Orca Prompt Vault
Browse versioned system prompts, tool definitions, and harness records from AI products.
- Source repo
- Continuum-AI-Corp/OrcaPromptVault
- Stars
- ★ 72
- Last updated
- 11d ago
- License
- AGPL-3.0
- FA score
- 51/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificChatGPT · Codex · Claude Code · Claude.ai (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Researchers compare the instructions and tool sets a single model receives under different harnesses such as OpenCode and Hermes.
- Not a fit if
- Teams needing a prompt API or runtime integration
- Researchers who require captured evidence for every artifact
- Source review
- 51/100 · Major gaps
What does this agent do, and when should you use it?
Orca Prompt Vault is an archive of system prompts, developer instructions, tool schemas, and harnesses organized by the product that ships them. It distinguishes material captured from local proxy traffic with OrcaReplay from reported text recovered from model responses, and describes the different evidence limits of each. The README lists 106 artifacts: 44 captured files and 75 inherited from CL4R1T4S. Coverage includes products such as ChatGPT, Codex, Claude, Claude Code, Gemini, Cursor, and OpenCode, with folders organized by prompt publisher rather than underlying model vendor. It is suited to inspection, research, and version comparison; the repository does not document a query API or hosted service for applications to consume.
Readers can browse product folders for prompts, tool schemas, skill definitions, and harness instructions, and consult docs/CAPTURES.md for captured artifact sizes, harness versions, and reproduction commands. The capture workflow uses OrcaReplay: install orcareplay with npm, clone the OrcaReplay repository, then run capture/capture.mjs to start a local proxy and launch an unmodified agent; the tool extracts the prompt, replaces personal paths and identity details, and writes a file. Reported artifacts are archived as received and dated when the date is known. Readers can rerun a capture command and compare the output, or contribute a reproducible capture or dated transcript following CONTRIBUTING.md.
- Researchers compare the instructions and tool sets a single model receives under different harnesses such as OpenCode and Hermes.
- AI product teams track prompt changes over time and identify the lines that changed.
- Developers inspect how Claude Code prompts differ between interactive terminal and scripted modes.
- Tool authors study existing tool schemas, skill definitions, and harness instructions for interoperability research.
- Contributors capture agent requests on their own machines with OrcaReplay and submit artifacts with reproduction commands.
How do you install or deploy this agent?
No installation is needed to browse the files on GitHub. To capture prompts yourself, the README provides the steps below; you need npm, a Node.js runtime, Git, and the target agent available locally. Captures read traffic from your own sessions through a local proxy. The README does not specify costs or account requirements.
How do you use this agent?
This example follows the README to capture Claude Opus 5. After it runs, inspect the generated prompt file and use docs/CAPTURES.md and CONTRIBUTING.md to check provenance details and contribution format.
What are this agent's strengths and limitations?
- Includes both captured and reported material, with an explicit explanation of their different evidence limits.
- Captured entries include harness versions, file sizes, and reproduction commands for repeat collection and comparison.
- Covers products from multiple vendors and independent projects, organized by prompt publisher.
- Preserves CL4R1T4S history and indexes inherited files by their upstream paths.
- Of the 106 artifacts listed in the README, 44 are local captures, and some products have no captured entries.
- Prompts assembled server-side by services such as ChatGPT, Claude, and Gemini can only be archived as reported text, which has a different evidence basis.
- Users must run OrcaReplay and the target agent locally; no hosted capture service or prompt API is documented.
- The README says some harnesses require --tls-intercept but does not provide full per-product operating details here.
How does this agent compare with similar options?
The README describes the archive's relationship to CL4R1T4S: Orca Prompt Vault preserves its unedited files and commit history while adding OrcaReplay captures. It describes OrcaReplay as the companion tool for capturing and replaying requests byte-for-byte.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Orca Prompt Vault This agent | 51 · Major gaps | Web appFree | ★ 72 | 11d ago | — | — |
| Unstract: Unstructured Document to Structured Data | 44 · Major gaps | Self-hosted serviceFreemium | ★ 7.3k | today | Python | OpenAI API · Claude API |
| Agent Lightning | 40 · Major gaps | CLIFree + model costs | ★ 19k | 3d ago | Python | OpenAI API |
| Agents 2.0 Symbolic Learning | 25 · Major gaps | Library / SDKFree + model costs | ★ 6k | 2y ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes a local proxy reading traffic from the publisher's own sessions, scrubbing it before storage, excluding credentials and personal data, and allowing takedown requests. This gives some visibility into data handling and removal. Deductions: proxy permissions and user confirmation steps are unspecified; dependency security is not addressed; the stated external effects lack implementation evidence; source indexing and capture details are described but the referenced indexes and captures were not supplied for review.
The README distinguishes Captured from Reported artifacts and explains their different evidentiary status. Deduction: artifact counts conflict: the README says 119 artifacts, while Credits says 106. Availability safeguards for OrcaReplay, npm, and external services are not described, and the supplied material says nothing about error handling or failure messages.
The archive covers products from multiple vendors and distinguishes use contexts such as terminal and script modes; it also states that the archive is for understanding agents rather than bypassing their safeguards. Deductions: reader prerequisites and unsupported environments are not specified; the broad product coverage is backed here by only brief capture instructions and references to external documentation.
The repository describes a product-based layout, file naming format, contribution process, upstream index, and capture index, providing a strong information structure. Deductions: installation guidance is limited to a short global install and capture command; examples and FAQ material are sparse; the AGPL file and archive licensing statement are clear, but collecting third-party prompts under fair use leaves rights and license scope uncertain; maintenance is partly attributable to the named team and contribution path, without a clear responsible maintainer or update commitment.
The archive provides prompts, tool schemas, and harness material, with concrete comparisons that illustrate research uses. Deductions: as a static archive it does not itself provide agent capabilities; reproducing or validating its value depends on OrcaReplay and archive contents not supplied in this review.
The README separates captured material from reported material, describes capture metadata and scrubbing, and claims to provide reproducible commands and a path-by-path upstream index. It also clearly distinguishes evidence from testimony. Deduction: only the README and LICENSE were supplied, so the referenced indexes, captures, commands, and findings could not be cross-checked.
- CAPTURES.md, UPSTREAM.md, and the archived artifacts were not supplied, so capture provenance, reproduction commands, and upstream mappings could not be checked.
- The README's artifact totals conflict, and its fair-use statement does not resolve uncertainty about rights or license scope for third-party prompts.