Apastra
What it is
Apastra is a self-hosted, file-based PromptOps framework and skill pack for managing prompt versions and evaluating agent instructions. It stores prompt specifications as YAML files with stable IDs, variable schemas and output contracts that can be versioned in Git. For regression checks, Apastra compares evaluation results with baselines; it also validates evaluation evidence and recalculates scorecard means. Users supply the executable harness that makes model calls, because the project does not yet include a production provider adapter. Apastra can run without a hosted service, telemetry sink, hidden database or CI infrastructure. Optional GitHub Actions workflows support regression gating, promotion and immutable releases, and the documentation names Claude, Cursor, Amp and Codex as compatible IDE agents. Installation is documented through Git clone or npm, with setup actions controlled by opt-in flags. Optional request logging is off by default and stores complete request bodies locally, without authentication headers or uploads. Its rolling retention policy is seven days or 250 MB and can be changed. Apastra is free and open source under Apache-2.0; model and API use may incur costs during evaluations.
For and against
For
- Version prompt specifications in Git.
- Compare evaluation results with baselines.
- Can run without hosted or CI infrastructure.
- Optional GitHub Actions workflows support regression gating.
- Request logging is off by default.
Against
- No production provider adapter is included.
- Users must supply an executable model-call harness.
- Model and API usage may incur costs.
The verdict
Apastra provides local-first prompt versioning and evaluation workflows, with optional GitHub Actions integration. Its main prerequisite is a user-supplied harness for running model calls.
Who it suits
It suits developers, platform teams and agents that need versioned prompts and evaluation checks. Users must bring their own executable harness for model calls.
Plans and prices
Apastra
Apache-2.0 · local-first framework · model/API usage may incur costs when evaluations run
Read 4 Oct 2026 · github.com
On the record
- Agent compatibility
- The docs name Claude, Cursor, Amp, and Codex as IDE agents that can run Apastra workflows.bintzgavin-apastra-14.mintlify.app · read 4 Oct 2026
- Evaluation approach
- Apastra validates evaluation evidence and recomputes scorecard means, while a user-controlled adapter executes the model calls.github.com · read 4 Oct 2026
- Install options
- The README documents installation by Git clone or npm, with setup actions controlled by opt-in flags.github.com · read 4 Oct 2026
- Integrations
- The docs describe optional GitHub Actions workflows for regression gating, promotion, and immutable releases.bintzgavin-apastra-14.mintlify.app · read 4 Oct 2026
- License
- The project is licensed under Apache-2.0.github.com · read 4 Oct 2026
- Local-first design
- Apastra says it can run without a hosted service, telemetry sink, hidden database, or CI infrastructure.github.com · read 4 Oct 2026
- Logging retention
- The request-logging setup offers a rolling retention policy of seven days or 250 MB, which users can change.github.com · read 4 Oct 2026
- Model adapter limit
- The README says Apastra does not yet ship a production provider adapter, so users bring their own executable harness.github.com · read 4 Oct 2026
- Prompt versioning
- Prompt specs are YAML files with stable IDs, variable schemas, and output contracts that can be versioned in Git.bintzgavin-apastra-14.mintlify.app · read 4 Oct 2026
- Regression checks
- Apastra compares evaluation results with baselines to help catch quality regressions before shipping.bintzgavin-apastra-14.mintlify.app · read 4 Oct 2026
- Request logging
- Optional provider request logging stores complete request bodies locally; it is disabled by default and does not persist authentication headers or upload logs.github.com · read 4 Oct 2026
- What it does
- Apastra is a file-based PromptOps framework and skill pack for versioning prompts and evaluating agent instructions.github.com · read 4 Oct 2026
- Who it is for
- The getting-started guide identifies developers, platform teams, and agents as its audience.github.com · read 4 Oct 2026
