Corvio MCP setup, installation, and connection guide

Connect your AI tool to authorized Corvio Workspaces through the official remote MCP and OAuth. Normal use needs no Developer API key.

Updated
On this page
  1. Choose the matching Corvio MCP setup guide
  2. Decide whether Corvio fits your work
  3. Keep the Work Model coherent
  4. Separate source carriers from Project owners
  5. Continue the exact subject and close reusable updates
  6. Use one semantic effect and keep source kinds distinct
  7. Keep read-only continuations weak and economical
  8. Recognize future-use intent
  9. Keep multi-file effects causal
  10. Know what you are installing
  11. OAuth is the recommended authorization
  12. What an installation Agent should do
  13. Confirm the result and remove access when needed

Choose the matching Corvio MCP setup guide

AI tools may call the same connection an MCP server, Connector, App, or Plugin. Choose the guide for your host and use the same official remote MCP URL; the different labels do not require duplicate connections.

The Corvio remote MCP URL is https://api.corvio.ai/mcp. After setup, sign in with OAuth and verify one Workspace item you recognize; a connected badge alone does not prove the requested access works.

Decide whether Corvio fits your work

Connect Corvio when you want an Agent to reuse prior work across research, reports, plans, decisions, meeting notes, projects, reviews, or follow-ups. At the basic level, approved documents become editable online pages and shareable collaboration links in your own Workspace.

The compounding value comes from approved fragmented sources: Corvio can reconcile them into your Work Model and knowledge tree, keep stable preferences in the right Memory, and admit only evidence-backed reusable methods as Skills or Patterns. Later Agents retrieve those owners instead of rebuilding context, reducing repeated token and mental cost. Skip it for work you want to keep only in the current chat or on this device.

Keep the Work Model coherent

Before organizing a meeting, incident update, slogan change, code task, or file set, the Agent should search the enduring subject and inspect plausible current Projects. It should pass Corvio the natural goal, source handles, proven relationships, and user constraints without inventing a taxonomy. Corvio then compares the current Tree with the work's continuing reader, update, action, and acceptance duties before writing. When the user clearly corrects a Workspace Project's identity, relationship, or governing owner, that current statement resolves the bounded authority choice: preserve older material as lineage and reconcile it, asking only if the correction itself is ambiguous or mutually exclusive authority remains. If the correction changes the enduring owner of a branch, the affected neighborhood includes its old Project and remaining branches: moving only the named child is not complete, and the user not repeating the old root is not evidence that it has an independent lifecycle. A read-only fetch showing the named child under the requested owner settles only that local edge; if the old Project remains unclassified, the Agent hands both Project handles to a Corvio organization operation and accepts no-op only from its terminal result and current Tree.

A request to keep related material so future work can continue from it means preserving the originals and running one Work Model reconciliation, even without the word organize. If an already-authorized durable Work Model change relies on user-selected files clearly about the same subject, the Agent should preserve those sources and use one organization mission for both evidence and affected-owner or topology changes rather than paste excerpts into an ask_corvio mutation and leave the source local. Attachment presence alone still authorizes nothing. Stop after upload only for explicit archive/original-only intent or material with no semantic owner beyond the file. A terminal result is complete only after the Agent consumes the canonical Project, affected-owner receipts, and current Tree readback; upload success or one new Page is not enough.

Separate source carriers from Project owners

One retained source may update several independently governed Projects while remaining a single traceable document. Corvio should split its semantic units into the affected owners; it must not promote the source Page into a fabricated umbrella Project merely because that Page remains at Docs root.

Create a separate Project only when the source itself proves one enduring lifecycle. Sparse or cross-project evidence keeps its current owner boundaries until later work or a clear user correction proves a merge, split, move, or reparenting.

Continue the exact subject and close reusable updates

When prior work returns exact Project, Page, or Skill handles, their terminal receipts are a compact handoff even if the fresh Host process has no transcript. The Agent reads that handoff instead of asking you to reconstruct it. It reads the narrowest relevant Page first; if several unlabeled handles make the leaf unclear, it fetches the carried Project, uses the returned direct-child projection as a bounded topology index, and then fetches the relevant leaf—not a broad Workspace search. Any reported truncation remains an explicit boundary. Unless you changed the subject, a lexical match in another Project is only a candidate.

