Skip to Content
Skilder

Release notes

What changed in Skilder, newest first.

v1.2.1

Fixed

  • Connecting ChatGPT to Skilder no longer fails at sign-in.

v1.2.0

Changed

  • Workspaces have three roles: Admin, Contributor and Member. A Contributor creates and edits skills, roles and Apps, as members could until now. A Member uses what is published, in Skilder and through their agent, and changes nothing: they see no drafts and no authoring controls, and their agent has no manage tool. Everyone who was a member is now a Contributor. People who join from now on are Members unless you choose another role when you invite them, and you can change anyone's role in Settings > Users.
  • Only an Admin or a Contributor can own a skill, role or App. A Contributor who becomes a Member stays the owner on record of what they owned, but can no longer change it.
  • On Tools, the banner naming failing MCP servers links admins to each server, and Fix in MCP Servers opens their settings. Everyone else is asked to have a workspace admin reconnect them.
  • Coming back to the Playground reopens the conversation you left in that browser tab. New conversation starts a blank one.
  • In the Playground, consecutive tool calls and the thinking between them fold into one line, such as Ran 4 steps. Open it to see each step, what it sent and what it returned.
  • Home's This week figures and the Value tab in Activity say how fresh they are, such as Updated 2 min ago. They can run up to five minutes behind the latest calls.
  • When you accept an invitation, the Join screen lists the workspace's six most-used roles under Roles your agent can learn, and how many more there are.

v1.1.2

Fixed

  • The ADLC Workshop template guides each step more reliably: the first story is always a running skeleton, the Product Owner is asked to publish the Project brief after requirements and again after refinement, a review gives one verdict per item, and running the project uses its own start command. A workspace that already installed the template keeps its earlier text; install the template in a new workspace to get these fixes.

v1.1.1

Changed

  • Email invitations and the invite link now live in one Invite people dialog, opened from Settings > Users, the account menu or Home.
  • Workspace setup speaks of templates instead of industries, such as Install template on Roles.

Fixed

  • The Value returned tile on Home shows large amounts in short form. Hover it for the exact amount.

