Overview
Coventra MCP connects Claude Desktop, Codex, or any MCP-compatible client directly to your review project. The agent reads the project context you explicitly share — screening criteria, study records, extraction fields, quality templates, and review status — and works from it. Not a copy of the data. The actual project state.
The design is simple: the AI proposes, the researcher decides. Every output lands in a reviewable queue. Nothing becomes part of your canonical record until a human explicitly accepts it.
This is not a chatbot overlay. It is a controlled connector with scoped permissions, per-project consent, and a redacted audit trail on agent activity.
- Compatible clients: Claude Desktop, Codex, and any MCP-compatible AI agent.
- Access model: per-project, per-client consent granted from Project Settings, revocable at any time.
- Visibility: the AI can only see projects you have explicitly connected. Reviewer identities during blind review are never exposed.
- Audit: agent activity is logged with operational metadata while source text, private prompts, credentials, and private evidence are not written into public-facing logs.
The AI as a research member
With MCP connected, the AI is not something you query in a separate window — it has direct access to your project and can work through it with you. Open Claude Desktop or Codex, connect to the project, and talk to it: ask what the screening queue looks like, tell it to work through a batch of records, ask why a particular study was flagged as borderline. It answers using your actual project data.
In practice: a researcher describes what they need, the AI does the mechanical work, and the researcher reviews the output. Locating records, navigating PDFs, transferring values — the AI handles those steps. Judgment stays with the researcher.
The conversation carries context across the session. The AI tracks which studies it has already screened, which proposals are pending, and what the current project state is. You can follow up, redirect, or ask for a summary of what has been done.
- You can ask direct questions about your project: 'How many studies are still unscreened?' or 'Which included studies have incomplete extraction?' and the AI reads the live project state to answer.
- You can give multi-step instructions: 'Screen all unscreened records for RCTs, tag the likely eligible ones as RCT candidates, and flag borderline ones for my review.'
- The AI retains context within a session. If you correct its reasoning on one study, that correction informs how it handles subsequent ones.
Screening
Coventra MCP reads each study's title, abstract, and your project's inclusion and exclusion criteria, then returns a structured recommendation — Include, Exclude, or Maybe — with a one-sentence reason citing the specific criterion that determined the result. For a project with several hundred unscreened records, the AI can process the full queue in a single session within minutes.
The AI's decisions occupy a dedicated agent lane tracked separately from human reviewer decisions. Agent assessments are invisible to co-reviewers, excluded from PRISMA counts, and never counted in conflict detection or agreement statistics. Blind review is unaffected. A small indicator appears on each study the AI has assessed; a reviewer can reveal the recommendation at any time by clicking it, but it is never shown by default.
Beyond pass/fail screening, the AI can create specialised tags and filters on your behalf. Tell it to tag all single-arm studies, all studies published before a given year, or all papers with a specific comparator — it screens the relevant records, creates a labelled group, and applies a custom filter you can activate in the screening workspace to view exactly that subset. Tags generated by the AI are visually distinct from researcher-created labels and can be bulk-removed in one click.
Screening conflicts are also addressable through conversation. The lead reviewer opens Claude Desktop or Codex, describes the conflict to the AI, and works through the reasoning with it directly. Once they reach agreement, the AI submits the resolution through Coventra MCP as a pending proposal. The lead reviewer then accepts it in the Coventra conflict queue. The reasoning and the final call both originate from the conversation — not from a button in the app.
- Focus your own review time on the Maybe category, which typically accounts for ten to twenty percent of records.
- For reviews with precise eligibility criteria, the AI can reliably classify the large majority of records. For exploratory reviews with broad criteria, treat its output as a triage aid.
- Specialised tags and filters created by the AI carry an AI indicator and a timestamp noting when they were generated.
Data extraction
Data extraction is the most labour-intensive phase of most systematic reviews. Coventra MCP helps an agent work from the uploaded full text and propose values for your extraction sheet: sample sizes, means, standard deviations, event counts, participant totals, timepoints, effect sizes, and free-text fields including population description, intervention, comparator, and study design.
Every proposed value must be source-backed. If a value cannot be tied to a visible location in the current PDF, it is rejected before it reaches the review queue. The agent does not get to estimate, invent, or silently fill gaps — absent fields stay absent until a researcher confirms otherwise.
Proposed values appear as pending overlays in the same column layout as your canonical extraction grid. Acceptance is field-level: you can accept the sample size, manually correct the mean, and leave the standard deviation for your own verification in a single pass through the row. Accepted values retain source provenance so they can be audited later.
The same reviewer-assist pattern applies to baseline characteristics and study characteristics. The agent proposes; the researcher verifies and accepts.
Every proposed extraction value requires human verification before it enters your canonical record. The AI cites its sources precisely so verification is fast, but it does not replace independent confirmation against the original text.
Risk of bias assessment
Coventra MCP reads a study's full text and completes an independent risk-of-bias assessment using your uploaded assessment template. For each domain question in the template, it returns a judgment with a written justification drawn directly from the PDF. The template — including all question text, domain structure, and answer options — is loaded from your project configuration.
This assessment is structurally isolated from your human reviewers. It does not appear in the conflict queue, does not count toward your required reviewer threshold, and is not visible to your team unless they explicitly open the AI assessment panel on a specific study.
A reviewer can open the AI assessment for any study and apply the entire assessment to their own record with one click, or accept individual domains selectively. Risk-of-bias domain conflicts follow the same conversational resolution model described in the conflict resolution section: the lead reviewer discusses the disagreement with the AI in Claude Desktop or Codex, reaches a reasoned conclusion, and the AI submits it as a pending resolution in Coventra for final acceptance.
