Keyless evidence gathering
The core gather_bundle workflow searches and extracts public sources without making an API key the entry ticket.
OPEN-SOURCE · MCP · WEB RESEARCH
Sibyl.
Give agents evidence they can trace and revisit.
Sibyl is an evidence retrieval layer for AI agents. It returns ranked passages with provenance and explicit gaps, and supports bounded follow-up retrieval while the host agent owns planning and synthesis.
PyPI release: v0.3.0. This case study describes the v0.4.0 source preview at commit 684cdfe, reviewed October 19, 2026. The newer evidence loop and tool profiles are not claimed as part of the published package.
HOW IT WORKS
A bounded path from request to evidence.
- 01
Search
Gather public-web candidates without requiring a search API key for the core retrieval path.
- 02
Extract
Turn pages into compact, ranked passages that retain URLs, offsets, identifiers, and content hashes.
- 03
Distinguish
Cluster syndicated copies so repeated text is not mistaken for multiple independent sources.
- 04
Hand off
Return typed evidence and a next action: synthesize, refine, decompose, retry, or revise the request. Follow-up retrieval stays bounded.
DESIGN
The product is the contract.
The interface is useful because it exposes what happened, where confidence comes from, and where the system stops—not because an agent produced a confident paragraph.
Passage-level provenance
Citation IDs, character offsets, source URLs, and hashes let downstream agents inspect the exact evidence they received.
Independent-support signals
content_cluster_id groups syndicated or substantially duplicated material instead of counting domains alone.
Honest failure states
Typed status and diagnostics make missing, partial, or low-quality support visible instead of quietly filling the gap.
A bounded evidence loop
gather_evidence allows up to four retrieval calls within ten minutes. The host supplies atomic follow-up queries; Sibyl checks evidence state without invoking a hidden model.
A focused tool surface
The source preview exposes five keyless retrieval tools by default. Report and finance extras are opt-in; the full profile has fourteen tools. Missing configuration fails at startup.
QUICK START
Inspect the real thing.
Requires Python 3.10+ and Claude Code. These commands install the reviewed source preview. PyPI still serves v0.3.0; optional report synthesis requires the report extra and a configured model.
TERMINALPUBLIC BETA · SOURCE PREVIEW
$ git clone https://github.com/chriswu727/sibyl.git
$ cd sibyl
$ git checkout 684cdfe967440984b44cfc73ed3b8d16cec58938
$ python -m pip install .
$ claude mcp add sibyl -- sibyl-mcp --profile keylessVERIFIABLE SCOPE
- 85.2%
- answer coverage
- 100%
- trap safety
- 11.0s
- p95 retrieval latency
- 3
- repeats per question
Recorded July 21, 2026: 66 questions per repeat, including 54 answerable questions and 12 traps. This keyless web-retrieval run failed the overall launch gate: stability was 62.1% and the report's ready-bundle rate was 64.6%. Answer coverage measures retrieved evidence, not final-answer accuracy. Live search results change over time.
HONEST BOUNDARIES
What this does not prove.
- 01Retrieval is evidence collection, not a truth oracle; important claims still require judgment and source review.
- 02Search snippets are discovery leads, not equivalent to evidence extracted from the underlying page.
- 03Different domains can repeat the same wire copy, which is why Sibyl exposes content clusters rather than treating domain count as independence.
- 04One-shot report quality depends on the configured model, while the evidence bundle remains inspectable without synthesis.
- 05The keyless path depends on public search and publisher availability, so repeated results can vary or be throttled; the project remains a public beta until its dated live launch gate passes.
- 06Optional third-party extraction services receive the URLs they are asked to process; use the local path when that disclosure is inappropriate.
FAQ
Before you try it.
Does Sibyl require a search API key?
No for its core public-web retrieval path. Optional providers and model-backed report synthesis may have their own credentials, but structured evidence gathering is designed to work keylessly.
Does Sibyl always write a finished research report?
No. Its primary product is a typed evidence bundle. A configured model can synthesize that bundle into a cited report, but downstream agents can also inspect or compose from the evidence directly.
Do three different domains count as three independent sources?
Not necessarily. Sibyl clusters substantially duplicated or syndicated content so repeated copies can be recognized as one underlying line of support.
Does a sufficient status mean the claim is true?
No. Sufficiency describes whether the retrieval contract found enough usable evidence for the requested task. It does not replace source criticism or factual judgment.
OPEN SOURCE