v1.1.0

Added

  • A new account gets a guided start: name your workspace and pick your industry, and Skilder sets it up with ready-to-use roles, skills and tools, then helps you connect your agent, or try it in the Playground with a first prompt for each of your teams. Every workspace also gets a new Home with your agent's activity this week, and admins can invite several colleagues at once from Invite a colleague in the account menu.
  • Admins can install an industry after onboarding from the ⋯ menu on Roles, or from an industry's category in Add Role. It adds the industry's teams with their roles, skills and tools, marks the teams you already have as installed, and lets one workspace hold several industries.
  • Add Role and Add Skill offer roles and skills for six industries: Transport & Logistics, Construction & Real Estate, Wholesale & Retail, Manufacturing, Professional Services, and Software & IT. Find them under Browse all roles and Browse all skills.
  • The ADLC Workshop industry sets up a Product, a Workshop and an Engineering team whose roles take a small project from requirements to a running app, and builds a Workshop status App that ticks each step of the workshop.
  • Workspace admins can connect Amazon Bedrock as an AI provider, with an AWS region and either an AWS access key or a Bedrock API key. The model picker lists only the models that account can use.
  • Workspace admins can connect Exoscale as an AI provider with an Exoscale AI API key. Its models run in Vienna (Austria); the service is in private preview.
  • The Skilder AI provider offers Qwen3.8 — Exoscale (Austria), hosted in Vienna, beside Qwen3.8 on Infomaniak. It is off until an admin selects it on the Skilder card, and it counts against the daily budget like the other Skilder models. Every Skilder model is now named with its host and country.
  • The Skilder AI provider offers Claude Sonnet 4.6 — Amazon Bedrock (EU). Every call starts in an AWS European region (EU, Switzerland or UK) and is routed within AWS's EU geography. It is off until an admin selects it on the Skilder card, and it counts against the daily budget at Amazon's published Bedrock price for Europe (Zurich).
  • An admin's agent can browse the featured catalog and install a featured skill, role or MCP server. A server installed this way waits on its page for you to fill in its secrets.
  • Every member's agent can create and edit skills, roles, Apps and skill files, install featured skills and roles, and read teams, guardrails and MCP servers, as they can in the app. Each agent is offered only what its user's role and API key allow.
  • After you approve your agent's first connection to a workspace, Your agent is connected opens in a new tab. Its checklist ticks as your agent signs in and makes its first call, and it suggests a first prompt.
  • Refresh tools beside a server's tool list asks the server for its tools again and says how many it listed.
  • Every tool list marks a tool whose MCP server is disabled, disconnected, in error or has an invalid configuration, and names the reason, such as Server disabled.
  • A tool its MCP server no longer lists is marked No longer offered. Insights lists these tools under Tools no longer offered, with their server and the skills and Apps that still use them.
  • Admins can resend an invitation with Resend invitation in Settings > Users. Each resend sends a new link valid for 7 days, and an invitation nobody accepted in time stays listed as Expired until you resend it. When an invitation email fails to send, Skilder says so instead of reporting it sent.
  • Admins can turn on an invite link in Settings > Users: anyone who opens it joins the workspace as a member. Copy it or download it as a QR code, reset or turn it off, and see who joined through it.
  • On the Claude Desktop setup steps, Add to Claude adds Skilder to Claude Desktop in one click. The manual steps stay below it.
  • Admins can open Workspace settings from the account menu at the bottom of the sidebar, below Personal settings.
  • An App's page has an owner picker and a ⋯ menu with Duplicate app and Delete app. The delete icon is gone from the Apps list.
  • When your agent learns a role, skill or reference, MCP clients that render cards show a one-line Learned card with its name, and a role's skill count or a reference's skill. In a workspace that already has content, Skilder joined this conversation is one line too.
  • Playground and Studio show the model's reasoning in a collapsed row that reads Thinking… while it streams. Expand it to read the reasoning.
  • Your agent's workspace summary card lists guardrails with their status, and admins can publish a draft guardrail from it. A session trace names the guardrail rule that refused a call.

Changed

  • Only a skill's owner or a workspace admin can turn the skill on or off. Other members see its status without the switch.
  • Changing an AI provider's API type or URL asks for its key again.
  • When your workspace has several teams, a featured role in Add Role first asks which team gets it, and then says where it was installed. If the install fails, the reason stays in the dialog.
  • On Connections, the Any MCP Client page gives step-by-step setup like the other main clients, and every client page copies a link that names your workspace.
  • An agent's connection request stays open for 30 minutes. When it expires, Skilder shows This request has expired: go back to your agent and connect again.
  • A server's panel says when tool discovery is pending or has failed, and why. A server that offers no tools shows Lists no tools instead of an error.
  • The add-tool dialog shows its steps, and the server is added only once Test the server passes. Closing the dialog before that asks you to confirm.
  • A skill's Value section lists its value per use, all uses, priced uses and total value separately. A role's overview shows Total value, and Top skills by value shows how many of each skill's uses are priced.
  • On the Roles org chart, the role graph and the activity map, a two-finger swipe or the mouse wheel pans the view, and a pinch or Ctrl and the wheel zooms.
  • The Roles org chart hangs each level off vertical lines and shows an opened team's roles as cards. A team with 13 or more roles lists them as a roster you can filter by status, and on a phone the chart is a page that scrolls.
  • The invitation email names the person who invited you by their display name.
  • In Studio, a reply keeps running when you leave, reload or sign out. Reopening the conversation shows its live status until the reply lands, and a reply cut short is kept and marked as interrupted.

