Changelog

What shipped, newest first. Dates are when the work landed, not when it was announced.

    • Projects are public or private, and the people on one run it

      A project is now either open to the whole team or limited to the people named on it. You set that once on the project instead of narrowing every member one at a time. Anyone in the team can create a project and runs the one they made: they decide who is on it, which toolkits it holds, and whether it stays. Switching a project to public and back keeps its member list.

    • A project can turn individual actions off

      Next to the switch that installs a toolkit there is now a switch for each action, so a project can keep a connector's reads and refuse its deletes for everyone in it. When an action is refused, the message says the project turned it off, not that the agent's own grant blocked it.

    • An agent never reaches more than the person who owns it

      An agent could already be limited to certain projects. It can now also be limited to particular toolkits inside a project, and to particular actions inside a toolkit. The gateway also used to treat its two ways in differently: a raw agent key was capped at its owner's access, and an OAuth connection was not. Both now follow the same rule, so removing someone from a project removes their agents too, however those agents connect.

    • Per-person tool grants are gone

      Access used to be set per member, per toolkit and per action. That control has been removed. Limits now belong to the project and to the agent. Anyone who can see a project can call every action it leaves enabled. Existing installs are updated automatically on their next start, and no agent key changes what it can reach.

    • Twenty-five connectors, including Jira, Microsoft, Salesforce and Sheets

      The catalogue gained Jira, Confluence, GitLab, Outlook, Teams, Azure, Google Sheets, Search Console, Anthropic, Gemini, Datadog, Grafana, PagerDuty, Kubernetes, Google Cloud, HubSpot, Salesforce, Zendesk, Intercom, Airtable, Snowflake, BigQuery, PostHog, Figma and Shopify. Installs that already exist get them on their next start, not only new ones. Outlook, Teams and Salesforce need their OAuth applications registered before they can sign in.

    • A project's URL follows its name

      A project created as "New project" and named afterwards used to keep its original URL name, and that is the name agents were told the project had. Renaming a project now changes it, and the old one keeps working as an alias, so bookmarks and anything an agent still sends continue to resolve. Existing installs are updated once, on their next start.

    • Team membership is editable

      A workspace can now move someone between admin and member, or remove them, from inside Toolspoke.

    • The marketplace remembers which project you came from

      Every way into the marketplace now carries the project you were in, so an install goes to that project and not to whichever one was first in the list.

    • The toolkit builder became a conversation

      The builder used to be a form that took one document. It is now a conversation: it can search the web, read a documentation page, ask you a question with real choices, and put the toolkit together beside the chat while you watch. Extraction is still deterministic underneath, so the model cannot write an endpoint.

    • Bitwarden, Infisical and OpenRouter

      Three connectors added to the catalogue. Installs that already exist get them on their next start, not only new ones.

    • ElevenLabs, Higgsfield and Google Drive

      Three more, added the same way.

    • A toolkit can require its guide be read first

      Some toolkits only work if their steps are done in a certain order, and a tool's input schema cannot say that. Those toolkits now mark themselves as needing their guide read first, and search results say so, so an agent reads the guide before its first call and not after its first failure.

    • The marketplace split into three sections

      Toolkits your team built now sit above the catalogue instead of mixed into it, and the separate templates page was folded into the marketplace it duplicated.

    • Tool search can be answered by meaning

      Toolkits and their tools now carry a numeric representation of what they mean, so a search can match on meaning as well as on words. This is on when an OpenRouter key is configured and the catalogue has been backfilled. Otherwise search matches on words only, and setting SEMANTIC_SEARCH=off forces that without a redeploy.

    • list_toolkits

      A search with no query already returned the whole catalogue by name, but only an agent that had read that sentence in a tool description ever found it. There is now a tool of its own for asking what is installed.

    • The gateway answers every protocol revision it supports

      It answers 2026-07-28, 2025-11-25, 2025-06-18 and 2025-03-26 on one endpoint. It used to list only two of them, which made it refuse Claude's connector with a 400 before sign-in, because that connector opens with 2025-11-25.

    • server/discover is answered before sign-in

      Behind a 401, a client cannot tell a server it has to sign into from one it cannot talk to at all. The gateway now answers server/discover before a token is presented, with the same versions and capabilities for every caller, and it reveals nothing the 401 response does not already.

