Reproducible examples and evidence for Snipara Project Intelligence.
This repository keeps public integration examples, MCP discovery metadata, and benchmark evidence in one place. It is intentionally small: no customer data, private runbooks, internal ranking code, or credentials are included.
server.jsondescribes the Hosted MCP server for the official MCP Registry.io.github.Snipara/sniparais the published official MCP Registry entry (1.0.0).examples/hosted-mcpshows the primary HTTP setup path without hard-coding a project or API key.benchmarks/hosted-context-2026-06contains the frozen June 19, 2026 context-selection evidence pack and its committedresults.json.proof/project-continuity-2026-07mirrors the measured summary from the public Project Intelligence proof replay, with its owned-repository and negative-control limits kept explicit.
Verify the benchmark evidence with only Python's standard library:
python3 benchmarks/hosted-context-2026-06/verify_results.pyThe verifier checks the case manifest hash, case coverage, reported means, deltas, and token-reduction arithmetic. It does not claim that a live LLM run will be byte-for-byte deterministic.
Snipara is a Hosted MCP server for project intelligence and context optimization across AI agents. Each user connects a project-specific endpoint:
https://api.snipara.com/mcp/YOUR_PROJECT_SLUG
Authentication is supplied at runtime through X-API-Key or a bearer token.
Never commit an API key to this repository.
The frozen evidence pack compares focused Snipara context against the first 32,000 tokens of the same indexed documentation corpus on 12 project-context questions. It preserves case-level scores so the published aggregates can be independently recomputed.
It is one hosted run per task, not a universal model-quality guarantee. Model behavior and the live hosted index can change, so future reruns should publish a new dated evidence pack instead of rewriting the historical result.
The Project Intelligence proof replay measures a different question: whether continuity context changes what a coding agent finds before implementation and whether continuity-heavy scenarios pass across local, Codex CLI, and Claude runtimes. It is kept separate from the hosted-context benchmark because it uses a Snipara-owned repository and does not yet include a negative control where cold agents should already succeed.
snipara-evalsprovides the deterministic Project Intelligence scoring library and CLI.snipara-companionprovides local workflow continuity and code-impact gates.snipara-memoryprovides the standalone open-source memory engine.- Snipara documentation documents the Hosted MCP, Project Intelligence, integrations, and security boundaries.
Apache-2.0. See LICENSE.