Fixed

  • Your agent receives Skilder's full welcome text, including the roles it can learn, in clients that show the model only a tool's structured result, such as Claude Code, Claude Desktop and Codex.
  • An idle MCP server no longer fails every call after the first. Re-enabling a server asks it for its tools again, and a server whose tool list fails or comes back empty is asked again before it shows an error.
  • A member deleting their account no longer stops the skill, role, App and guardrail lists, or the Call Log and Clients tabs in Activity, from loading for everyone else.
  • In the skill editor, selecting a file opens and edits that file when another file in a different folder has the same name.
  • Review changes on a role no longer shows a renamed skill as one skill removed and another added.
  • Activity counts the same calls over the same window on every tab, and a call whose tool was deleted reads Deleted tool.
  • Asking your agent for calls or failures broken down by client returns figures instead of an error.
  • The setup callout on a shared skill counts its recommended MCP servers as well as its tools.
  • Review old versions in Settings > Storage no longer lists versions whose removal would free no space.
  • Email me a code instead on the password step sends the code and opens the code screen.
  • An SSO sign-in that fails under load lands on the sign-in error page.
  • The Codex CLI setup page says Codex signs in with your Skilder account.
  • A Studio tool call that runs longer than 60 seconds no longer ends in a network error. A self-hosted install behind its own reverse proxy must let /api/chat stream unbuffered.
  • Dialogs and icon buttons work with a keyboard and a screen reader: Enter saves a team name, and closing Add Skill, Add Role or Add Tools returns focus where it was.

v1.0.1

Added

  • The Help menu shows the version of Skilder you run, on its last line. Report Issue sends it with your report.
  • Continue with SSO signs you in through your company's provider when your work email belongs to a configured SSO domain.
  • Settings > Storage shows workspace admins how far over the limit the workspace is and lets them remove older skill-version files, which leaves the version's text but prevents restoration.
  • Guardrails shows when a rule has an open draft in the list.

Changed

  • Agents use a skill file's path to learn references, run scripts or fetch assets; only binary assets can be passed as files.
  • Agents can address a skill's file by name or folder path in the Script SDK, so a file ID changing on publish does not break the reference.
  • Agents can search and page through the workspace's skills, and see a role's or skill's publish status and open draft when listing them.
  • You can publish an App from the version card your agent shows when your MCP client renders cards.
  • Activity counts one use per skill per session and leaves test calls out of the count.

Fixed

  • A remote MCP server that ends a Streamable HTTP session no longer leaves its tools failing until restart; Skilder opens a new session and retries the call.
  • The Guardrails tester accepts every field of a sample tool call.
  • An agent learning a skill gets the descriptions of the tools attached to that skill, even when another MCP server has a tool with the same name.
  • Script errors keep their full multi-line message instead of showing only the first line.

v0.15.5

Fixed

  • Test Server fills in a configuration value written as a template, such as Bearer {token}, before it tests the server. A server that takes its token in a header no longer fails the test with rejected credentials.
  • When Skilder does not answer as you open the app, the app keeps retrying. Connections and Settings > AI Providers fill in once Skilder answers, instead of after a page reload.

v0.15.3

Added

  • Settings > Storage shows how much your workspace stores against its storage limit, split by kind, with its largest skill assets and the space older published versions still hold. Only workspace admins see it.

Changed

  • Suggesting skills for a role, getting AI help with an MCP server configuration, and the sub-agents a skill script delegates to count toward the Skilder AI provider's daily budget. Once the budget is spent, they stop with the daily limit message, as chat does.

Fixed

  • A custom provider with the OpenAI API type streams its whole answer when its endpoint adds file annotations or citations to a reply, instead of ending on "An error occurred."

v0.15.2

Added

  • In Playground, attach files to a message with the paperclip button or by dropping them on the message box. Each file can be up to 10 MB, and Skilder can read the images, PDFs and text files you attach. A message with files still needs some text.

