agent adjusted during the thinking process, how many tokens he used, and how much money he spent.
• Non-auditable - when it comes to customer data and fund operations, the compliance team cannot trace them back afterwards.
• Not reusable - each employee "talks to the Agent once", and the same business logic is invented N times
• Ungovernable - Credentials, model quotas, and approval nodes are all in everyone's own chat history, out of control

Workflow fills all five gaps, so Agents can enter a company's real production processes.

Three Platform Advantages

• Visual canvas + YAML two-way editing — Canvas and YAML edit the same workflow definition. Structures and values are convertible into each other; comments, blank lines, and original typesetting are not part of the stable preservation contract.
• Primitives for first-class citizens of AI workflows — single / group_chat / agent_based / classifier / state_machine / sub_workflow are all native step types, while manual_approval is an approval modifier you can add to any step. Describing multi-agent collaboration feels as natural as writing code.
• Production and operation capacity — Code isolation execution, credential encryption, execution recording, approval and debugging capabilities are all on the same platform. Automatic continuation after service restart is also controlled by deployment switches, workflow policies and step idempotent tags.

How can this help you?

The three most common user personas:

① Business Operations · Automate Things That Are Repeated Every Week

• At 8 a.m. every day, grab a batch of industry news from RSS/Twitter, classify it by topic, write a Chinese summary for AI, and submit it to Slack/Feishu.
• Every Monday, Apple Search Ads + Google Ads data is automatically pulled, Excel reports are generated, 5 key suggestions are issued, and emails are sent to operations.
• Before a new e-commerce product is put on the shelves, it undergoes a round of multi-agent review (product selection, naming, copywriting, pricing strategy). If one item fails to pass, the product will be stuck.

② Engineering Team · Make the LLM Pipeline an Operable and Maintainable Asset

• Code review: Multiple agents independently review a PR from different perspectives (readability/security/performance/test coverage), and give scores and improvement suggestions after summarizing.
• Online alarm classification: Connect to Webhook, LLM determines whether the alarm is a "real problem" or "noise", and automatically creates a ticket or swallows it directly.
• Release Note Automation: CI triggers, captures commits, filter breaking changes, and organizes them into user-facing Markdown.

③ Agent Developer · Orchestrate Multi-Agent Collaboration

• group_chat: Let the three role agents of PM/Engineering/QA criticize each other until a plan is aligned.
• agent_based: Let the orchestrator dynamically assign subtasks to the respective workers of researchers/writers/coders after reading the user's questions.
• state_machine: Multiple rounds of dialogue + state transfer, writing a complex business process (contract approval/work order life cycle) into a replayable state chart.

Three Key Differences from Other Tools

1. Interactive Breakpoint Debugger

Set breakpoints before or after any step, on failure, or when a condition is met; step through execution, inspect every variable, and edit variables at runtime before continuing. The experience matches a debugger in an IDE.

2. the Service Restarts and Automatically Resumes from the Interrupted step.

The workflow reached step 7, and the server happened to be restarted by OOM - other tools had to be run from the beginning. After Braidrun configures the recovery field, it can continue running directly from step 7 after restarting, and the previously completed steps (marked with idempotent) will be automatically skipped.

3. YAML and canvas bidirectional editing

Canvas is suitable for adjusting topology and step relationships, and YAML is suitable for precise field editing. After YAML parsing is successful, it will be applied to the visual model, and changes to the canvas will also generate corresponding YAML; the exported file is suitable for entering Git, but do not rely on comments and layout to be preserved between the two views.

Who It’s Not For

Be honest: In the following scenarios, you may choose another tool that is more suitable.

• Pure ETL/data transfer — Airflow / Dagster / dbt, these professional ETL orchestrators, are still better at it. We focus on "workflow with AI steps".
• Millisecond-level real-time inference — A workflow engine inherently carries scheduling overhead, so it isn't suited to online inference with millisecond latency requirements; for those cases, calling the LLM API directly is a better fit.
• Just need chat conversation — If you just want a conversational Agent, Claude Desktop / ChatGPT are enough; Braidrun is "multi-step orchestration", and using it for a single-step task is a bit overkill.

30 Second Price Preview

• Free — 5 workflows / 3 basic step types / BYOK / unlimited AI assistant — enough to try things out on your own.
• Pro — 50 workflows / all 8 plan step types (plus the system output-read node) / schedules and webhooks / the breakpoint debugger.
• Team — 200 workflows/team shared credentials/collaborative editing.
• Enterprise — Unlimited / SSO / Audit / SLA.

For detailed package comparison and upgrade path, see Subscriptions and Pricing.

30-Second Data Security Preview

• All API Keys are stored AES-256-GCM encrypted in the credential center and never appear in logs or YAML exports.
• In a BYOK setup, model calls use your own key directly and are billed by the provider.
• The Enterprise plan supports fully self-hosted deployment — data, run records, and artifacts never leave your boundary.

What To Do Next

• Quick Start — Sign-up is free—get your first workflow running in 10 minutes.
• Core Concepts — Before you start, spend 5 minutes understanding the relationship between workflow / step / agent / module.
• Template Library Overview — First check if there are any ready-made templates that are close to your needs.