01↗
Compose text models into multimodal capability
Use DeepSeek V4 for reasoning, GLM 5.2 for synthesis, and visual tools for perception, then publish the whole workflow as one model.
DeepSeek V4 + GLM 5.2 + vision tools → multimodal model
1flowbase
Conversation is the moat. The native foundation for AI apps.
Use Cases
AI Gateway
Virtual Model
MCP Gateway
Protocol Translation
Compose, publish, observe
Build a better virtual model from the models you already trust.
1flowbase sits between agent clients and model providers. The client still calls one familiar model name while your workflow decides what happens behind it.
02↗
Run a multi-model review panel
Fan out to several reviewers, preserve every branch, and synthesize one stronger answer behind a single model endpoint.
parallel reviewers → synthesis → final answer
03↗
Debug cost, latency, and failures
Inspect model calls, tool callbacks, tokens, duration, and errors as one connected execution trace instead of scattered logs.
request → trace → evidence → improvement
A small surface, a powerful runtime01 02 03 04
Your agents see a normal model. You see the whole system.
Publish once and keep existing client integrations unchanged. 1flowbase handles orchestration, protocol compatibility, and observability behind the endpoint.
Agent client
Claude Code, Codex, OpenCode, SDKs
Virtual model API
OpenAI- and Claude-compatible endpoints
Workflow runtime
Models, branches, tools, and synthesis
Execution evidence
Traces, tokens, latency, cost, failures
Control without lock-in
Designed for local agents and real production debugging.
1flowbase / workflow-editor
01
Protocol compatible
Use the clients and SDKs you already have. Publish workflows through familiar OpenAI and Claude APIs.
02
Provider flexible
Mix OpenAI-compatible providers, Claude-compatible models, vision models, and private endpoints in one workflow.
03
Visual workflow editor
Make branching, tools, and synthesis explicit. Reuse a workflow instead of rebuilding orchestration in every client.
04
Trace every call
Connect the final answer to every model invocation, tool callback, token count, duration, and error.
05
Self-hosted by default
Keep model credentials and execution data under your control with a one-command Docker deployment.
06
Open extension surface
Build workflow nodes, model providers, tools, and frontstage experiences on an open-source base.
Evidence, not guesswork
See why an answer was slow, expensive, or wrong.
A final answer is only the last line of a workflow. 1flowbase keeps the path that produced it visible, so you can compare branches, inspect tool results, and improve the system with evidence.
- ✓Model and tool call timeline
- ✓Token, latency, and failure attribution
- ✓Inputs and outputs connected to workflow nodes
execution / trace-7b14● complete
Run it on your own machine
From zero to a workflow endpoint with one command.
Deploy the full stack with Docker, open the visual editor, and publish your first virtual model.
Linux / macOSbash
curl -fsSL https://raw.githubusercontent.com/taichuy/1flowbase/main/scripts/shell/docker-deploy.sh | shFrequently asked
What developers usually want to know.
Is 1flowbase another LLM proxy?+
No. A proxy mainly selects or forwards to a model. 1flowbase composes models and tools into a workflow, publishes that workflow as a virtual model, and preserves the complete execution trace.
Do I need to change Claude Code or Codex?+
Usually no. If a client supports a custom OpenAI- or Claude-compatible endpoint, it can call a published 1flowbase workflow through the same model API shape it already understands.
Can I use local or private model providers?+
Yes. 1flowbase is designed for self-hosted deployments and provider-compatible endpoints, so credentials and traffic can remain in infrastructure you control.
How is this different from coding an agent graph?+
Agent frameworks help you write orchestration in code. 1flowbase gives that orchestration a visual editor, a reusable runtime, protocol publishing, and connected observability for local agent clients.
1f
Make every conversation compound into your AI application moat.
Start with the multimodal composition, adapt the workflow, and publish it as your own model endpoint.