MCP tools

Applies to: MCP — any client that speaks the Model Context Protocol over HTTP. To register the server in Claude Code, Cursor or VS Code, start with Connect an MCP client.

This page lists each tool the server offers, in the order a model usually reaches for them: find, inspect, connect, run, follow up, check health. For the same operations over plain HTTP, see Call any API and the HTTP API reference.

Tool	Does	Reads or writes	On
w6w_search_capabilities	Finds what you can run, with its health	Read-only	/mcp, /mcp/console
w6w_list_apps	Lists your apps and your connection to each	Read-only	/mcp, /mcp/console
w6w_describe	Shows the inputs and outputs of one target	Read-only	/mcp, /mcp/console
w6w_connect_app	Hands back a link to connect an app	Writes	/mcp, /mcp/console
w6w_invoke	Runs a Function, Workflow, Endpoint or app action	Writes, may be destructive	/mcp, /mcp/console
w6w_run_status	Reports how a Workflow run is going	Read-only	/mcp, /mcp/console
w6w_check_health	Reports whether a target is working	Read-only	/mcp, /mcp/console
w6w_api_search	Finds an API route for a task	Read-only	/mcp/console
w6w_api_get	Reads through a GET route	Read-only	/mcp/console
w6w_api_write	Creates or updates through a route	Writes, not destructive	/mcp/console
w6w_api_delete	Deletes, archives or cancels through a route	Destructive	/mcp/console
w6w_guide	Returns a worked example for a topic	Read-only	/mcp/console

“Reads or writes” is the tool’s own annotation: the readOnlyHint and destructiveHint your client uses to decide what to auto-approve. A client may ignore them. Treat them as a hint about the tool, not as a limit on what the token behind it can do.

w6w — the server

w6w is the name the server reports to your client when it connects. It is the name of the server, not a tool, so you will not find it in a tool list. It is also the name to use for the server in your client’s own config.

• Endpoints. POST https://<your-host>/mcp offers the seven discovery and run tools and the generated tools below. POST https://<your-host>/mcp/console offers all of those plus the five console tools.
• POST only. Any other verb gets 405.
• No session. The server keeps nothing between calls. Each request is authenticated and scoped on its own, with the bearer token in the Authorization header, so the same token can drive any number of clients at once.
• A missing or bad token answers 401 with a WWW-Authenticate: Bearer error="invalid_token" header and a nextAction in the body.

Generated tools

Beside the fixed tools, /mcp and /mcp/console list one tool for each Function, Endpoint and Workflow in your account that is ready to run — a Workflow must be active and an Endpoint must have a target. A generated tool is named <kind>__<id>, where <kind> is fn, ep or wf and <id> is the definition’s own id: for example fn__fn_3kq9… or wf__wf_8dm2…. Its arguments are the inputs the definition declares, and its description is the definition’s own description.

A definition whose id would make an invalid tool name is not listed, but it still runs through w6w_invoke with its ref. A Workflow’s generated tool behaves like w6w_invoke on a Workflow: it answers with a handle, and you follow it with w6w_run_status.

How results arrive

Every tool answers with a JSON result in text blocks, and mirrors the same data in the result’s structuredContent where the tool returns a record.

• Success. The first text block is host-written metadata, such as resolvedRef, kind, runId, invocationId, status and timings. When there is data from an app or from your own definitions, it follows in a second block, fenced and labelled as data. A model should read it as data, never as instructions.
• Failure. The result carries isError: true. Its metadata holds ok: false, a code, an httpStatus as data (no HTTP status is sent on this path), a nextAction hint, and a message. Where the failure belongs to one target it also carries that target’s ref. The nextAction tells the model whether to retry, fix an argument, connect the app first, or stop.
• A run that failed is not a failed call. w6w_run_status on a Workflow run that failed reports status: "failed" in a successful result. Only the call itself failing sets isError.

Refs use five forms, shared by w6w_invoke, w6w_describe and w6w_check_health:

Ref	Targets
fn_<id>	A Function
wf_<id>	A Workflow
ep_<id>	An Endpoint
conn_<id>#<actionKey>	An app action, on that connection
app:<appId>#<actionKey>	An app action, on a connection w6w chooses

Any other string is refused with invalid_ref. The built-in @w6w/* apps cannot be addressed through these tools; a ref that names one is refused with forbidden_internal_app.