    • The URL says which agent is calling

      A client that signs in carries no key, and a key was the only thing that had ever identified the agent, so those connections were attributed to nobody. The gateway now accepts the agent id in the URL path, which is the one thing a person reliably pastes.

    • OAuth tokens are checked against the deployment's own keys

      We moved to better-auth 1.7, which replaces the bundled MCP provider. An access token is now a signed JWT that the deployment issues and verifies against its own set of public keys, so a token minted for one deployment is accepted only by that one.

    • /api/health says which build is answering

      The health endpoint now reports the commit the running build came from, so you can confirm a deploy without comparing asset hashes.

    • A pinned CLI is asked what it can do

      When a toolkit is built from a command line tool, extraction and packaging read the tool's own help output to learn its commands. Before, they guessed from the README, which usually documents installation and not usage.

    • An error names whose credential was rejected

      A rejected credential used to be blamed on whichever one happened to be on screen.

    • A CLI toolkit installs the binary it needs

      The release file a command line tool needs is found, pinned, downloaded and verified before installation. A file that is not pinned or not confined is refused and never run.

    • Any URL resolves to the repository behind it

      A documentation site, a homepage or a package page now resolves to the GitHub repository it belongs to, so you can paste whatever link you have without hunting for the repository.

    • Coolify, as a CLI

      Offered as a CLI toolkit pinned to a release we checked.

    • Signing in became an install type

      A toolkit can now say that it connects by signing the user in, and what that needs. Before, OAuth was a special case built for one connector.

    • Gmail and Google Calendar

      Both connect by signing in and share one OAuth client, so a single consent covers both. Limits are left to each person's grant and are not set by the connector.

    • A call can be recorded without keeping what it returned

      Redacting by key name cannot hide a message body or a document, because those arrive under ordinary names. A connector can now declare, per operation, that its response is never stored. The call is still recorded in full: person, agent, tool, duration, status and credit.

    • Templates in review can be approved or rejected

      A template waiting for review can now be approved or rejected. A rejected one stays private.

    • One product, several ways in

      A marketplace app can offer a REST API, an MCP server and a CLI at once, and its card shows every route it offers. Notion was the first to get its REST API alongside its MCP server.

    • A missing credential answers 401, on the right host

      The 401 response now points to the host the client actually asked for, not to whatever address the deployment was configured with.

    • The project became the unit of scope

      One credential is now the unit of configuration. It replaces a per-tool setup that made two installs of the same product impossible to tell apart.

    • Clarity, Ploxs and Bash

      Three connectors added to the catalogue.

    • Sign in with Google or GitHub

      Offered when the provider's client ID and secret are configured, and absent otherwise. Email and password remain the fallback.

    • Claude can authorize itself

      Claude no longer has to be handed a key by hand. The connector registers itself, the person approves it in Toolspoke, and it is issued its own token.

    • Publishing, with verified still meaning verified

      A creator can put a toolkit forward for publishing. The verified badge is granted by review, not by publishing.

    • Relaydeck became Toolspoke

      The product was renamed throughout.

    • Toolspoke builds its own sandbox image

      Toolspoke now builds the sandbox image its tools run in. Before, the runtime assembled one, so the image a tool ran in was decided somewhere we could not see.

    • Projects

      Tools, credentials and sandbox namespaces are scoped to a project, with a page to manage them. Tests cover the isolation between projects.

    • One agent key can reach several projects

      Or all of them, as a standing grant that includes a project created after the key was issued.

    • Automated tests on every change

      Tests and coverage run on every pull request, each run with its own database so two jobs on one host cannot adopt or tear down each other's containers.

    • Deterministic extraction

      OpenAPI 2.0 through 3.2 is normalised into one canonical document, Postman collections go through the same extractor, MCP servers are read from their own tools/list, and command line tools are read behind a command allowlist. Every capability comes out with the details needed to execute it.

    • Made-up endpoints cannot get into a toolkit

      The model sorts, names and describes capabilities by id, and every id it returns is checked against what extraction found. An id that does not exist fails the build, so an invented endpoint cannot get through.

Why there are no version numbers

Toolspoke does not publish version numbers yet, so entries are grouped by the date the work was merged into the main branch.

A new connector is dated by the day it became available to installs that already exist, because that is when a running deployment actually gets it.

For how any of it works, start at the documentation.