Your documents were written to be read. Shaper rebuilds them to be retrieved—so agents find the right answer, in the right shape, every time.
Human prose is built for sequence. Agent retrieval is built for self-contained answers. The gap is where good agents go wrong.
Every document passes through the same rigorous system. Each issue is visible, explainable, and paired with a precise recommendation.
Confirm the source preserves headings, lists, tables, and links. Markdown or semantic HTML is preferred; Word must use native styles and converted PDFs need manual verification.
Agent impact
Weak formats force the parser to infer structure, producing unreliable chunks.
Recommended action
Convert to clean Markdown or semantic HTML and verify the extracted structure.
Find skipped heading levels and headings used only for visual emphasis.
Agent impact
Poor hierarchy creates arbitrary chunk boundaries.
Recommended action
Normalize heading levels and use them to express real section structure.
Find list-like text typed into paragraphs instead of native numbered or bulleted lists.
Agent impact
Items can merge or lose their sequence during extraction.
Recommended action
Convert list-like paragraphs into semantic lists.
Find aligned text masquerading as a table or tables without header rows.
Agent impact
Cell relationships can disappear when the content is extracted.
Recommended action
Use real tables with clear headers and captions.
Identify large sections that combine unrelated topics.
Agent impact
One retrieved chunk may mix irrelevant or conflicting intents.
Recommended action
Split long sections into focused, single-topic subsections.
Verify every internal reference points to an existing section containing the expected information.
Agent impact
A broken reference leaves the retrieved passage incomplete.
Recommended action
Repair the target or inline the required context.
Identify referenced documents, forms, appendices, or systems absent from the knowledge base.
Agent impact
The agent cannot follow a dependency it cannot retrieve.
Recommended action
Include the source under a stable identifier or inline the required content.
Find sections that point to one another without either defining the rule.
Agent impact
Retrieval reaches a loop instead of an answer.
Recommended action
Replace the loop with one explicit, canonical rule.
Find references to missing, unavailable, or retired attachments.
Agent impact
The agent may fill the missing evidence with unsupported content.
Recommended action
Restore the attachment or remove and replace the reference.
Find terms such as approximately, typically, generally, or roughly.
Agent impact
The agent must repeat the ambiguity or invent precision.
Recommended action
Use concrete values, explicit ranges, or a stated exception.
Find decisions left to an unnamed person or case-by-case process.
Agent impact
The source provides no defensible decision path.
Recommended action
Name the decision owner and state the criteria they apply.
Find specialist or recurring terms treated as already understood.
Agent impact
Retrieved passages can be interpreted differently across queries.
Recommended action
Define each term where it first appears or in a linked glossary.
Find different names that may refer to the same concept.
Agent impact
Search and retrieval signals fragment across multiple labels.
Recommended action
Choose one canonical term and map necessary aliases explicitly.
Find conditional statements that never resolve to a concrete rule or outcome.
Agent impact
The agent may supply a condition or result not supported by the source.
Recommended action
State the trigger, exception, and resulting action explicitly.
Compare repeated numbers, dates, durations, and thresholds for the same rule.
Agent impact
Different retrieved chunks can produce different confident answers.
Recommended action
Reconcile values and keep one canonical statement.
Find competing precedence or authority statements.
Agent impact
The agent cannot determine which instruction governs.
Recommended action
State a clear source hierarchy and conflict rule.
Find current and obsolete guidance presented together without status markers.
Agent impact
Old guidance can rank as highly as the current rule.
Recommended action
Remove obsolete content or label its status and replacement clearly.
Compare terms used across sections and authors for the same subject.
Agent impact
Drift weakens retrieval and can imply distinctions that do not exist.
Recommended action
Standardize terminology throughout the source.
Find empty or absent appendices, tables, definitions, and enumerated items.
Agent impact
The evidence cannot support a complete answer.
Recommended action
Supply the missing material or state the exclusion explicitly.
Find TBD, not included, to be completed, and similar placeholders.
Agent impact
The agent may infer details that were never approved.
Recommended action
Resolve placeholders before publication or mark the topic unavailable.
Find rules that vary by jurisdiction, product, or team without the required breakdown.
Agent impact
A general chunk can be applied to the wrong audience.
Recommended action
Provide the complete breakdown and label each scope.
Find pilots or programs without current status, start date, or end date.
Agent impact
Temporary guidance can appear permanently valid.
Recommended action
Add status and dates, then remove expired guidance.
Confirm the document exposes a clear effective date and version.
Agent impact
The agent cannot independently determine recency.
Recommended action
Record both values in metadata or the document header.
Confirm one identifiable source of truth exists for each topic.