Changed

  • The Skilder AI provider offers Qwen/Qwen3.8-27B-FP8 and swiss-ai/Apertus-v1.5-70B, and new workspaces default to Qwen3.8. Workspaces that used Qwen/Qwen3.5-122B-A10B-FP8 or mistralai/Mistral-Small-4-119B-2603 on it move to Qwen3.8, and so do their conversations. You can select Mistral Small 4 again on the Skilder card.
  • Each tool row in Playground names what the tool acted on, such as the role a learn call learned.

Fixed

  • Usage of the Skilder AI provider is counted at its list price, a tenth of what was counted before. Your daily budget now covers about ten times as many tokens, and past usage is recounted at the list price.
  • Replacing a file from Asset Settings while you edit a published skill replaces it in your draft, instead of in the published skill.
  • Replacing an asset with a file whose name has spaces or brackets gives the asset a valid name, as Add Asset does, so you can still rename it later.
  • A file named only in non-Latin letters, such as 报告.pdf, keeps its extension when you add it to a skill. It is added as unnamed.pdf, as the right kind of file.

v0.15.1

Added

  • Build an App, under Authoring, shows how to write an App, give it tools, preview and publish it, and offer it to your agent through a skill.

Changed

  • Studio opens from its row in the sidebar, which shows when Studio is turned on for your workspace. The flask icon in the List view of Roles and Start in Studio on an empty Roles page are gone. A direct link to Studio still opens it.
  • The product tour stops at Playground instead of Studio.
  • When your workspace has no AI provider, Playground links admins to Settings > AI Providers instead of Studio, and tells members to ask a workspace admin.

v0.14.4

Added

  • Apps, under Capabilities in the sidebar and marked Beta, are in every workspace. An App is an HTML component you write, with its own list of tools, that your agent opens in the conversation for you to use. You need an MCP client that renders cards.
  • Playground, under Agents in the sidebar and marked Beta, is a chat with Skilder about your workspace, where tool results can show as cards. It uses your workspace's AI provider.

Changed

  • Studio no longer has a row in the sidebar. Open it from a role's flask icon in the List view of Roles.

Fixed

  • When a remote MCP server refuses the token or header in its configuration, Test Server says the server rejected the credentials in the configuration, instead of blaming OAuth, and suggests the Authorization: Bearer <token> form many servers expect.
  • After a failed Test Server, trying again no longer shows an error about cleaning up the test server.
  • When your OAuth sign-in in the Add Tools dialog has expired, the dialog offers Connect again instead of failing every test until you reopen it.

v0.14.3

Added

  • When a tool needs a scope your OAuth connection to its MCP server does not hold, the call fails with a message to reconnect, and the server's Status in Settings > MCP Servers reads Needs reconnect. Click Reconnect to grant the missing scope. Your other tools on that server keep working meanwhile.

Changed

  • Connecting an MCP server with OAuth sign-in asks for the scopes the MCP server names, instead of every scope its sign-in service offers. A reconnect asks again for the scopes you granted before.

Fixed

  • An MCP server with OAuth sign-in that your agent added gets its tools when you first connect to it, instead of showing 0 tools. A server already stuck at 0 tools gets them after you disconnect it in Personal integrations and connect again.
  • Connecting an MCP server that names no scopes no longer ends on an invalid_scope error when its sign-in service refuses its own scope list. Skilder retries the sign-in without a scope.
  • When your browser blocks the sign-in window while you add an MCP server with OAuth sign-in, no server is created.

v0.14.2

Changed

  • With session tracking on, an agent runs a workspace tool only when it belongs to a skill its session learned, directly or through a learned role. Any other tool is refused, and the error names the tools the session has learned.
  • A skill script that calls a name two of the skill's attached tools share is refused, and the error names their MCP servers. Detach all but one.
  • On a bundled MCP server, such as the Google server, the Configuration summary and the Fields tab of Edit configuration no longer show the script the server runs. The JSON tab still does.

