HivemindOS manual
Apps And Agent Providers
HivemindOS can connect outside apps, tools, and specialist agent services while keeping Fleet, Work, Hive Superbrain, approvals, and task history in control of the overall job.
Open More → Integrations to connect an account, or More → Apps & Services to browse compatible services available across your machines. A task that needs a missing provider can open the relevant setup in place.
Browse And Connect Providers
The catalog lets you and your agents discover an existing provider before building a new integration. Entries explain the job they support, required setup by key name, whether the provider reads or changes outside data, and where the service runs.
- Agent2Agent (A2A) as the open protocol for working with agents that are not part of your fleet, in both directions.
- Browser Use for browser automation tasks such as navigation, forms, extraction, and screenshots.
- Awesome MCP Servers as a curated MCP discovery lane for tools such as GitHub, Slack, Linear, Stripe, Postgres, and Notion.
- Agent Mailboxes as the product-facing one-click mailbox flow for persistent agent email addresses.
- AgentMail as a hosted inbox API backend for one-click agent mailboxes, plus its hosted MCP server for inbox/message tools.
- Cloudflare Agentic Inbox as one provider backend for routable email-agent Workers.
- MCP Email Server as an advanced local stdio bridge for existing IMAP and optional SMTP mailboxes.
- OpenHands and Aider as optional coding runtime adapters.
- Hivemind Office as an optional local visual editor companion with HivemindOS-owned document inspection, review hashes, confirmation gates, and copy-first writes.
- Palmier Pro as an installable macOS video editor with a local MCP endpoint for timeline and media workflows.
- RentAHuman as a REST and MCP provider for real-world human work through profile search, conversations, bounties, service bookings, escrow, and transfers.
- X API MCP as an official user-account bridge for X search, timelines, bookmarks, trends, news, and Articles.
- Robinhood Trading MCP as the official OAuth bridge for Agentic brokerage accounts, portfolio and market reads, and governed equity orders.
- n8n as an installable workflow automation service.
- Listmonk as a localhost-only campaign and transactional-email manager that uses a separately configured SMTP provider.
- NocoDB as a private local database with spreadsheet-style tables and governed agent record access.
- Cal.com, local Cal.diy, and Managed Hivemind Bookings through provider-neutral, confirmation-gated agent booking tools.
- Queen Bee PRD decomposition for turning product requirements into linked Work Board tasks.
Catalog entries name required credential keys only. They do not store secret values. Installing or connecting a provider does not run it, spend money, send messages, or approve external changes.
Advanced Provider Details
The sections below describe individual protocols and specialist providers. Most users can connect them from Integrations and stop there.
Agent2Agent (A2A)
A2A is an open protocol for agents built by different people, on different stacks, to work together. Where MCP gives one model a set of tools, A2A lets two agents hand each other whole units of work. HivemindOS speaks it in both directions.
Outbound, HivemindOS can read another agent’s published capability card, send it a task, and poll that task to completion — no wrapper or custom integration per agent. Outbound calls refuse endpoints on private or loopback addresses unless you configure that peer deliberately, and require an encrypted connection for anything on the public internet, because the request carries your credential.
Inbound is off until you turn it on, and it is off by default. When enabled, HivemindOS publishes a capability card describing what the fleet can take on and accepts work from outside agents. Three properties are worth knowing:
- Inbound work is ordinary work. It lands on the Work Board, routes through the same agent selection, and has to satisfy the same completion gates, evidence requirements, and human approvals as anything you queue yourself. There is no privileged path for external callers.
- Inbound work requires your operator credential. A scoped agent token cannot accept work from outside the fleet.
- The published card describes the product, not your setup. It lists the kinds of work the fleet takes and how to authenticate — never your machines, agent names, models, or network identity. What an external caller can read back about a task is its state and its finished result, never how or where it ran.
Runtime Adapters
OpenHands and Aider appear as optional runtimes. HivemindOS checks for their CLIs, can install them through uv, and exposes a run-task runtime action for background coding work. Runtime task launches include uv/Homebrew CLI paths and load shared hive env credentials into the child process so provider keys saved with hive-env-add are available without restarting the dashboard.