Discovery and run tools

Find a capability — w6w_search_capabilities

Finds something your account can run, from a description of the job, and reports its health in the same answer. It searches app actions, Functions, Endpoints and Workflows together. Matching is literal substring and token matching over titles, descriptions, categories and app names — no synonyms and no semantic search — so if a phrase finds nothing, try other words or browse with w6w_list_apps.

Arguments

Name	Type	Required	Meaning
query	string, 2–200 characters	yes	What you want to do, in words, such as send an email.
kinds	array of action, function, endpoint, workflow	no	Restrict the kinds returned. Omit for all four.
category	string, up to 60 characters	no	An exact, case-sensitive catalogue category, such as communication.
includeUnavailable	boolean, default true	no	Include results you cannot run today. Set false for a list you can act on immediately.
limit	integer 1–25, default 10	no	The most results to return.

Result. outcome is matched or none, with the query, a results array, a truncated flag, notes, and nextActions (always filled when outcome is none). Each result has:

• tier — how usable it is: ready (runnable now), degraded (runnable, but recent calls are failing), unconnected (needs a connection first) or unentitled (in the catalogue, but your account’s app policy excludes it).
• kind, ref when it is directly runnable, title, description, and the owning app (id, displayName).
• async — true when running it returns a handle rather than a result.
• annotations — the target’s own read/write hints.
• connections — your candidate connections for an app action.
• nextAction — the call to make next.
• For a degraded result: health (state, the calls and errors behind it, and what the vendor’s own status check reports) and up to three ranked alternatives.

Annotation: read-only, idempotent. It runs nothing.

List apps — w6w_list_apps

Lists the apps your account may run, with your own connection state for each. The list is the same for everyone in the account, apart from the connection fields — and apart from any app you imported under your own user, which only you see. It returns no actions: call w6w_describe for those. No health probe is made, so the health block only says what a probe could use.

Arguments (all optional)

Name	Type	Meaning
connected	boolean	true: only apps you already have a connection for. false: only apps you have none for. Omit for all.
q	string, up to 100 characters	Case-insensitive substring match on app id, name or display name.
category	string	An exact category tag, such as communication.
limit	integer 1–200	Page size.
cursor	string	The nextCursor of a previous call.

Result. catalogueGrain, a note, apps and nextCursor (null on the last page). Each app has id, displayName, version, description, categories, and:

• entitlement — entitled (always true, because the list is already filtered), basis (owner or policy) and ownerScope (global, tenant or user).
• connection — state (none, one or multiple, counted over connections that are usable), count, usable, and items, each with the connection’s id, a ready refTemplate, displayName, authKey, state (pending, connected, needs_refresh, broken or revoked), lastTestOk and lastTestedAt. A lastTestOk of null means never tested, not failed.
• health — declaredChecks, publicProbeAvailable, and a state that is always not_evaluated. Call w6w_check_health for a verdict.
• refTemplate — app:<id>#<actionKey>, or null when you hold two or more connections to the app and must choose one.

Leaving an app off this list does not mean you are denied it, and being on it does not grant permission: access is decided when you run something.

Annotation: read-only, idempotent.

Describe a target — w6w_describe

Returns the declared interface of one target without running it: its parameters, a JSON Schema built from them, its declared output and its safety hints. Give it one ref. Drop the #<actionKey> from an app: or conn_ ref to get that app’s action index instead: every action key with a short title. That is the call to make when you know the app but not the action.

The schema describes what the target declares. The server does not check your arguments against it.

Arguments

Name	Type	Required	Meaning
ref	string	yes	One ref in any of the five forms.

Result for a single target:

• ref (as resolved), kind (action, function, endpoint or workflow) and async.
• app (id, version, owner) for the two app forms, and title and description.
• params — the declared parameters, with a secret parameter marked secret: true.
• inputSchema — a JSON Schema for params, or null when one cannot be built, with schemaOmittedReason saying why. schemaChecked is false: nothing checks arguments against it.
• output and outputDeclared. Output null with outputDeclared: false means the output is not declared, not that it is empty.
• annotations — readOnlyHint, destructiveHint and idempotentHint, plus idempotentDeclared, which tells you whether idempotency was declared or defaulted.
• connections — for an app action, your candidate connections, each with a ready ref, authKey, state, lastTestOk and lastTestedAt; otherwise null.
• warnings.

