The problem with most agent builders
They're sequential. They fail without recovery.
They require code. They route your data through
their servers.
Parlance is built to fix all four.
What it is
A free, open-source, no-code visual workspace
for orchestrating multi-agent Claude workflows.
Build on a canvas, connect nodes with edges,
run with real API calls.
What the engine actually does
Parallel execution
Modified Kahn's algorithm computes topological
waves. Each wave dispatched with Promise.all.
A→B→C runs sequentially. A and B feeding into
C runs A and B simultaneously, then C. Automatic.
Three-tier retry
Self-retry → upstream escalation → Overseer
intervention. When an agent fails, the engine
tries to resolve it without user input.
Targeted feedback
Overseer evaluates output against your conditions.
On rejection: parses TARGET: [agent name], re-runs
only that agent and its downstream forward chain.
Agents upstream of the fix point untouched. Cost
savings proportional to where in the chain the
problem is.
Workflow chaining
Chain Input nodes pull output from other Parlance
workspaces. Optional caching avoids re-running
expensive upstream workflows.
Architecture
- 1,258-line execution engine (orchestrator.js, runEngine.js, claudeClient.js, mockClient.js, engineEventBridge.js)
- React 19 + ReactFlow canvas
- Single useReducer, 26 pure-function action types
- Backend: 3 Vercel serverless functions (CORS proxy only — all orchestration runs client-side)
- BYOK — key in localStorage, sent per-request, never written server-side
56 source files. 8,310 lines. MIT licensed.
The full suite
This is the fifth tool built in two weeks.
All free, all open source, all BYOK:
- Reality Check: realitycheck-project.vercel.app
- Singularity: singularity-workspace.vercel.app
- LoopLens: looplens-project.vercel.app
- Claude Batch: virerra.github.io/Claude-Batch
- Parlance: parlance-project.vercel.app
Links
Live: parlance-project.vercel.app
Source: github.com/Virerra/parlance
License: MIT
Top comments (0)