The MCP server is for operations: answering questions about how your workspace runs and
carrying out operational commands. Asking Hive, reading your Ask chats, and running or testing
agents are not available to AI clients — use Hive directly for those. See What stays in the
Hive app.
This page is about AI clients connecting to Hive. For the opposite direction — Hive connecting
to another tool’s MCP server — see Connect MCP servers.
Your server URL
If your workspace’s Settings → Connected AI clients shows the MCP server URL, the server is available to you — copy the URL from there. It is:Connect your client
- Claude Code
- Claude.ai and Claude Desktop
- GitHub Copilot in VS Code
- Gemini
- Cursor and others
/mcp in Claude Code and choose hive to sign in.find_tools and run_tool — which is what lets a
client with a tool limit work at all. Add ?toolset=full to the URL if your client can hold
several hundred tools and you want each listed separately. See Toolsets.
Clients register themselves automatically, so you never create a client ID or secret by hand.
Clients from known MCP client platforms — Claude (claude.ai, anthropic.com), GitHub
Copilot (github.com), VS Code (vscode.dev), Cursor (cursor.com), ChatGPT (chatgpt.com,
openai.com) and Windsurf (windsurf.com) — may identify themselves with a client ID metadata
document published on their own domain. Every other client uses OAuth dynamic client
registration, which MCP clients do on their own when they connect.
If you build an MCP client, register it dynamically. Hive only reads a client ID metadata
document whose client ID is an https:// URL on one of the platform domains above or a
subdomain of one, on the default port, with a path — for example
https://claude.ai/oauth/mcp-oauth-client-metadata. A client ID URL on any other domain, an IP
address or another port is refused as an unknown client, and Hive makes no request to it.
Sign-in requests that name a client ID metadata document Hive has not seen before are limited to
10 every 15 minutes per network (an IPv6 /48 or an IPv4 /24).
A dynamically registered client cannot send members back to a shared host where anyone can
publish a page under a path of their own: Google Apps Script (script.google.com,
script.googleusercontent.com), cloud storage (Google Cloud Storage, Firebase Storage, Amazon
S3, Azure Blob Storage) and raw GitHub content (raw.githubusercontent.com,
gist.githubusercontent.com). Registering such a redirect_uri is refused with
invalid_redirect_uri. A per-user subdomain such as <user>.github.io is allowed.
Sign in and consent
1
Your client opens Hive
The client sends you to a Hive consent page in your browser. Sign in if you are not already
signed in.
2
Check who is asking
The page shows the client’s name, where you will be sent back to, the workspace it will act in,
and the access it asks for. A client that registered itself automatically is labelled
unverified: its name is whatever it chose to call itself, so judge it by where you will be
sent back. For such a client the page shows the path of that address under its host, shortened
in the middle when it is long. If that is not a place you expect, deny the request. Hive removes
invisible and text-direction characters from client names and shortens long ones, so a name
cannot hide or rewrite the return address.
3
Choose the workspace
The connection acts in one workspace: the one you are working in. If you belong to more than
one workspace, pick the one you want from the Workspace list on the page. Choosing another
workspace also switches your active workspace in Hive, and Allow waits until the switch is
done. If you switch workspace in another tab while the page is open, the page asks you to reload it before
you decide.
4
Choose the access
The page lists the access the client asked for, area by area — see Access by
area. You can narrow it to any part of that request, for example read-only, or
a single area. You can never grant more than the client asked for. A client that asks for
nothing specific gets read-only access to all areas.
5
Return to your client
Approve, and Hive sends you back to the client, which can now call Hive’s tools. Deny also sends
you back, telling the client you refused. For your security, the Allow button only becomes
available once the page has been open and in focus for a moment without taps, clicks or key
presses anywhere on it, and it pauses again whenever you switch away. Once it is ready, move
your pointer onto it (or move keyboard focus to it) before you select it. On a touchscreen or
with a pen, press and hold Allow briefly, then lift; a quick tap does not approve.
https://oauth.get-hive.ai. Your client never sees your Hive password or session.