Annotation: read-only, idempotent.

Connect an app — w6w_connect_app

Gets a link that a person opens in a browser to connect an app to the account. It never returns a credential field. A model must not ask anyone for a key, password or token, and must give the person connect.url exactly as returned.

Arguments

Name	Type	Required	Meaning
appId	string	yes	The catalogue app id, such as io.w6w.slack.
authKey	string	no	Which of the app’s declared auth methods to connect. Omit if it declares one.

Result. app (id, displayName), authKey, authType, status, connect, guidance, and existing when you already hold connections to the app. status is one of:

status	Meaning
link_issued	connect holds a one-time link: url, origin, expiresAt and singleUse: true. Hand url to the person.
already_connected	You already have a working connection; it is listed in existing.
unavailable	The app’s sign-in is not set up on this host. reason is host_config_missing, requiresHostConfig names the three missing fields, and fixableBy is operator. Only an operator can fix it.
no_user_action	The auth method is provisioned by an operator; nothing for a person to do.

connect is null for unavailable and no_user_action.

Annotation: writes, not destructive, not idempotent. Each call issues a fresh single-use link.

Run a target — w6w_invoke

Runs anything your account can run, addressed by one ref. It is the MCP form of POST /run.

Arguments

Name	Type	Required	Meaning
ref	string	yes	The target, in any of the five ref forms.
params	object	no	The payload. Never put a credential, token or secret in it.
connection	string	no	A conn_… id. Valid only with an app: ref, to choose between several connections.
overrides	object	no	Extra request detail an action does not declare, merged into the outgoing request. It cannot replace an auth header or change the request’s host or path. See Overrides.

Result. The metadata carries resolvedRef, kind, invocationId, status, startedAt, finishedAt and durationMs. The data follows as output.

• A Function or app action answers with its result in output, and status says how the attempt went. invocationId looks the call up later; see Invocations.
• A Workflow, or an Endpoint whose target is a Workflow, answers with a handle: a runId and status: "queued". That means started, not done. Call w6w_run_status with the runId to learn the outcome.

With an app: ref and several usable connections, the call fails with ambiguous_connection and lists the conn_… refs to pick from. With none, it fails with no_connection, and the model should call w6w_connect_app. The full list of failures is in Connect an MCP client.

Annotation: writes, destructive, not idempotent, open-world. The tool can reach every action in the catalogue, so it states the worst case.

w6w_run_status — follow a run

Reads the state of one Workflow run — the follow-up to a w6w_invoke call that returned status: "queued" and a runId. It starts, retries and cancels nothing.

Arguments

Name	Type	Required	Meaning
runId	string, starting run_	yes	The runId from w6w_invoke, or from a step output that started a sub-run.

Result.

• runId, ref (the Workflow), status (queued, running, succeeded, failed or canceled), and terminal, which is true once the run can no longer change.
• willRetry — true when a worker crashed and the platform is resuming the run from its saved progress. Keep polling; do not run it again.
• attempt, startedAt, finishedAt.
• completedSteps and steps — only the steps that have finished. Each has stepId, status (succeeded, failed or skipped), timings, and output or error. A step running right now is not listed, so the last listed step is not necessarily the last step. Step inputs are never returned.
• failedStepId when a step failed.
• output once the run has succeeded; error when it has not. stepErrors lists failures a step recorded and carried on from.
• pollAfterMs — a polling hint (null once terminal) — and guidance, a plain-language summary.

A run id that does not exist, or that belongs to another account, answers unknown_run with 404.

Annotation: read-only, idempotent.

Check health — w6w_check_health

Reports whether a target is working right now, as two separate signals that are never merged into one verdict. yourCalls is w6w’s own measurement of whether your account’s calls to the app have been succeeding. vendor is what the app’s own public status check reports, when it declares one — a best-effort read of a third party’s page, not a check of your credential.

