Live · in daily use
Maestro
Self-hosted multi-model research orchestrator — an open alternative to closed deep-research products.
- Role
- Sole author
- Stack
- LangGraph · OpenRouter · sandboxed workers
- Code
- Repository →
≥2
independent sources required per claim
1 : n
one conductor, a model bound per node
Live
self-hosted, domain-agnostic engine
How it works
- 01 Conductor goal → task DAG
- 02 Model binding a model per node by task type
- 03 Sandboxed fan-out sub-agents run in parallel
- 04 Disk handoff path + 2-line abstract returned
- 05 Corroborated synthesis ≥2 sources per claim
The conductor never sees raw sub-agent output — only the index it wrote.
Model binding per node
Task type decides the model, so cost tracks difficulty instead of the longest step.
The problem
Hosted deep-research tools are a black box on two axes: you cannot see the plan they followed, and you cannot choose the model that runs each step. Everything gets the same expensive general-purpose model, and the trace of what was actually read is discarded when the answer appears.
The approach
A conductor model decomposes the goal into a task DAG, then binds a model per node by task type — cheap and fast for retrieval, stronger for synthesis. Sub-agents run sandboxed and in parallel, and nothing enters the final answer without corroboration from at least two independent sources.
What broke
Context. Sub-agents returned full findings to the conductor and the orchestrator window filled long before the DAG finished. The fix inverted the contract: sub-agents write bulk output to disk and return only a path plus a two-line abstract, so the conductor reasons over an index rather than the corpus. Planning stays lean no matter how much gets read.
What worked
The same DAG machinery drives literature review, competitive research and technical due diligence with no task-specific code. Because the plan and the reading trail are both on disk, a run can be audited after the fact — which is the part hosted tools cannot offer.