DeterminFlow is a runtime for AI workflows whose steps are already known. It organizes LLM calls, scripts, API calls, database operations, and human approvals into versioned workflows that can be validated, retried, resumed, and audited. Each agent handles one bounded node: it reads only the context that step needs, uses only the tools granted to it, and returns a result that can be checked. The runtime owns control flow, data flow, retries, and recovery.
The project positions this against handing a whole known process to a single general agent such as Codex or Claude, which carries a growing history, holds every tool, and has to be restarted by hand after a failure. In DeterminFlow a task freezes its workflow definition and inputs at start, agent nodes run in isolated sessions with their own token ledgers, invalid output is repaired or retried at that node, and execution checkpoints survive process restarts so a run continues from the failed node.
In one production novel-writing pipeline with 11 separate model sessions, a run used 176,584 tokens. The project estimates a single long-running agent would use 3.4 to 9.1 times as many tokens for the same work, a saving of about 70% to 89%.
Workflows are edited in a visual editor and served through a FastAPI backend and React console with cron automations, WebSocket events, and health checks. Plugins package workflows, agents, prompts, skills, rules, scripts, and database migrations into installable business services, and a community skill and plugin directory is built in. Desktop builds ship for Windows and Apple Silicon macOS, and you configure your own model API.
Features
- Visual workflow editor: variables, conditions, parallel branches, loops, human approval, and subprocesses
- Four core node types: agent, script, approval, and subprocess, each with its own inputs, outputs, model, and failure handling
- Frozen task definitions: a task locks its workflow and inputs when it starts
- Recovery: automatic or manual retry, skip, and resume from the failed node, with checkpoints across restarts
- Per-node tool scoping: tool allowlists and blocklists, workspace, and maximum turns set per agent node
- Structured output handling: JSON detection, parsing, repair, and model retry, with downstream nodes able to send work back upstream
- Per-call accounting: status and token usage by workflow, task, node, attempt, and model call
- Service delivery: workflows exposed as APIs, background services, cron automations, or plugins
- Reusable assets: MCP servers, agent and prompt templates, skills, and rules usable inside workflows
- Plugin system: official and community plugin sources, including an open-source novel-writing workflow set with 7 workflows and 84 nodes
- Skill marketplace: browse, install, and submit community skills with verified versions and sources