Access by area
Access is granted per product area. Read lets the client see what you can already see in that area; Write lets it take the actions you are already allowed to take there, and includes Read for the same area. Your role still applies on top: access you grant never lets a client do more than you can do in the app. A client also sees only the signals, decisions and matters you hold yourself, as you do in the app, whatever your role. See Who can see a signal.
Two more options cover every area at once, including areas Hive adds later:
- All areas (read) —
hive:read. - All areas (write) —
hive:write, which includeshive:read.
approvals:write as well as the signals or
diligence scope.
Agents and workflows carry out the actions selected for them as each account’s level allows (by
themselves on Act), apart from the few
that always wait for a member’s click and any action whose destination did not come from words
you trust: those wait as an approval card (see
What agents and workflows do by themselves).
Saving, scheduling or deploying an agent that has selected actions needs only the
Agents scope (plus the read scopes its sources and attached apps already need), and running or
arming a workflow with an active Action step needs only the Workflows scope. Neither needs
approvals:write.
An AI client cannot add where an agent or a workflow sends. Creating, updating, restoring or
duplicating an agent, or changing its bindings, is refused with destination_in_hive (403) when it
would add an email address, a phone number, a link, a channel or a repository the agent does not
already have. So is creating, updating, duplicating or importing a workflow, saving its draft,
applying a workflow chat change or asking Describe for a new workflow, when it adds one the
workflow does not have. The error reads “Add where an agent or a workflow sends in Hive itself: an
AI client cannot add an address, number, link, channel or repository to it.” Removing or keeping
the ones already there works, and so does adding your own connected mailbox’s address to an
agent; deploying is unaffected. Add other new ones in the Hive app. The same refusal covers a
workflow send that goes to the answer of an Agent step pinned to an app with no tools (or of such a
row in a Parallel agents step), since that answer can decide where it sends: an AI client cannot
create or import such a workflow, point a step at such an answer, move a send to another row or
output field of the same step, or change that Agent step’s prompt, apps, tools or model.
An AI client also cannot approve a card Hive held because of where its action sends:
approvals_resolve answers approve_in_hive (403), “Approve this in Hive itself: an AI client
cannot approve an action Hive held.” It can still deny or snooze one. What an AI client writes is
never your words: a schedule it saves does not vouch for where an agent sends, and the form values
of a workflow run it starts do not count as yours, so a step that sends to one waits for approval.
See AI clients and drafts.
Ask projects is marked sensitive on the consent screen, because a project’s files can hold
confidential material.
An agent answers from the knowledge it is given, so over MCP that knowledge needs its own area too.
Creating, editing, duplicating, restoring, deploying or scheduling (including changing or
re-enabling a schedule) an agent
whose attached knowledge, file inputs or source slots reach the Library needs library:read; one
bound to connected apps, connected accounts, connected documents, a knowledge workspace or a knowledge document (both can be files synced from connected apps) needs connectors:read, and one that delegates to other agents needs what they need. The
same applies to running, scheduling or publishing a workflow with a saved-agent step that uses
one. An agent with no Library or connected-app knowledge needs only agents:write. Runs that
start on their own — a schedule or a trigger — are not affected.
Reading an agent’s answer over MCP needs the same areas as running it, because the answer quotes
the knowledge it drew on. This covers a run’s result, the answer and steps on a run, run history,
evaluation runs, and a workflow run with a saved-agent step:
- With
agents:readalone (orworkflows:readfor a workflow run), you can read the answers of an agent that has no Library or connected-app knowledge. - An answer that draws on the Library, or cites a Library file, also needs
library:read. - An answer that draws on connected apps also needs
connectors:read. That includes one that cites a connected document, a record from a connected app or a file synced from one (such as a OneDrive or Google Drive file), and one from an agent with a bound knowledge workspace or knowledge document. - An agent that delegates to other agents needs what each of them needs, as they were when the
run started. If a delegate’s version can no longer be read (it was deleted), only the
hive:readorhive:writeumbrella scope reads the answer. - A run recorded before Hive pinned agent versions needs both
library:readandconnectors:read.