Fixed

  • A skill script can no longer call a tool whose Status is Disabled.
  • After a network loss or a computer waking from sleep, Skilder reconnects on its own instead of showing Something Went Wrong until you reload.
  • Connecting an MCP server with OAuth sign-in no longer fails with "OAuth provider not configured" when the browser reports the sign-in window closed early. The dialog shows Waiting for sign-in… with Cancel and Try again.
  • When renaming an asset fails, the message says why.
  • Agents no longer time out connecting when several connect at once to a workspace with many tools.
  • When GitHub limits import requests, the error says when to try again.

v0.14.1

Added

  • File retention in Settings > Workspace sets how long files from in-app conversations and agent sessions are kept after their last activity: 10, 30 or 90 days, 10 by default. Skill assets never expire. In Studio, a file that retention removed shows as expired, and the conversation text stays.
  • A file your agent uploads with upload_skilder_file, and every file a tool or script returns, is kept in your workspace under an @skilder-file: path. Your agent passes that path to a later tool or script call, downloads the file with read_skilder_file, or attaches it to a skill as an asset without uploading it again.
  • Agents on MCP revision 2026-07-28, such as current Claude Code and claude.ai, connect to Skilder on that revision. Older clients keep working on the same address. Skilder also connects to an installed remote MCP server on 2026-07-28 when the server offers it.
  • A remote MCP server whose sign-in accepts Skilder's published client details is offered OAuth sign-in with no client to register, and shows as Auto-configured.
  • When your own OAuth connection to an MCP server is revoked or its refresh fails, the server's Status in Settings > MCP Servers reads Needs reconnect, with a Reconnect button.

Changed

  • An @skilder-file: path in the older form, which names a runtime instead of a stored file, is refused, and the message points your agent at the current form. If a skill's instructions spell out such a path, replace it with one upload_skilder_file returns.
  • A skill's scripts, and the sub-agents they delegate to, call only the tools attached to the skill, and a sub-agent gets no Skilder tools such as learn. Skilder attached to each skill the tools its TypeScript scripts already called by name, where exactly one tool in the workspace has that name, so those keep working. A name several tools share, or one no tool has, was left unattached and its calls are refused until you attach the right tool. The script editor warns about a tool the skill lacks.
  • The sandbox database server (@skilder-ai/sandbox-db-server) keeps its data when Skilder restarts or redeploys the runtime, and when your workspace moves to another runtime.

Fixed

  • Accepting an invitation no longer sends you back to the sign-in page about 15 minutes later.
  • A custom provider whose API Type is openai answers in Studio. When a provider call fails, the chat shows the provider's status and response instead of a generic error.
  • In Activity, a session no longer picks up a role that another session learned on the same agent connection.

v0.14.0

Added

  • MCP servers can be edited in place from Settings > MCP Servers: add, remove or rename arguments, environment variables and headers, mark a value secret, edit the configuration as JSON with Ask AI, and change the repository and the authentication method. The server keeps its tools and skill links. A server showing Invalid config is repaired the same way, with Edit configuration.
  • A Python skill script can declare its dependencies, inline or in a requirements.txt reference, and the sandbox installs them before it runs.
  • Your agent can attach a text file it holds on disk, such as a large CSV or JSON file, and it becomes an editable reference, the same as a text file added with Add asset.
  • Your agent can download any file of a skill with read_skilder_file, references and scripts included, not only assets.

Changed

  • You stay signed in while you use Skilder. A session ends after 7 days without use, instead of 7 days after you signed in.

Fixed

  • When your session has ended, Skilder sends you to the sign-in page once, instead of reloading it over and over.
  • An MCP server that fails to start shows the error that stopped it, such as a timeout or a crash, instead of an unrelated line from its output.
  • Inline code in skill and role instructions no longer shows a stray backtick on each side.
  • The version history keeps words whole and fits its card at every width, and on phones its header no longer covers the first versions.

v0.13.2

Added

  • Connected apps is a new tab in Personal settings. It lists every AI application you signed in to Skilder over MCP, with its workspace, when you granted it and when it was last used. Click Revoke to end an application's access, and follow the reconnect steps when a row says Reconnect needed.
  • Your agent can now edit an installed MCP server in place: its name, description, configuration and auth mode. The server keeps its tools, skill links and enabled state. Secrets read back as ********, and the transport, the remote URL and the program a server runs cannot change: install a new server for that. Only admins can make the edit.
  • On a role page, the breadcrumb now ends with the role's name. When the role's team holds other roles, click the name to switch to one of them.
  • The Google server (@skilder-ai/google-server) has a draft-reply tool. It saves a reply as a draft inside the original Gmail thread, and the recipient's mail client threads it too.

