Bay Run
io.github.barneywohl/bay-run · v1.2.0 · MCP 2025-11-25
Free OpenAI-compatible inference with signed provenance receipts and 3 focused MCP tools.
Reachability
reachable
checked 2026-08-27 05:43 UTC
Registry status active
Tools pinned 32
f8e9e149e7ea
Tools last changed
2026-08-13
Repo last push 2026-08-18
GitHub stars 0
Provenance
Registry namespace io.github.barneywohl
(GitHub-account verified by the official registry)
Repository
github.com/barneywohl/bay-run
Website
https://run.huggingbay.xyz
Remote endpoints
https://bay-run-mvp-889989800693.us-central1.run.app/mcp/ (streamable-http)
https://bay-run-mvp-zfmlsu2yla-uc.a.run.app/mcp/ (streamable-http)
https://run.huggingbay.xyz/mcp/ (streamable-http)
Observed changes
| When (UTC) | Event | Detail |
|---|---|---|
| 2026-08-27 05:36 | tools changed | 0 added, 0 removed, 3 changed |
| 2026-08-26 05:33 | tools changed | 2 added, 4 removed, 1 changed |
| 2026-08-20 05:31 | tools changed | 0 added, 0 removed, 2 changed |
| 2026-08-19 05:30 | tools changed | 0 added, 0 removed, 2 changed |
| 2026-08-18 05:30 | first capture | 5 tools pinned |
| 2026-08-13 05:31 | tools changed | 2 added, 0 removed, 15 changed |
| 2026-08-13 05:31 | tools changed | 2 added, 0 removed, 15 changed |
| 2026-08-12 05:31 | tools changed | 9 added, 0 removed, 15 changed |
| 2026-08-12 05:31 | tools changed | 9 added, 0 removed, 15 changed |
| 2026-08-11 05:30 | tools changed | 1 added, 0 removed, 1 changed |
| 2026-08-11 05:30 | tools changed | 1 added, 0 removed, 1 changed |
| 2026-08-10 05:30 | tools changed | 5 added, 0 removed, 1 changed |
| 2026-08-10 05:30 | first capture | 20 tools pinned |
| 2026-08-09 05:30 | first capture | 15 tools pinned |
Pinned tool definitions (32)
| Tool | Description |
|---|---|
| try_bay_run | Run Bay Run's zero-credential activation proof. This tool accepts no arguments and performs no caller-controlled compute: it always evaluates the fixed expression 6 * 7, returns 42, and links to OAuth, MCP, A2A, document/PDF... |
| discover_models | Search the catalog for a small, cheap, open specialist model that does ONE narrow job better/cheaper than a general LLM — embeddings, reranking, text classification, NER/extraction, routing, guardrails, transcription, vision,... |
| eval_models | Prove which candidate model actually wins on YOUR data before committing — a head-to-head bake-off, not a public leaderboard (MTEB rank does NOT predict your-domain performance). `task` = 'embedding'|'rerank' (rank a labeled... |
| embed | Turn text into embedding vectors using ANY open embedding model. Configured common specialists are pinned warm; long-tail models load on demand. No caller packaging or deployment is required. `model` = any Hugging Face embedding... |
| rerank | Reorder candidate documents by true relevance to a query using an open scalar-output cross-encoder/reranker, served on demand — the standard move to sharpen RAG / search precision after a vector search returns a noisy top-k.... |
| extract | Turn messy HTML/text (e.g. a scraper/Firecrawl dump) into STRUCTURED JSON using a small CPU-served generative specialist — the reliable "extraction" layer downstream of scrapers, no GPU, cheap at volume. `model` must be a... |
| find_specialist_for_task | ONE call to find the best small specialist model for your task, proven on YOUR examples. `task` = plain-language description (e.g. "multilingual sentence embeddings for FAQ search", or "extract product fields from scraped... |
| route | RUNTIME auto-router: send a job WITHOUT knowing which specialist to use, and Bay Run picks the curated warm small specialist for the inferred task family when available. Unlike find_specialist_for_task (which needs LABELED... |
| classify | Classify text with a small CPU-served text-classification specialist — the guardrail / safety / moderation / sentiment / intent / NLI layer agents need and that frontier routers don't offer as tiny models. TWO modes: (1)... |
| speed_test | Run Bay Run's fixed, bounded FAST-PATH benchmark. It executes embedding, reranking, and classification through the curated small-model warm pool and returns per-stage model, latency, cold-load status, plus total server wall... |
| request_specialist | Ask Bay Run for a specialist for a task — and NEVER get a dead end. If a servable specialist EXISTS, this chains discover -> (eval, if you pass labeled `examples`) -> and returns a ready-to-call serve pointer (the proven winner,... |
| memory_context | FAST durable working CONTEXT in one call. With no arguments, load the newest values from your default private context. Pass `updates` to atomically save decisions, plans, tool results, or handoff state and receive the refreshed... |
| remember | DURABLE cross-call MEMORY: store a small key->value so future runs of YOUR agent keep context between calls (agents' #1 repeated pain — no built-in memory between tool calls). Upserts (namespace, key) -> value, scoped to YOUR... |
