@fivexer/mcp exposes a Fivexer workspace to Claude Code, Claude Desktop, Cursor, ChatGPT, hosted agent runtimes, and any other MCP client. The full set covers tasks, workers, skills, teams, rotas, decisions, learning, and stats. A smaller set of seven tools is there for agents that only hand work off.
It is how an agent asks for work to be routed ("file this as a billing task, high priority") and how it reads the result back, without you writing any integration code.
Hosted connector (no install)
Hosted clients — claude.ai connectors, ChatGPT connectors, anything that cannot spawn a local process — connect over HTTP and authorize with OAuth. Add a custom connector pointing at:
https://api.5xer.com/mcp
No API key is involved and there is nothing to install. The OAuth server advertises openid
and email scopes and exposes a UserInfo endpoint, so Claude Enterprise and ChatGPT Enterprise
workspaces can restrict the connector to users with a verified corporate email domain.
For Claude Code over the same transport:
bashclaude mcp add --transport http fivexer https://api.5xer.com/mcp
Most people should use the plugin instead — it bundles the routing skill alongside the tools and needs no config at all. See Claude Code setup.
Staying connected
Authorizing once is meant to last. The connector holds a short-lived access token and refreshes it in the background; the underlying connection only lapses after 180 days with no use at all, and every refresh resets that clock. You should never be asked to sign in again just because a token aged out, two chats refreshed at the same moment, or a refresh request timed out.
If a client does start asking for re-authorization repeatedly, it is a real signal rather than routine churn — the grant was revoked, the workspace was suspended, or the account that authorized it was removed.
Console → Settings → Connected apps lists every app that has been authorized, what it may do, and when it last checked in, and is where you disconnect one. Disconnecting kills every token that connection ever minted, not just the one in hand, so the app has to be authorized again from scratch.
With a workspace key instead of OAuth
Server-side agent runtimes cannot click through a sign-in screen. Claude Managed Agents, OpenAI's Responses API and Google's ADK all send a fixed bearer token, so the hosted endpoint also accepts a workspace API key:
Authorization: Bearer sk_live_...
The key reaches exactly what it reaches on the REST API. Every tool call is recorded in the audit log against the key's fingerprint, never the key itself. Setup for each runtime is in Connect hosted AI agents.
The agent toolset
An agent that only needs to hand work off should not have to read past 190 tool descriptions. Ask for the agent toolset and you get seven:
| Tool | What it does |
|---|---|
delegate | Creates a task and returns its id straight away. dryRun: true checks who could take it instead. |
await_result | Waits up to 25 seconds for the task to finish and returns the outcome. Call it again until done. |
explain_decision | Says who got the task and why: the routing mode and the top five candidates with their scores. |
get_task_context | Reads the task's title, description, context and result contract. |
cancel_task | Withdraws a task that is no longer needed. |
list_skills | Lists the skill catalog, for requiredSkills. |
check_quota | Shows how much room the workspace has left. |
Select it with a query parameter on the hosted endpoint, or an environment variable locally:
https://api.5xer.com/mcp?toolset=agent
bashFIVEXER_MCP_TOOLSET=agent npx -y @fivexer/mcp
Three things make delegate safe for an agent to use:
- Retries are safe. Pass your own
taskId. If a task with that id exists, live or already finished,delegatereturns it withdeduplicated: trueinstead of creating a second one. - You say what done looks like.
resultContract.schemais a JSON Schema the result must match. Whoever does the work sees it, and a result that does not match is refused withresult_invalidand stays with them to fix.resultContract.acceptanceCriteriais the same thing in plain words. It is shown to the worker but not checked by machine. - A failure is never refused. A worker who cannot do the job completes it with
success: falseand a reason, and that always goes through.
delegate also takes the same targeting as create_task. teamId limits the task to a crew,
and preferTeamId only offers it to that crew first. vetoedWorkers keeps it from named workers,
people or agents. latitude, longitude and maxDistanceKm limit it to executors nearby, and
requireGeo leaves out anyone without a known location. On the hosted endpoint the coordinates are
offered to workspace-key callers only; a chat client signed in with OAuth does not collect a precise
location.
The executor can be a person, an AI agent run by the agent daemon, or a crew. Your rules decide who gets it, not the vendor of the agent asking.
Local stdio server
For a local process that should use a workspace key directly:
bashclaude mcp add fivexer \ -e FIVEXER_API_KEY=sk_test_... \ -e FIVEXER_BASE_URL=https://api.5xer.com \ -- npx -y @fivexer/mcp
Both variables are required — the server refuses to start without them rather than guessing an
origin. npx @fivexer/mcp init walks through it interactively and prints ready-made snippets for
Claude Code, Cursor (~/.cursor/mcp.json), and Claude Desktop (claude_desktop_config.json).
The key you supply is a workspace key, so the tools reach everything that plane reaches. Use an
sk_test_ key against a sandbox workspace while you are exploring.
What the tools cover
| Area | Tools |
|---|---|
| Tasks | create, bulk create, get, list, cancel, accept, reject, complete, escalate, unpark, context, comments, attachments |
| Workers | upsert, list, remove, availability, queue, identities, invites, join links |
| Skills | create, list, update, delete, suggest |
| Teams | create, list, delete, members, presence |
| Routing | suggest_workers, list_decisions, list_parked_tasks |
| Learning | status, worker stats, preview weights, apply weights, record feedback, reset |
| Observability | get_stats, get_sla_stats, break_metrics, check_quota |
Read-only tools are annotated as such, so a client that distinguishes them can run those without prompting.
The concepts resource
The server also publishes fivexer://concepts — a short mental model of the platform: the task
lifecycle, how routingWeights and hard vetoes work, wildcard syntax, skill thresholds, and geo
limits. Agents that read it before calling tools make far fewer wrong-shaped requests, because the
rules that surprise people (weight 0 is a veto, wildcards are suffix-only) are stated up front.
Related
- Connect hosted AI agents — Claude Managed Agents, OpenAI, Google ADK and Gemini, and where A2A stands.
- Claude Code setup — the plugin, OAuth sign-in, and troubleshooting, step by step.
@fivexer/agent— the other direction: run an AI agent as a worker that receives routed tasks.- Auth & keys — what a workspace key can reach.