- OpenHands runs documented headless automation with
openhands --headless --json --override-with-envs -t <task>and maps the selected model plus shared OpenAI key into OpenHands’ expectedLLM_MODELandLLM_API_KEYvariables. - Aider runs one-shot scripting with
aider --message <task> --yes --no-auto-commits --no-dirty-commits.
Runs are recorded under HivemindOS runtime logs and are visible through the runtime runs/log APIs. Autonomous coding runs should still use project-specific safety rules, including disposable worktrees when the task can mutate code.
Browser Use Service
The Apps & Services view includes Browser Use as an installable provider. HivemindOS installs a pinned CLI release through uv tool install browser-use[cli]==0.13.8 and checks that the installed binary still speaks the interface HivemindOS drives before using it, so a mismatched or same-named different program is reported plainly instead of failing on every action. Browser actions run against a browser HivemindOS launches and owns, never the browser you are using: it keeps its own profile, so your own windows, tabs and sessions are untouched. HivemindOS does not run Browser Use’s upstream curl | bash installer, or its setup or install commands, silently from the app.
Browser Use also offers cloud browsers — a browser running on their infrastructure instead of your machine, useful when a task needs to run somewhere isolated or several need to run at once. HivemindOS drives a cloud browser exactly like a local one, so the same approvals, limits and activity records apply. Cloud browsers cost money for as long as they run. Starting one needs a Browser Use API key and Full permissions, every one is given a time limit so a forgotten browser stops itself, and HivemindOS keeps a list of the ones it started so you can see what is still running and stop it.
/api/browser-use exposes the executable bridge for bounded local CLI actions by default: doctor, open, state, click, input, type, screenshot, and close. HivemindOS launches those commands with Browser Use anonymized telemetry disabled, only opens http/https URLs or about:blank, and stores screenshots inside the HivemindOS Browser Use run directory instead of accepting arbitrary output paths.
The Browser Use provider card has a Full permissions toggle. Enabling it requires a slide-to-unlock warning and persists server-side under the HivemindOS home directory. Full permissions let agents use the high-agency actions wired through the HivemindOS bridge, including Browser Use Cloud task creation, real Chrome profile/CDP launch options, file upload, and JavaScript eval. Setup/install commands remain controlled by the app installer rather than the runtime bridge.
MCP Catalog
GET /api/mcp/catalog returns a curated MCP server list with capability tags, credential key names, side-effect classes, install hints, and safety notes. Agents should verify credentials and side effects before installing or calling a server.
AgentMail’s hosted MCP endpoint is included for runtimes that prefer MCP. Use OAuth-capable clients against https://mcp.agentmail.to/mcp when available, or pass AGENTMAIL_API_KEY through the MCP client when OAuth is unavailable. Live send/reply tools remain write-capable email side effects and should confirm recipients, subject, body, and attachment intent before use.
Robinhood’s official Trading MCP
at https://agent.robinhood.com/mcp/trading is included as robinhood-trading. Connect it from
Integrations → MCP Servers with Robinhood browser authorization, then choose the
dedicated Agentic brokerage account. HivemindOS encrypts the local OAuth session and
reconnects it after restart. Agents can use the explicit read allowlist for accounts,
portfolio, orders, positions, market data, watchlists, options context, earnings, and
scans. Raw mutation tools are not generally exposed: equity orders call Robinhood’s
pre-trade review and then pass through HivemindOS wallet caps, task-scoped company governance when the order belongs to an active company task, explicit
CONFIRM_BUY or CONFIRM_SELL, and the unified activity ledger before placement.
Robinhood Agentic brokerage and Robinhood Chain are separate venues. The brokerage MCP operates a dedicated Robinhood account. Robinhood Chain uses a self-custody wallet, USDG, ETH gas, and on-chain Stock Token liquidity; it is not a brokerage-account bridge.
