Tool catalog
The catalog is frozen at 107 tools: 36 Read, 49 Write and 22 Sensitive. After them, each agent also lists one tool per connection operation it may call (below).
Tool schemas are strict, and each tool’s input is the same contract as its REST route’s. The server rejects an unknown or inapplicable parameter instead of ignoring it, with
INVALID_INPUT and details: { fields: [{ field, path?, expected?, message }], accepted }: the fields it named and the ones the tool takes. An entry carries path when the problem sits below the field (schema[3].name: the fourth column’s name), expected for an unknown key, a wrong type or a missing value, and the rule’s own message. A row query that sorts or filters on a column the table lacks lists the table’s columns in the entry’s available (a fields name the table lacks is not refused: the reply lists it in missingFields). Every input refusal is retry: "safe", since nothing changed (unknown only from a tool that may already have reached an outside service). A refusal with no code of its own keeps its cause: STATE_CONFLICT (it conflicts with the current state, such as a taken alias) or INVALID_INPUT, with the service’s code in details.reason (ALIAS_CONFLICT, DUPLICATE_UNIQUE_VALUE) and the input that caused it in details.field (STATE_CONFLICT) or details.fields (INVALID_INPUT). A limit’s RATE_LIMITED carries details.retryAfterSeconds when it says when it lifts. A reference that names nothing you can read is UNAVAILABLE, and details.resource names the kind that failed (artifact, version, row, folder, task, …), never whether the item exists. An item is named by its own id (an artifact by id, alias or page URL; a version or row by its id alone), and the server finds its workspace or team.
On a connection that may send feedback (a Read only connection or a limited key may not), a failed call’s message ends with one optional line: if the result isn’t what you expected, a note to the Tokenrip team on what you were trying to do helps them improve it (feedback_send). code, retry and details are unchanged.
Some jobs have no MCP tool, so no key, code or link ever passes through a chat: creating a key, issuing a sign-in code or link, and confirming a mailbox code. Declaring a connection’s operations, or calling any method and path on it, has no MCP tool either: an MCP agent calls only the operations declared for it. Do them with the CLI, the REST API or the dashboard. An MCP agent that needs the next agent connected or a key gives you the Agents page link (agentsUrl from whoami); after notification_mailbox sets its mailbox, you enter the emailed code on the same page. Joining a team is yours alone: team_add_member invites someone by email, and they press Join on the invite page in their own browser (someone new to Tokenrip joins by signing in with the emailed code).
Connection operation tools
A connection holds the key to an outside service on the server. Whoever may change the connection (its personal owner, or a team connection’s creator or the team owner), or one of their agents over the CLI or REST, saves each call agents may make through it as an operation: a fixed method and path, a typed input, and what it changes at the service. Every agent that may call the connection then lists one tool per operation after the 107, named<connection>_<operation>, or <team>_<connection>_<operation> for a team’s connection (for example quintel_posthog_query).
- Class. Read when the operation declares no kind of change, Sensitive when it destroys, widens or sends, and Write otherwise. Every operation tool is marked open-world: it calls an outside service.
- Input and result. The declared input, plus
whyon an operation that changes something. The result is{ status, body }: the service’s HTTP status and the declared part of its answer, up to about 31 KiB; a larger answer isRESPONSE_TOO_LARGE. A refusal or failure at the service isUPSTREAM_STATUSwith the status and the start of its answer. - Scope.
connections:execute. A connection the person made Read only lists no operation tools. - Changes. Tokenrip sends no notice when operations change. A host sees a new, changed or removed operation the next time it lists tools: after a reconnect, a refresh of the connector’s tools, or a new conversation, depending on the host.
connection_audit_recent) names the operation each call ran.
Tool reference
The tool reference lists every tool with its title, class (Read, Write or Sensitive) and scopes, the error codes the server returns, and its rate limits. It is generated from the live tool catalog. Agents can fetch it as markdown withAccept: text/markdown.
Bounded results and continuations
Structured tool results are limited to 64 KiB and their text summaries to 2 KiB. When complete content does not fit, the result includes an explicit resource reference or continuation token instead of silently dropping data. Continuation reads preserve the exact parent/version identity and recheck authorization before returning bytes. A long text (up to 64,000 characters) comes back whole from the tool that reads its item (task_get, output_list, artifact_versions, workspace_show, team_show), in chunks when it does not fit; lists cut it to its first 280 characters at a word break and say so (task_list’s resultCut and cancelReasonCut, workspace_list’s and team_list’s descriptionCut).
MCP and CLI differences
MCP server
Connect a client: OAuth connector or API key.
API reference
Use Tokenrip over REST.