| recall | Read back DURABLE MEMORY you stored with `remember`. Pass `key` to get one value, or omit `key` to LIST live keys in bounded, newest-first pages. Scoped to YOUR principal (bearer-token identity) — you can only ever see your own... |
| search_memory | Search YOUR durable memory without crossing principal boundaries. Matching is a deterministic, case-insensitive LITERAL substring over keys and stored JSON; %, _, and ! are ordinary characters, not wildcards. Results are... |
| export_memory | Export a versioned, bounded JSON page of YOUR live durable memory. Omit namespace to walk every namespace for the authenticated principal. Each item carries a stable entry receipt and content digest; next_cursor continues the... |
| set_memory_ttl | Set, refresh, or clear expiry on one live memory entry while preserving its stable entry receipt. `ttl=null` clears expiry. Missing or already-expired entries return found=false and are never resurrected; call remember to start... |
| forget | Delete DURABLE MEMORY. Pass `key` to remove one entry, or omit `key` to clear the ENTIRE namespace for YOUR principal. Scoped to your bearer-token identity, so it can never touch another agent's memory. Returns {deleted: N}. |
| calculate | Evaluate arithmetic/math EXACTLY and instantly — the reliable calculator agents need because LLMs silently botch multi-step arithmetic. SAFE: no code execution, a hardened whitelist parser (not eval). Supports + - * / // % **,... |
| validate_json | Validate JSON instantly: if `data` is a STRING, confirm it PARSES; if a `schema` (JSON Schema) is given, validate the value against it and return every violation. The layer agents need to check tool arguments / LLM-generated... |
| resolve_link | Check whether a URL is ALIVE (fast HEAD/GET); if it's dead/unreachable AND it names a model Hugging Bay has MIRRORED, return the mirrored copy's serve pointer — a fallback UNIQUE to Bay Run (we hold the mirror). Great for... |
| verify_provenance | Authenticate a Bay Run result provenance receipt and optionally match its output, source artifacts, document chunks, and citations. Verification is deterministic, bounded, and side-effect-free. It returns only status booleans... |
| parse_pdf | Extract embedded text from a PDF with native PDFium, FAST and model-free. Pass a public URL or small base64 PDF; every fetch, redirect, byte, page, character, timeout, and concurrency dimension is bounded. Returns page-aligned... |
| parse_document | Parse DOCX, HTML, Markdown, CSV, or TSV into deterministic RAG-ready text. The source is MIME- and size-gated; DOCX ZIP expansion/XML, HTML complexity, CSV rows, characters, parse time, and concurrency are independently bounded.... |
| summarize | EXTRACTIVE summarization, FAST and CPU-cheap — pull the most central (and, if you pass a `query`, most relevant) sentences straight out of a document, in their original order. No generative model, no hallucination: every... |
| rag_search | Turnkey RETRIEVE for RAG, FAST — hand it a query and your documents and get back the most relevant passages, ranked, with scores + original indices. The one-call retrieval layer for agent RAG: embeds the query and passages with... |
| submit_job | Submit a DURABLE bounded batch for asynchronous execution. The same principal + idempotency_key + body returns the same job while its bounded record is retained; reusing the retained key with a different body returns... |
| get_job | Get principal-scoped job STATUS without returning stored inputs or results. States are queued, running, completed, failed, or cancelled; completed_items reports bounded progress. |
| cancel_job | Cancel a job ONLY while it is still queued. Claim and cancellation are atomic, so once a worker has changed the state to running this returns job_not_cancellable and never interrupts in-flight model work. Repeating cancellation... |
| get_job_result | Retrieve the bounded result of a principal-scoped terminal job. Each batch item has an independent succeeded/failed outcome with a machine-readable error. While queued/running, returns job_result_not_ready; results follow the... |
| retry_job_callback | Retry a principal-owned blocked or exhausted callback receipt. If signing or the terminal snapshot is still unavailable, abandon_job_callback releases its retained capacity. |
| abandon_job_callback | Abandon a principal-owned blocked or exhausted callback receipt when delivery is unsafe or no longer useful. Abandoning never deletes through the runtime role and releases callback capacity immediately. |
Get alerted when this changes.
Email the moment io.github.barneywohl/bay-run drifts, dies, or revives.
Status badge
[](https://toolpin.dev/servers/io.github.barneywohl/bay-run)
Maintain this server? Add the badge to your README. It links your users to this live status page.