X’s official API MCP is included as xapi. The X API server is hosted at https://api.x.com/mcp, but user-account access runs through the local xurl bridge because X requires the user to provide their own OAuth 2.0 developer app. In Integrations → MCP Servers, users can save X_MCP_CLIENT_ID and X_MCP_CLIENT_SECRET into shared env, start the browser sign-in, and sync the xapi MCP entry into installed agent runtimes. Runtime configs launch HivemindOS’s local bridge wrapper and do not contain X secrets or OAuth tokens.
Users who do not want to bring their own X developer app can use the managed X API path instead. That path signs in through HivemindOS-hosted infrastructure, keeps X OAuth token custody server-side, and deducts vetted X API and X MCP calls automatically from the same HivemindOS account balance used by HivemindOS Models. Consumer requests always use that account. The app keeps the account’s server-issued credit token encrypted under the operating-system user’s home directory; operator tokens are not bundled into customer installs. The hosted /api/x/pricing policy is authoritative; the current default policy uses upstream X API unit cost plus 25% markup with a $0.001 minimum debit. A default X MCP tool call is $0.005 upstream and $0.00625 retail. The downloaded app cannot choose official X pricing, redirect revenue, or mint credits locally.
The X sign-in cache is owned by xurl under the user’s home directory. HivemindOS reports whether that cache exists, but it does not read or copy token bodies. Agents should treat X tools as acting with the signed-in account’s scopes; write-capable tools such as bookmark changes and Article publishing need explicit user confirmation.
The read-only X docs MCP is also included as x-docs for clients that can connect to Streamable HTTP servers at https://docs.x.com/mcp.
RentAHuman REST API
/api/rentahuman is the HivemindOS REST facade for RentAHuman. It reads RENTAHUMAN_API_KEY and optional RENTAHUMAN_API_URL from shared env at runtime and reports only whether those keys are present. The default upstream base URL is https://rentahuman.ai/api.
GET /api/rentahuman?action=status returns credential presence and the supported action matrix. Read-only discovery actions such as search-humans, get-human, list-bounties, get-bounty, browse-services, and get-service-availability can be called through GET or POST. Authenticated read actions such as listing conversations, rentals, transfers, escrow state, and bounty applications require RENTAHUMAN_API_KEY.
Actions that can message people, create bounties, book services, open escrow, release payments, or send transfers are prepare-first. A POST without confirmation returns the prepared request and the confirmation string. To execute, repeat the POST with confirmation: "RENTAHUMAN_ACTION" after the user has reviewed the human-facing task, payment amount, deadline, evidence requirements, and destination.
The curated MCP catalog also includes RentAHuman. Use npx -y rentahuman-mcp for runtimes that prefer MCP stdio, or the local /api/rentahuman route when HivemindOS should enforce shared-env loading and confirmation gates.
n8n Service
The Apps & Services view includes n8n as an installable provider. The install action starts n8n through Docker, binds it to 127.0.0.1:5678, sets localhost-safe HTTP settings, and keeps workflow execution outside the HivemindOS process. Once n8n is running, the existing connected-app discovery can surface its UI and API handles.
Listmonk Service
The Apps & Services view includes Listmonk for self-hosted newsletters, subscriber lists, campaigns, and transactional templates. HivemindOS creates a dedicated Docker Compose project under its local service directory, binds the admin UI to 127.0.0.1:9000, keeps PostgreSQL off host ports, preserves database and upload volumes across stops, and pins both container images by immutable digest.
Listmonk is not an inbox and does not replace AgentMail, Cloudflare Agentic Inbox, or an IMAP mailbox. It also does not deliver mail by itself. Before a live send, the operator must configure an SMTP provider inside Listmonk, verify the sending domain, review recipients and content, and configure bounce handling. Keep administrative access limited to trusted operators because campaign HTML, templates, uploads, and advanced subscriber permissions are powerful.
NocoDB Service
The Connections view can connect a NocoDB workspace on the current machine or a private fleet machine. HivemindOS keeps the connection private and limits how much information an agent can read at once.
Create a dedicated connection key in NocoDB and connect it from Integrations. Agents can discover bases and tables, read at most 100 records at once, and retrieve one record. Creating, changing, or deleting a record requires an exact human confirmation and a unique change reference, so repeating the request does not repeat the change.
