PaloSpring
PaloSpring is a self-hosted agent runtime built on structured continuity. Development runs as a sequence of gated stages, each consuming the artifact produced by the stage before it, so intent and rationale are preserved rather than discarded when the context window scrolls. Connect Palo Bloom to carry the shape of a session across the boundaries that would otherwise erase it.
Open source. Self hosted. Structured autonomy.
Local control first
Runs on your hardware, on your model keys, against your repositories. Each stage gates before the next begins, and at round close the Council reviews. Nothing leaves your machine unless you route it out.
Type "palospring".
Open a terminal, choose a workspace, write the directive. The pipeline stages activate in sequence, each gated on the artifact produced by the one before it.
Research before implementation.
Prior artifacts in the workspace are reused when they exist. When they do not, pre-research runs against the evidence base before implementation is permitted to begin. The system will not start coding just because it was asked to.
Architecture as artifact.
The research brief becomes a blueprint. Preflight checks verify consistency, critique examines the architecture on its own terms, and only then does implementation begin. Each stage receives the prior artifact rather than a conversation log, because context that must be held in attention across a hundred turns is context that has already been lost.
Implementation and validation.
Code is generated from the architecture artifact, merged through an integrator, and run through the validation ladder: lint, test, review. QA passes or the round does not close.
The Council
Deliberation is structurally separate from execution. When a stage produces ambiguity or a judgment call that exceeds model authority, the issue routes to the Council, where votes are recorded and overrides are written to an audit log. Disputes are resolved by structured decision rather than by averaging opinions or letting the most confident model win.
Stage routing
Each stage can route to a different model. Local models for sensitive work where data must not leave the machine, remote providers where throughput or capability matters more than locality. A single model can carry the full pipeline if the developer prefers it that way.
Open Dataset
Opt in research data
Anonymised task and output pairs, off by default. Contributors opt in, and the resulting dataset is public. It is intended to improve understanding of structured agentic development, not to train models that replace the developers who contributed it.
Effort Mode
Low to Extra High. Controls how many research loops run, how many critique passes each artifact receives, and whether the Council reviews at round close or only on dispute. More effort costs more time and compute, and catches more errors before they compound.
Start here
Evaluate it yourself.
Open source, self hosted, documented, available. Run it on your own hardware. Add Palo Bloom when continuity across sessions matters.