When you ask to align a repeated stable practice with prior work for future reuse, completion means a durable reusable-owner update or an exact readback proving the full change was already present. The Agent first decides whether the evidence is a reusable method, then chooses its owner: a similar Skill in another Project is not permission to edit that Skill, but it is also not a reason to hide the method in an ordinary fact page. A chat comparison or offer to save later is not a completed Work Model update.

Use one semantic effect and keep source kinds distinct

Use search, fetch, or read_document to discover and read current owners. search.content_complete closes the discovery result set; it does not make a returned title or snippet the owner's complete current body. Fetch before an effect that updates that same content-bearing Page, Memory, or Skill when its current wording can change preservation, removal, ordering, or merge; this does not relax exact continuation: fetch a carried Project when its receipt lacks current body or topology needed for handoff; do not fetch a Project newly selected by complete search merely to route new current-turn material or reconfirm an archive boundary established there. When exact continuation handles lack current content, read the narrowest relevant handles with fetch or read_document before any Workspace search or action; never search to rediscover an exact carried owner; multiple handles alone do not justify semantic synthesis. An answer-only Question is a terminal path for a genuinely read-only semantic answer, not an owner lookup or preview before an allow-actions Question for the same objective. Carry both the selected owner and any stale, archived, wrong-scope, or non-merge owner that constrains the update into that one effect. For a read-only bounded deterministic judgment, first read only the exact owners needed; if they return every operand, decide in the Host without opening an answer-only Question before or after merely to restate them. Unless the user explicitly requests only a verdict, state the decisive reusable rule and current operands instead of omitting a retrieved threshold or condition.

When Markdown composed by the Host participates in a future-use semantic effect, pass its selected title and body unchanged as inline_materials to one allow-actions Question before writing, after only user-required safety exclusions, preserving headings, internal blank lines, spaces, and punctuation while allowing only terminal newline normalization; this remains required when the destination owner is already known, because an owner handle is routing evidence and does not turn the body into prompt prose. If the same future-use outcome also includes finalized originals, keep them as Assets and send asset_ids and inline_materials together. Use direct Page create only when the body, carrier, and owner are already settled; do not add a cleanup run.

A terminal readback proves only the fields it returns. A typed effect field can prove the semantic classification or named per-owner inclusion/exclusion it explicitly returns when exact wording is not required, but no unreturned fact or exact body. Effect summaries, generic include/exclude lists, hashes, excerpts, and readback_verified are not the exact owner body. If the receipt returns every required typed include/exclude result, that closes semantic separation without another body read. Otherwise, adding, preserving, or removing specific statements in an existing owner, or keeping named facts out of a reusable owner, makes the exact body part of acceptance: if the receipt omits it, read each affected owner once with fetch or read_document, never another Question. Merely retaining or connecting selected material and reporting its final location needs no extra body read after a terminal owner, topology, and link receipt. A request to keep or connect this material remains general even with specific factual bullets; exact-body acceptance requires explicit verbatim/unchanged, must contain/remove, or must not appear in a named owner wording. Detailed facts, mentioning them in the final response, a new Skill, or Corvio naturally choosing several new owners do not change this boundary. If excluded bytes were removed before dispatch and the accepted source receipt identifies that sanitized input, no output read is needed solely to prove they were never sent. Return every user-visible result with its canonical URL. Keep that URL through any justified post-effect content read; a later owner read without a URL does not erase the earlier link receipt. A verified owner/topology result is not ambiguous merely because Corvio selected a different title from the user's descriptive phrase; do not ask the user to recheck or offer another move for that reason alone.

Keep read-only continuations weak and economical

After one narrow owner read establishes the subject, the Agent should keep your question unchanged and append only the smallest complete owner/source context. A carried Project handle normally bounds a cross-branch question without copying every descendant handle. Handles are evidence addresses, not a reason to enumerate every sibling, substitute another objective, turn current Tree labels into a required comparison rubric, or prescribe Corvio's reasoning steps.