insufficient_scope rather than returning it with
parts left out. A list of runs or run history still lists every run, but a run your access does
not cover shows its status, timing and who ran it without its answer or steps. Starting or
stopping a run still works, and its response leaves out an answer your access does not cover. An
agent template’s report in your notifications, or in a delivery approval, shows “Agent report
ready — open Hive to read it.” without connectors:read. A workflow Approval step’s
notification, and a workflow action step’s approval, can quote an earlier saved-agent step, and
Hive does not work out which agent at that point: without the hive:read or hive:write umbrella
scope they show “Workflow content may include an agent answer — open Hive to read it.” The
weekly brief lists those approvals the same way, as “Workflow action” or “Deliver the agent
report” rather than their step label. Approving still sends exactly what the workflow prepared. Reading answers in the Hive app is
unchanged.
A workflow reaches whatever its steps do, so over MCP its steps need their own areas as well as
workflows:write. An Agent step (or an agent in a Parallel agents step) that reads connected
apps, and a trigger that starts on a connected app’s events, need connectors:read. Create
PDF and Generate image steps file what they make into the Library, so they need
library:write. Hive checks the whole workflow when you run or test it, deploy it or switch it
to an earlier release, submit its form, and create or change a schedule for it. A step that is
switched off is not counted. Runs that start on their own are not affected. The deploy
preview shows its warning about a Create PDF step’s folder only when the connection’s scopes
allow Library reads (library:read, library:write, hive:read or hive:write) and the member
it acts for can use the Library.
A test run started with workflows_run checks each send’s recipients as a production run does.
When a production run would refuse a send, for example to a Slack channel the connection cannot
see, the tool returns an error instead of the run: action_not_allowed (403) or
workflow_connector_unavailable (409) when the Slack connection is the problem. The error carries
the run’s id and the failed step’s message, so you can read the run with workflows_list_runs. See
Test runs check recipients.
Company Overview and Diligence sit under Diligence & company overviews, but a project on a
connected OneDrive, SharePoint or Excel file reads that drive with your connected account:
listing your connected drives, folders and workbooks, creating such a project, and syncing,
running, reading or acting on it each need connectors:read as well. Without it, listing projects
leaves out those projects. A project on an uploaded data room needs only the diligence scope.
Listing your connected MCP servers over MCP shows each server’s address only (for example
https://mcp.example.com/…), never its full URL, because some services put a connection secret in
the URL path. The Hive app still shows the full URL.
Developers configuring a client request these as OAuth scopes, space separated — for example
signals:read workflows:write. find_tools and your client’s tool list only show the tools your
connection can use.
What your client can do
Hive’s operational actions are available as MCP tools — Signals, Inbox and approvals, Business Brain, Workflows, setting up agents in Agent Studio, Library, Ask projects, connectors, Company Overview, Diligence, settings and the rest of the product. When Hive adds an operational action, it appears as a tool automatically. Tool names start with the area they act on, such assignals_list and signals_snooze.
Your client reaches them with find_tools and run_tool, or lists them individually with
?toolset=full — see Toolsets. Either way it sees only what your connection was
granted.
Ask and agent runs stay in the Hive app
Hive’s MCP server answers questions about operations and carries out operational commands. It does not give AI clients Hive’s Ask or agent pipeline, whatever access you grant (hive:write
included). These are not tools, and a client that calls one by name is refused:
- Asking Hive — a question, a follow-up, editing and resending a question, regenerating an answer, and AI edits to an Ask document.
- Reading Ask chats — your conversations and chats shared with you, their messages, comments and files, chat search, and the documents Ask wrote in them.
- Running or testing agents — running an agent, the agent runs you start from an agent’s page or from Ask, Agent Studio previews and tests, evaluation runs and delegations. Public agent links are never tools either.