The Apps & Services card does not currently install NocoDB because the current release did not pass security review. This is still a user-operated integration: HivemindOS does not provide hosted NocoDB workspaces, and NocoDB’s current license requires a separate commercial agreement before a company may offer it as a hosted service to customers. See NocoDB for connection and recovery guidance.
Booking Services
HivemindOS supports connected Cal.com accounts, an audited local Cal.diy Docker stack on a selected owner-authenticated fleet machine, and an official managed booking worker for eligible subscriptions. Agents use booking_read and booking_manage; mutations require action-specific confirmation and a stable idempotency key. Company calls additionally require a working company task assigned to a member agent and reserve against company integration limits.
See Agent-Managed Bookings for local publishing, backups, subscription limits, provider selection, confirmations, Zero Human Company governance, and the current live-service boundary.
Palmier Pro Service
The Apps & Services view includes Palmier Pro as an installable macOS video editor for agent-assisted timeline work. HivemindOS installs the reviewed GitHub release DMG by downloading the pinned asset, verifying its SHA-256 digest, mounting it read-only, and copying the app bundle into Applications. It does not run a shell installer.
Palmier Pro requires macOS 26 Tahoe on Apple Silicon. When the app is open, it exposes a local MCP endpoint at http://127.0.0.1:19789/mcp. Agents should connect to that loopback endpoint only after the user has opened the intended Palmier project and confirmed that timeline or media mutations are in scope.
Hivemind Office Companion
The Apps & Services view discovers existing Hivemind Office, HermesOffice, and GenOffice desktop bundles on macOS and Windows. HivemindOS does not automatically install HermesOffice: the audited source revision does not provide a signed, immutable binary artifact with a reviewed digest, so the card reports Install blocked instead of cloning or building mutable third-party code.
Agents use the existing bundled HivemindOS MCP server rather than an editor-owned model provider. Five tools separate status, inspection, desktop opening, review preparation, and final application. HivemindOS restricts paths to the local home/vault boundary, snapshots full file hashes, defaults to a non-overwriting copy, and requires separate exact confirmations for saving a copy and replacing the original. Replacement creates and verifies a sibling backup first.
See Hivemind Office Companion for the supported formats, exact source provenance, review workflow, MCP tool names, conflict behavior, recovery path, and credential boundary.
Agent Mailboxes
Agent Settings exposes a mailbox action for existing agents. The intended user flow is: select an agent, press Create mailbox, and receive a persistent address for that agent. The app stores mailbox ownership under HivemindOS state and does not ask the user for per-agent IMAP, SMTP, password, or host settings in the primary flow.
HivemindOS does not currently advertise a first-party managed mailbox broker as live. Reserved broker endpoint settings remain blocked until the app and hosted service share a reviewed provisioning contract. Today, hosted mailbox provisioning uses AgentMail, while Cloudflare Agentic Inbox provides the self-hosted Cloudflare path.
Once agents have mailboxes, the outreach threads they send and receive are streamed onto a company’s cockpit in the Zero Human Companies Emails tab, which reads across AgentMail and Cloudflare Agentic Inbox and lists the crew’s mailboxes with per-mailbox status.
GET /api/agents/mailbox?agentId=<agent> returns existing mailbox records plus provider readiness. POST /api/agents/mailbox with action: "create", agentId, and agentName attempts live provisioning. A mailbox is marked ready only when the selected provider can both receive mail for the address and send live internet email from the same domain. If no provider is live-ready, the API returns a blocked provider report with concrete setup blockers instead of creating a fake mailbox.
AgentMail is one hosted backend for this contract. Configure AGENTMAIL_API_KEY in shared env; optional AGENTMAIL_DOMAIN or HIVEMINDOS_AGENTMAIL_DOMAIN selects a verified custom domain, otherwise the provider defaults to agentmail.to. HivemindOS creates each inbox with a deterministic client_id so retries do not create duplicate AgentMail inboxes. Optional AGENTMAIL_API_BASE_URL or AGENTMAIL_API_URL can point at another AgentMail-compatible API host.