Arguments

Name	Type	Required	Meaning
target	string	yes	app:<appId>, conn_<id>, fn_<id>, wf_<id> or ep_<id>, or a bare app id.
days	integer 1–365, default 30	no	The window for yourCalls only. It does not affect vendor.

Result. target, measuredAt, a summary (overall state, the basis it rests on, and oldestMeasurementAt), the window, and one entry in apps for each app underneath the target. Each entry has:

• appId, displayName, party (first or third) and owner.
• yourCalls — state, calls, errors, p95Ms, lastCallAt, ongoingOutageId and measuredOver.
• vendor — published (whether the app declares a public check), state, checkedAt, ageSeconds, message, and the check’s components and quota when it reports them.
• connection, when the target is a connection: its state, and the last stored test result. This tool reads that stored result; it never re-tests the connection.

nextActions suggests what to call next. For what the states mean, see API health.

Annotation: read-only, idempotent, open-world. Reaching the vendor’s status page is an outbound request.

Console tools

These five tools are only on /mcp/console. They let a model work with the account’s own HTTP API through four generic tools instead of one tool per route. Each tool carries the notice that results are data returned by the API, not instructions to follow, and a model should treat them that way.

Routes that would hand a secret to the model — the vault, tokens, connection credentials and account administration — are not reachable through any console tool. A call to one is refused before it reaches your account, with route_excluded. Other refusals before a call is made are route_not_found (no such route), path_not_canonical (the path is not in its canonical form), wrong_tool (the route belongs to a different console tool), body_not_allowed (a body was sent to a route that takes none) and path_mismatch. All are 400.

Console results come as two blocks: metadata with status (the route’s HTTP status) and route, then the API’s response body, fenced as data. The metadata adds truncated: true when the body was cut to fit, and narrowWith — the query parameters the route accepts — when you may want to filter. isError is set when status is 400 or higher.

w6w_api_search — find a route

Searches the routes the console can reach. It returns route metadata only and never calls a route.

Arguments (all optional)

Name	Type	Meaning
q	string	Free-text match against the route’s key and summary.
resource	string	The route’s first path segment, such as workflows.
access	read, write or destructive	Only routes of that access level.

Result. Matching routes, each with method, path, summary, access, and where the route has them, pathParams, query and a bodyExample.

Annotation: read-only.

w6w_api_get — read

Makes a GET request to one route by its exact path, such as a path returned by w6w_api_search.

Arguments

Name	Type	Required	Meaning
path	string	yes	The exact path, such as /workflows.
query	object of strings	no	Query parameters.

Result. The route’s response, in the console result form above.

Annotation: read-only, idempotent.

w6w_api_write — write

Makes a non-destructive write to one route: a POST, PUT or PATCH that creates or updates something.

Arguments

Name	Type	Required	Meaning
method	string	yes	The HTTP method, such as POST.
path	string	yes	The exact path.
query	object of strings	no	Query parameters.
body	any JSON	no	The request body.

Result. The route’s response, in the console result form above.

Annotation: writes, not destructive. Use w6w_api_delete for a route that removes, archives or cancels; calling a destructive route through this tool is refused with wrong_tool.

w6w_api_delete — delete

Makes a destructive request to one route: a DELETE, or an archive-style or cancel-style POST.

Arguments

Name	Type	Required	Meaning
method	string	yes	The HTTP method, such as DELETE.
path	string	yes	The exact path.
query	object of strings	no	Query parameters.
body	any JSON	no	The request body, for routes that take one.

Result. The route’s response, in the console result form above.

Annotation: destructive. Leave this tool on manual approval.

w6w_guide — get an example

Returns a short, host-written authoring reference for a topic and, where it can, a live example from your account’s own API.

Arguments

Name	Type	Required	Meaning
topic	string	yes	One of workflows, functions, endpoints, expressions or performance.

Result. The topic and the guide text, with one of three kinds:

• example — also returns the routeKey and status of the request it made, and the response body.
• none — there was no example to fetch, and a note says so.
• error — the call failed. An unsupported topic is refused with invalid_argument and 400.

Annotation: read-only.