- Use the AI assessment as a structured starting point and give particular attention to domains where the AI flags concern — these are statistically the domains most likely to require careful human re-reading.
- The AI's written justification for each domain cites specific passages in the PDF. If the justification references a passage you disagree with, that passage is a productive starting point for your own assessment.
Conflict resolution
Conflict resolution in Coventra MCP is conversational, not automated. There is no button in the app that triggers an AI call. Instead, the lead reviewer opens their AI client — Claude Desktop or Codex — and discusses the conflict directly with the AI. The AI reads the relevant project data: the competing decisions, the stated reasoning on both sides, the source text or extraction values in question, and the applicable criteria or template question. Reviewer identities are never exposed.
The AI reasons through the conflict with the reviewer. It can ask clarifying questions, explain which piece of evidence is more probative, identify where the criteria are ambiguous, and work toward a conclusion. This is a dialogue, not a lookup. When the reviewer and the AI agree on a resolution, the reviewer instructs the AI to submit it. The AI then writes the resolution to Coventra through the MCP connection as a pending proposal — structured, justified, and attributed to the conversation.
The lead reviewer then opens Coventra, finds the pending resolution in the conflict queue, reads the submitted reasoning, and accepts it. That single acceptance step makes the resolution canonical. The reasoning and the conclusion were formed in the conversation; the acceptance in the app is the researcher's explicit endorsement.
- Screening conflicts: the AI reads the study abstract, eligibility criteria, and both decisions, reasons with the reviewer, and submits Include, Exclude, or Maybe with the criterion that supports it.
- Extraction discrepancies: the AI reads the competing values and the PDF source text, reasons with the reviewer about which value is better supported, and submits the agreed value.
- Risk-of-bias conflicts: the AI reads the domain question, both reviewers' answers, and the PDF sections, reasons with the reviewer domain by domain, and submits the agreed judgment.
Citation discovery
Forward citation searching — identifying papers that have cited your included studies — is a methodologically required component of a systematic review search strategy but is routinely incomplete in practice.
For included studies with enough citation metadata, Coventra MCP can help surface related publications and triage them against your project's eligibility criteria. The result is a reviewer-facing list of likely relevant, uncertain, and unlikely records with brief reasons.
Likely eligible citations can be bulk-imported into your study library from the triage view, where they enter the normal screening queue. Citations already screened or dismissed are excluded from subsequent runs.
- Run citation discovery after you have finalised your included study list — forward citations of excluded studies are rarely relevant.
- The triage list is a starting point, not a final screening decision. Likely eligible citations still require independent reviewer assessment before inclusion.
Analysis profile suggestion
Selecting an appropriate analysis model — fixed versus random effects, effect measure, heterogeneity estimator, subgroup structure — requires familiarity with meta-analytic methods that varies substantially across research teams.
Coventra MCP reads your project's evidence state after extraction is substantially complete and returns a structured recommendation: analysis kind, effect measure, heterogeneity model, tau estimator, and suggested sensitivity analyses. The recommendation is accompanied by a written explanation citing the specific factors that drove it — study count, outcome type, prior heterogeneity if a run exists, risk-of-bias profile, and identified caveats.
The suggestion is a structured starting point, not an automated conclusion. It can pre-fill the analysis setup flow, but the reviewer remains responsible for the final model choice.
The suggested analysis profile is a methodological starting point informed by your evidence state at the time of the request. Clinical heterogeneity, study design nuances, and domain expertise must inform the final model choice.
How Coventra MCP reads PDF content
Coventra MCP is designed so agents do not need to browse an entire PDF blindly. The app prepares controlled evidence views from the uploaded report and keeps those views tied to the current PDF version.
When the agent works on a specific extraction question, Coventra narrows the evidence to the most relevant parts of the report and preserves source anchors for review. This keeps the conversation efficient while still letting the reviewer inspect the original source.
The practical result is precise provenance without handing the agent a raw document dump. Reviewers can trace accepted suggestions back to the supporting full text inside Coventra.
Connecting an AI client
Coventra MCP is compatible with Claude Desktop, Codex, and any client that implements the Model Context Protocol. A user connects an AI client, grants project access inside Coventra, and can revoke that access later.
On each connection, Coventra checks signed identity, project consent, and scoped access before returning project information.
- Open Settings → AI Integrations in Coventra and copy the connection details shown for your account.
- Add Coventra as a connector in your AI client.
- Restart your AI client to load the server.
- In Coventra, open Project Settings for each project you want to share and grant access to the registered client.
- In your AI client, send: list my Coventra projects — and confirm the project appears before proceeding.
Research integrity model
Coventra MCP operates under a propose-review-accept contract. The AI proposes; a human researcher decides. This applies to every capability without exception.
Agent screening decisions are tracked separately from human reviewer decisions using a dedicated source marker. They are excluded from PRISMA counts, conflict detection, agreement statistics, and blind-mode logic throughout the application. Blind review integrity does not depend on display-layer logic.
Extraction proposals are checked before storage. Proposals that cannot be traced back to the current source document are rejected before reaching the review queue.
Accepted AI-proposed values keep source provenance as a permanent part of the record. Re-uploading a PDF flags old provenance for review.
Agent activity is recorded for audit without storing private source passages, private prompts, credentials, or hidden reviewer identity details.