- Reading agent answers — run results, run lists and run history.
signals_remind parks one of your signals until a time you choose and reminds you, and only you, in your
Inbox then (POST /api/signals/:id/remind, body { "remindAt": "<ISO date-time>" }). The time must be between 5
minutes and 30 days ahead; anything else is refused, not moved. See
Remind me.
matters_get returns one of your own matters. It includes actions: for each signal on the
matter that has a receipt, its signalId, when the occurrence started (startedAt), each action’s ladder (its
label, the rungs Hive has evidence for, the current state, a detail, and whether it is reversible or was
undone) and the outcome Hive read. Provider ids are never included. The conversation view of the review panel
is not available over MCP.
brain_get_context briefs your client on the customers, people or deals a question names: the
facts Hive learned about each with their source and date, their relationships, learned normals,
what changed in the last 30 days and how they connect. It adds their open signals only when your
connection also has signals:read.
No Brain tool returns the people, addresses, company links and latest threads Hive takes from the
From, To and Cc lines of your Gmail or Outlook mail (see
People and companies from your mail),
and Brain writes by id treat those entries as not found. Like the mail itself, they stay in the
Hive app and Ask. People Signals names on your matters are a separate part of the Brain and are
not covered by this.
Brain fact writes with brain:write (adding or correcting a fact) take a fact name Hive knows, such as email,
phone, domain, role or company, or a note named note:<your label> in lower case, such as
note:payment terms. Use a note for anything no known name fits. A write under any other name is refused and
answers 422 with one of these codes:
Writes you make through your own connection count as yours, like an edit in the Hive app.
Asking the workflow assistant —
workflows_generate, workflows_send_chat,
workflow_chat_threads_start and workflow_chat_messages_send — answers at once with a pending
request and its id, while the assistant works in the background. Call
workflows_get_assistant_turn with that id until its status is succeeded (its result is the
draft or the reply) or failed (its error says why). Only the member who asked can read it.
Ask Hive, in the Hive app, uses a small, fixed set of Hive’s own read routes itself, read-only,
to answer questions about how your workspace runs, such as workflow and agent runs and schedules.
That is unchanged: only AI clients lose Ask and agent runs. They run with
the asking member’s own role and never write. See
Ask about your Hive workspace.
Some things stay in the Hive app and are never tools, whatever access you grant:
- asking Hive, reading Ask chats, and running or testing agents or reading their answers (see Ask and agent runs stay in the Hive app);
- choosing, changing or cancelling a plan, buying seats, credit top-ups and add-ons, assigning seats, changing spend controls, auto top-up and member limits, accepting or declining an offer from Hive, approving or declining a workflow run paused at 50 agent steps (on its Inbox card), downloading the usage CSV, and your workspace’s model provider keys (reading your plan, usage and limits are tools: see Limits and credits);
- creating incoming webhooks, because the new signing secret would pass through the AI model;
- inviting members, changing their roles, and removing them;
- Hive staff platform tools;
- managing AI client connections and consent;
- connecting accounts to Hive (OAuth sign-in with your tools);
- changing your personal settings, such as your custom instructions for Ask, your default
model, your language and time zone, and the GitHub pull-request summary Hive posts to Slack, because they shape every later
Ask answer in the app or send your private pull requests to a channel (your client can still
read them with
account:read); - setting or changing an Ask project’s instructions, and adding, changing or removing its skills or skill packs, because they shape the Ask answers of everyone in the project (your client can still create and rename projects, and read their instructions and skills);
- changing a Library template’s generation instructions, because they shape every document generated from it by anyone it is shared with (your client can still set them when it uploads a new private template, read them, and generate a file from a template, which uses model credit);
- sharing an Ask project with someone and giving a project member more access, because that hands them your chats and files (your client can still lower a member’s access, remove members, and leave a project someone shared with you);
- archiving an Ask project, because it hides the project and its chats from everyone in it (your client can still restore an archived project you own);
- sharing an agent or a workflow with more people, adding a member or raising a member’s role, because that hands them its results and its reach into your connected apps and Library (your client can still see who has access and remove people);
- adding an address, number, link, channel or repository to an agent or a workflow, and approving an action Hive held because of where it sends, because an AI client’s words must not decide where your agents and workflows send (your client can still remove ones already there, and deny or snooze a held card);