Cloudflare Agentic Inbox is another backend for this contract. It requires a Cloudflare DNS domain with Email Routing and Email Sending ready, and an Agentic Inbox Worker target for routing rules. The app may create the per-agent routing rule automatically once the provider is ready; the user should not configure IMAP or SMTP for each agent.
MCP Email Server remains useful for advanced bring-your-own-mailbox deployments and regression tests. It is not the default mailbox creation UX.
Agentic Inbox Setup
GET /api/cloudflare/agentic-inbox returns the Cloudflare inbox blueprint plus setup status. POST /api/cloudflare/agentic-inbox supports:
action: "scaffold"or"install"to create a real Worker project under the HivemindOS home directory.action: "deploy"or"start"to install dependencies and runwrangler deploy.
The scaffold includes Email Routing handling, a SQLite-backed Durable Object inbox store, R2 attachment binding, Workers AI binding, an Email Sending binding, health/catalog routes, draft creation, and explicit approval before replies are sent. Generated package scripts use npx wrangler, so npm run check can validate the Worker dry-run before a local dependency install. Deployments still require Cloudflare account auth, an onboarded sending domain, and Email Routing configured to deliver the mailbox to the Worker before it receives real mail.
MCP Email Server
The Apps & Services view includes MCP Email Server as an installable local bridge for agent-readable mailboxes. HivemindOS installs the PyPI package with uv tool install mcp-email-server and reports setup readiness through /api/fleet/apps/installable-services.
The bridge is intended to be spawned by an MCP client over stdio, for example with mcp-email-server stdio after installation or uvx mcp-email-server@latest stdio when the operator prefers a per-run package launch. HivemindOS does not keep the email bridge running in the background and does not read mailbox credentials during status checks.
For advanced deployments, configure mailbox access through environment variables in the MCP client or shared env. The required read keys are MCP_EMAIL_SERVER_EMAIL_ADDRESS, MCP_EMAIL_SERVER_PASSWORD, and MCP_EMAIL_SERVER_IMAP_HOST. MCP_EMAIL_SERVER_SMTP_HOST and related SMTP keys are optional; when SMTP is omitted, the MCP server runs in read-only IMAP mode and hides outbound email tools.
Regression coverage is available with pnpm test:mcp-email. The default suite launches throwaway local SMTP and IMAP fixtures plus real mcp-email-server stdio processes to verify agent-to-agent delivery across distinct email addresses, reply threading, message search and date filters, read state, mailbox listing, moving, deleting, draft saving, Sent copies, HTML parsing, CC/BCC delivery privacy, attachment download, MCP process restart persistence, and read-only IMAP behavior when SMTP is not configured.
Production-provider coverage is available with pnpm test:mcp-email:real. It is opt-in and skips unless MCP_EMAIL_REAL_E2E=1 is set. Configure two live test accounts with MCP_EMAIL_REAL_ALPHA_EMAIL_ADDRESS, MCP_EMAIL_REAL_ALPHA_PASSWORD, MCP_EMAIL_REAL_ALPHA_IMAP_HOST, MCP_EMAIL_REAL_ALPHA_SMTP_HOST, and the matching MCP_EMAIL_REAL_BETA_* keys. Optional per-account keys include ACCOUNT_NAME, FULL_NAME, USER_NAME, IMAP_PORT, IMAP_SSL, IMAP_START_SSL, IMAP_VERIFY_SSL, SMTP_PORT, SMTP_SSL, SMTP_START_SSL, and SMTP_VERIFY_SSL. The real-provider suite sends unique test messages between the two accounts, verifies live delivery, reads, replies, attachments, HTML, mailbox listing, read-only mode, and deletes the test messages unless MCP_EMAIL_REAL_KEEP_MESSAGES=1 is set.
PRD Decomposition
POST /api/queen-bee with action: "decompose-prd" parses a PRD into an epic and linked implementation tasks. Use preview: true to inspect the generated task set without writing to the Work Board.