Agent impact
Multiple sources can drift and return incompatible answers.
Recommended action
Designate and maintain one canonical source per topic.
Find facts repeated across handbooks, wikis, slides, emails, and other sources.
Agent impact
Copies drift independently and compete during retrieval.
Recommended action
Consolidate the fact and link other sources to it.
Find references to the latest communication or verbal clarification without a source.
Agent impact
The governing evidence is unavailable to the agent and reviewer.
Recommended action
Replace informal references with a dated, retrievable source.
Check that overlapping sources state which one governs in a conflict.
Agent impact
The agent may select authority based only on retrieval rank.
Recommended action
Declare precedence wherever multiple sources legitimately overlap.
Find sections dependent on earlier context, generic headings, subtle duplication, and tables or lists that lose meaning alone.
Agent impact
A retrieved chunk can be relevant yet impossible to interpret correctly by itself.
Recommended action
Use descriptive headings, add local context and captions, and consolidate repeated facts.
Find text trapped in images, text boxes, SmartArt, embedded objects, fields, or annotations, plus broken characters, numbering, and reading order.
Agent impact
Invisible, reordered, or garbled content cannot reliably ground an answer.
Recommended action
Transcribe non-text content, run the intended parser, and compare its output with the rendered source.
The same fact, reshaped so an agent can retrieve and reason from it.
“Customers who have made a purchase may, subject to the conditions below, request a return...”
“You can return an unused item within 30 days of delivery.”
Answer firstHuman-written content
Bring the policies, guides, FAQs, and help content your agent already relies on.
29 deterministic checks
Shaper identifies structural and retrieval issues, then explains every recommended change.
Quality scorecards
Every shaped document gets a graded evaluation showing what passed, what changed, and why it matters.
Answer-shaped knowledge
Get newly shaped documents with explicit answers and semantic structure, ready for your agent.
Plug Shaper into any MCP-compatible client and let agents shape knowledge on demand — run checks, reshape documents, and receive structured output without a custom integration.
Any MCP client can ask Shaper to run the full 29-check assessment and receive structured, explainable findings.
Get back clean HTML, evaluations, and readiness scores as machine-readable JSON your agent can act on.
No bespoke API work. Add Shaper as an MCP server and every compatible client gains knowledge-shaping capabilities.
Any MCP client can call Shaper’s tools. The agent drives the workflow; Shaper supplies deterministic checks, exact proposals, and verified output.
Add Shaper as an MCP server in any compatible client. Supply your OAuth bearer token and the Streamable HTTP endpoint.
Create an estate and register sources. Files upload through the authenticated REST endpoint; everything else runs through MCP.
Call estate.discovery.start to run the 31 deterministic checks, then read structured findings, evidence, and agent impact.
Generate proposals, record decisions, transform approved documents, and retrieve exact hash-verified HTML.
// Start deterministic discovery
{
"name": "estate.discovery.start",
"arguments": {
"estate_id": "estate-ea66e438..."
}
}
// Read structured findings, evidence, and impact
{
"name": "estate.discovery.reports",
"arguments": {
"estate_id": "estate-ea66e438...",
"run_id": "<discovery-run-id>"
}
}{
"servers": {
"shaper": {
"type": "http",
"url": "https://<shaper-host>/mcp/",
"headers": {
"Authorization": "${input:shaper-authorization}"
}
}
}
}Replace the URL with your Shaper host. Supply a short-lived OAuth bearer token when the client prompts.
estate.createcompileCreate a distinct Knowledge Estate.
estate.discovery.startcompileRun deterministic discovery on staged sources.
estate.discovery.reportsqueryRead findings, severity, evidence, and remedies.
estate.recommendation.startcompileGenerate exact transformation proposals.
estate.proposal.decidereviewApprove or decline a proposal with rationale.
estate.transformation.startcompileTransform one approved proposal into shaped HTML.
estate.artifact.approvereviewApprove an artifact with exact finding acknowledgment.
knowledge.queryquerySearch the active approved evidence release.
Roles use collection-scoped grants such as shaper:<collection>:query, :compile, and :review. Estate admin actions remain REST or browser operations.
This page reads directly from the shaper repository — documentation and the latest changes are fetched live from GitHub.
Give chat agents answer-first content with explicit context, so every response is grounded instead of guessed.
Turn buried help articles into precise question-answer pairs that surface the exact detail a visitor needs.
Arm support bots with structured policies and troubleshooting steps that resolve tickets faster.
Make internal wikis, SOPs, and documentation retrievable so employees find answers without interrupting colleagues.
Shape your knowledge before it reaches the conversation.
Request early access