- sharing anything else with more people, whatever access you grant (
members:writeandhive:writeincluded): granting someone access to a chat, project, prompt, Library folder, file or template, raising their access, transferring a Library folder, or making a folder or template available to the whole workspace (your client can still lower someone’s access and remove people); - putting content where other people will see it: adding files to a project other people can see; uploading, creating or moving files, folders or prompts into a Library or prompt folder someone else can read; running, publishing or scheduling a workflow whose Create PDF step files into a shared Library folder, or whose Create PDF or Generate image step saves to the Library of a workflow owner other than you; saving a workflow as a workspace template; and publishing diligence knowledge to the workspace (your client can still do all of these in places only you can see — workflow PDFs and images can be filed into your own private folders — and move content out of a shared place);
- turning on chat sharing or data-sharing consent for the workspace, or widening diligence knowledge reuse (your client can still turn them off or narrow them);
- changing the workspace’s Ask instructions or company description, and creating, enabling, disabling or deleting workspace skill packs, because they shape every member’s Ask answers;
- creating, editing, deleting, sharing, mandating or switching skills,
approving or rejecting suggested skills, uploading a skill and Create with chat, because a
skill steers your own and, once shared, everyone’s Ask answers (your client can still read
skills with
library:read); and changing what the workspace shows every member as its own — the sidebar links, the new-chat info boxes, the chat disclaimer under the Ask composer, the sharing notice and custom onboarding tasks — because every member trusts them as the workspace’s own words and links (your client can still read all of them withsettings:read, delete onboarding tasks, and change other general settings withsettings:write); - engaging or releasing the kill switch, because
releasing it lets Hive act on its own again, and engaging it halts every automation in the
workspace and returns active learned behaviours to shadow, which releasing it does not undo
(your client can still read whether it is on with
autonomy:read); - public agent links and public workflow form links;
- live event streams and test utilities.
Toolsets
Hive has several hundred tools — one for every action in the product. Most AI clients cannot hold
that many at once, and some stop part-way through loading them and end up with nothing. So the
default gives your client two tools and lets it reach everything through them:
find_tools— describe what you want in plain language, such as “snooze a signal”. Hive ranks its tools by meaning and by keyword and returns the best matches with their input schemas. If meaning-based search is briefly unavailable, the answer says it used keywords only, so describe the task in English keywords.run_tool— run one of those tools by name, with its arguments. It runs reads and changes alike, within the access you approved on the sign-in page.
run_tool is one tool, your client cannot tell you which Hive action it is about to take,
so Hive marks it as making changes and a well-behaved client asks you before every call.
Add ?toolset=full to the server URL for a client that can hold hundreds of tools. Your client
then lists each action as its own tool, can ask you to approve each one separately, and can upload
files — the one thing run_tool cannot do, because it carries JSON rather than file contents.
find_tools marks those upload tools so your client knows to ask you to switch.
Switching an existing client between the two URLs connects it again, which leaves you with two
connections in Settings → Connected AI clients. Revoke the old one.
notifications/cancelled message is refused as an invalid request (-32600). A request that
has not been answered within two minutes and fifteen seconds is ended with HTTP 504 and a
JSON-RPC timeout error (-32001). If that call
changes data, Hive reports its outcome as unknown rather than failed, because the action
may still finish. Your AI client is told to check the current state, for example by listing
or reading the item, before it tries again, so the action is not done twice.
Limits and credits
Your AI client works inside your workspace’s plan and credits, like you do in the app. The client pays for its own model; Hive charges only for the work it does for you.Because the compact toolset is the default, your client starts most tasks with a
find_tools
search. Ranking tools by meaning uses a model, so it is charged as a knowledge search — see the
plan details below. Searches that fall back to keywords use nothing.- Database-only read and data tools — reading signals, the Brain, workflows, the Library and most other tools that only read or change Hive’s own records — are free up to 20,000 calls per Pro or Max seat a month, pooled across the workspace. After that they use 1 credit per 250 calls. Approving, rejecting and undoing use no model credits; like other database-only tools, they count toward the free 20,000 calls per Pro or Max seat a month, and after that use 1 credit per 250 calls.