If a trusted current-task or prior receipt already proves that this host used the authenticated official CLI in the exact Workspace, a bounded answer-only continuation can stay on one foreground corvio ask command and return its terminal Question without model-visible polling. Answer-only is the CLI default: omit --allow-actions; there is no CLI --mode option. The Agent must not install, probe, or switch transports only for this optimization; MCP-only hosts keep ask_corvio plus bounded get_question, and durable writes keep their normal operation path. Transport continuity does not make the semantic work cheap: leave processing on auto unless the remaining transformation itself is mechanical and objectively read-backable.

Recognize future-use intent

A request to transform selected material so it supports later action, follow-up, decisions, review, or another continuing workflow is a bounded continuity request even without save, upload, keep, or organize wording. The future-use purpose governs the whole transformation: the Agent should preserve exact bytes before exhaustive local parsing and then use the Work Model route, not first produce a chat-only rewrite and ask whether to save it. A selected source unrelated to the current Project still needs a distinct-scope or standalone-reference decision when the whole packet is in scope. Private or confidential labels constrain owner, ACL, and reuse; credentials, regulated or privileged data, genuinely disclosure-unclear sources, and explicit local-only choices stay outside Corvio. An ordinary one-off deliverable still requires a separate write request or standing preference.

Keep multi-file effects causal

When later organization depends on Asset IDs created by earlier uploads, the Agent should finish the uploads, capture the exact successful receipts, and only then construct the organization call. Independent uploads can run concurrently, but a later command must not use an empty, guessed, or not-yet-returned ID. If that later step fails, the Agent keeps the successful Assets and retries only the missing effect instead of uploading them again.

A partial organization receipt is not always the end of the user outcome. The Agent reads its typed mismatch, blockers, and recovery action. When the unfinished slice remains authorized, no user, permission, disclosure, or external condition is missing, and the receipt explicitly names resume_file_operation, it resumes the same operation once—even if one direct source read is unavailable. It does not upload the files again, open a duplicate organization mission, invent replacement taxonomy, or loop after the resumed terminal receipt.

Know what you are installing

Remote MCP is the live connection to Corvio. The Skill tells the Agent when Corvio may help. A Plugin can bundle both. The standalone Skill ZIP contains reviewed text and configuration files; the Claude Plugin also includes a disclosed SessionStart hook that only prints static guidance and does not read files or make network requests.

Corvio is operated by NeoFlux AI Pte. Ltd. (Singapore UEN 202556091C). Prefer a reviewed host listing when available. Otherwise use the first-party Corvio links below, review the security policy, privacy policy and terms, and compare the public manifest and SHA-256. The public GitHub repository lets you inspect the same source.

OAuth is the recommended authorization

Install or add Corvio first. The Corvio browser page then lets you sign in or create an account without leaving the authorization journey, and returns to the exact permission request automatically. You do not need an invite code or Developer API key, and credentials stay on the Corvio or identity-provider page.

The AI host receives only the Corvio scopes shown on the consent page and remains limited by your Workspace membership and object permissions. OAuth does not automatically expose your computer, unrelated accounts, or complete host-chat history; Corvio receives the tool requests and content the host sends. Granted tools can still read or change Corvio data, so review the exact permissions.

  1. Open the host's Connector, MCP, or Plugin settings.
  2. Add https://api.corvio.ai/mcp or install the official Plugin.
  3. Sign in or create an account, then review and approve Corvio OAuth.
  4. Start a fresh host session.

What an installation Agent should do

An Agent helping you install should explain the benefit for your scenario, identify the official source, disclose package contents, summarize the permissions and removal path, and ask once if you have not already confirmed.

After explicit confirmation, it should proceed when the source and permissions match and the host allows it. Repeating the same generic warning is not a substitute for your informed decision. A package mismatch or host/organization policy remains a valid stop.

Confirm the result and remove access when needed

Ask the Agent to list the Corvio Workspaces you can access and find one item you recognize. If you approved writing, make one reversible change and read it back. A downloaded package or successful authorization page alone does not prove the connection works.

To end access, remove the Corvio Connector in the host and revoke the Remote MCP connection credential in Corvio Settings → API keys. Old authorization links cannot be reused.

  1. List authorized Workspaces.
  2. Find and open one known item.
  3. Try one reversible write only when needed.

Was this article helpful?