- Reads that reach outside Hive — a live read of a connected app, a sync, a knowledge search,
or an upload Hive reads and indexes — use credits from the first call. (Hive reads Companies
House only while an agent is working, for example in a workflow’s Agent step: that work uses
credits as it does in the app, and each Companies House read it makes for your client uses a
small fraction of a credit.) Reads like these are measured on their summed cost: they add up,
and a tenth of a credit is drawn each time their cost reaches one.
A
find_toolssearch that ranks tools by meaning is a knowledge search too, measured the same way and charged to the member who connected the client. If the pool is empty, the search still answers, by keyword only. - Tools that run AI work — running a workflow, drafting with the workflow assistant or drafting an agent with AI, research — use credits exactly as they do in the app.
find_tools keeps working: when it cannot rank by meaning it ranks by keyword instead,
so your client can still find and run the tools it is allowed to.
See While the workspace is read-only.
Limits:
- AI client work can use at most £5 of AI credits per member a day (500 credits on an earlier plan, inside the member’s own monthly limit and the workspace’s pool).
- A workspace’s connections together make at most 600 calls a minute, on top of each connection’s own limit (see Security). Reads that reach outside Hive are limited to 30 a minute and 3,000 a day per workspace.
- Some outside services are shared by every Hive workspace, so each workspace’s AI clients may use at most a tenth of such a service’s limit in a minute. Past it, a read is refused until the next minute, and other workspaces are not affected. Companies House, which Hive’s agents read for you, allows each workspace at most 120 reads in any 5 minutes (Hive’s whole key allows 600), and AI clients at most 12 a minute of those.
- A member can hold at most 20 active connections.
- A tool for a feature your plan does not include is refused as it is in the app, with
feature_not_in_plan: on Starter, for example, adding a delegation to an agent or reading an agent’s analytics. - Connecting an AI client needs an active Basic, Pro or Max seat. A member without one, or whose seat is
suspended over the plan’s limit, is told on the consent page that they need a seat instead of being
offered Allow, and every tool call from a client they connected earlier is refused with
seat_required(“You need a seat to use this workspace. An owner or admin can give you one in Settings → Members.”) until they have a seat again. The client can still list Hive’s tools, which shows no workspace data. - A member on an Approver seat cannot connect AI clients either: the consent page offers them no Allow, every tool call from a client is refused, and the clients a member connected are disconnected when they move to an Approver seat.
billing:read, your client can read:
Anything that spends money or changes what the workspace pays or may use stays in the Hive app,
most of it under Settings → Billing. An Owner or Admin approves or declines a workflow run
paused before it takes more than 50 agent steps on its Inbox card (or, without a seat, under
Billing notices in Settings → Billing), and a public link’s Model for visitors is
chosen on the agent’s public link. Asking for a higher top-up limit, and approving or declining
that request, are done in the app too. See Billing and plans.
Security
- Short-lived access. Access tokens last 10 minutes and only work against the Hive MCP server. Your client refreshes them in the background.
- Rotating refresh tokens. Every refresh issues a new refresh token. If an old one is used again, Hive treats it as stolen and revokes the whole connection, however many refreshes ago it was replaced and whatever scope it asks for. Sending an old refresh token to the revocation endpoint ends the connection too. A client that retries a refresh within a few seconds, because it never received the answer, is refused without losing its connection. A token Hive did not issue is simply refused; it cannot end anyone’s connection.
- Connections expire. A connection ends after 14 days without use, and after 90 days at most however often it is used. An expired connection leaves the list. Sign in again to reconnect.
- Permissions are checked on every call. Your role and workspace permissions apply to each tool call, not just at sign-in, so a role change applies from your client’s next call and keeps the client connected. If you are removed from the workspace, every AI client you connected to it is disconnected straight away and leaves Connected AI clients; reconnecting needs a new sign-in once you are a member again. The same happens if the workspace is deleted: your client’s next refresh is refused and it asks you to sign in again.
- Malformed sign-in requests stop at Hive. If a client starts sign-in with a request Hive cannot accept, such as an unknown scope or a missing PKCE challenge, Hive shows the error itself instead of sending you back to the client’s address. Update or reconfigure the client and connect again.
- No staff elevation. Hive staff platform access never applies over MCP.
- Limits apply. Every MCP request a connection sends counts toward its per-connection rate
limit, including listing tools, initializing and
ping, not only tool calls. Each HTTP request counts once as soon as it arrives, and that count covers its first message; each further message in a batch counts once more. An HTTP request that arrives when the connection is already over its limit is answered429 Too Many Requestswith aRetry-Afterheader and a JSON-RPC error, without being read; a batched message over the limit gets its own JSON-RPC error telling the client how many seconds to wait. One HTTP request can carry a batch of at most 10 messages; a larger or empty batch is refused whole, without running any of it. A message that names no method, such as an empty object or a response, is refused as an invalid request (-32600). A request body can be at most 1 MB, except — on?toolset=full, the only toolset that offers upload tools — a single tool call (not a batch) to an upload tool, such as uploading a Library file or an Ask project file, which can be up to about 14.3 MB — enough for the largest upload, 10 MiB of files as base64 — so a larger request is refused by its declared size with413 Payload Too Large, before it is read. Past its first 1 MB, an upload request must be file content (base64), not more JSON structure, or it is refused with413too. A member can send one upload request larger than 1 MB at a time, a workspace one at a time across all of its members, and Hive processes only a few at once overall; another is answered429 Too Many Requestswith aRetry-Afterheader once its first 1 MB has arrived, so send large uploads one after another. Past its first 1 MB, an upload must keep arriving at 2 MB or more every 5 seconds and finish within 45 seconds, or it is refused with408 Request Timeout. After a workspace’s upload larger than 1 MB ends in anything but a stored upload — refused with408or413, or refused by the upload tool, for example because its files are over 10 MiB — that workspace can send no upload larger than 1 MB for 60 seconds. A workspace can also send at most 6 uploads larger than 1 MB every 10 minutes. Either way, another is answered429 Too Many Requestswith aRetry-Afterheader. A request body must arrive in full within 60 seconds, or it is refused with408 Request Timeout. A connection can have at most 4 requests in progress at once, and a member at most 8 across all of their connections; a request beyond that is answered429 Too Many RequestswithRetry-After: 1, without being read. A request with no access token at all always gets its401challenge, so clients can discover how to sign in however many connections they start from one address. Requests with an access token Hive cannot verify (malformed, or one Hive did not issue) are limited per network address and are answered with429 Too Many Requestsonce over that limit. Requests with a real Hive credential are limited per connection instead, never per network address, so hosted clients that send many members’ traffic from shared servers, such as Claude.ai and ChatGPT, do not throttle each other: an expired or revoked access token always gets its401challenge so the client can refresh, and sign-in code exchanges, token refreshes and revocations are not held back by other members’ traffic from the same address. Client registration is limited only against floods, at 3,000 registrations an hour per network address, so a hosted client registering once per member is not refused. Each connection can refresh its tokens at most 30 times an hour, andfind_toolssearches are limited to 30 a minute per member across all of your connections. A refresh over the hourly limit is refused with HTTP 503temporarily_unavailableand aRetry-Afterheader; the client keeps its refresh token and can use it once the limit allows. Only successful refreshes count toward the limit, and a refresh token sent by a different client is refused without counting. Your plan’s limits and credits apply as they do in the app: see Limits and credits.
Manage connected clients
Settings → Connected AI clients lists each client you have connected, with:- the client’s name and the host it returns to,
- the areas it can read or change,
- when it was connected, last used, and when it expires: the end of its 14-day idle window, or its 90-day limit if that comes first.
Related
Connect MCP servers
The other direction: Hive using another tool’s MCP server.
Roles and permissions
What each role can see and do — and so what a connected client can.
Security overview
How Hive isolates workspaces and protects access.
Developer overview
Every way to build on Hive.