Changed

  • A skill file's type now follows its name, whether it comes from the editor, your agent, a ZIP file or GitHub. Any text file becomes an editable reference, including SQL, HTML and CSS, and a .txt file imported from GitHub stays plain text instead of turning into Markdown. Files already in a skill keep their type.
  • An MCP server whose saved configuration is no longer valid now shows Invalid config and stops serving its tools. Saving an invalid configuration is refused with the field at fault, instead of reporting success.
  • On an MCP server's settings, clearing a secret field and saving keeps the saved secret instead of erasing it.

Fixed

  • Your agent can rebuild a skill's files as you have them: a file it creates can go in a folder, a .sql file keeps its name instead of becoming .sql.md, and a JSON or CSV reference reads back with its content instead of empty.
  • Bash skill scripts run. Before, every Bash script failed to start.
  • An agent sending wrong credentials no longer locks every other agent out of Skilder for up to an hour.
  • Renaming your workspace, changing its currency or switching session tracking no longer resets its timezone.
  • Agents see a replaced asset file, and a skill imported from GitHub or Community, right away. Before, they could keep the old version until they reconnected.
  • When you open a skill someone else is editing and you can read their draft (as an admin or the skill's owner), the banner now says you are viewing their unpublished draft, not the published version.
  • The Google server: reply-email with reply-all addresses each recipient once and never you, and a reply to a message you sent goes to its recipients instead of back to you. A sender name containing a comma no longer turns into an extra recipient.
  • The link to your MCP servers on the Tools page opens Settings instead of an error page. Members, who cannot open Settings, no longer see it.

v0.13.1

Added

  • Publish has a menu next to it now. Choose Publish with label… to name the version you are publishing, on a skill, a role or a guardrail. The label shows in the version history and in Activity.
  • MCP App: Your agent's cards now work in every workspace, with nothing to enable. A display or analytics card opens a session or a call in place, pages through long results, and offers a few follow-up questions. You need an MCP client that renders cards.

Changed

  • Guardrails are now in every workspace, still in beta. The Guardrails row in the sidebar carries a Beta tag, and there is nothing to ask us to enable. A workspace with no rule switched on is unaffected.
  • A team that is still in use can no longer be deleted. Skilder names what holds it: teams under it, roles published in it, drafts moving a role into it, or being your only team. Pick a replacement team and click Replace and delete to move those roles and drafts there, then delete it. See Organize roles in teams.
  • Accepting an invitation now opens the workspace on Home instead of Activity.
  • When a member leaves, is removed or has their account deleted, the skills, roles and guardrails they own pass to an admin of the workspace.

Fixed

  • Deleting a member's account no longer stops the Skills, Roles and Guardrails lists from loading.
  • Your agent can run the scripts of a skill draft you have not published yet. Before, execute_script said the script did not exist.
  • An edit that reaches a skill, role or guardrail while it is being published is refused, instead of being reported as saved and then lost.
  • A slow save on one role or guardrail no longer holds up saving and publishing the others. An AI rewrite that finishes after you discarded the draft no longer lands in the next one.
  • A change to your workspace, such as switching a guardrail off, now reaches your agents within minutes even when its update notice is lost.
  • Someone who was both an admin and a member is listed once in the member list, and removing them removes them completely.
  • Moving a team with your agent no longer reports an error when the move worked. The agent's role tools now say that a change waits on the role's draft until you publish it.
  • read_skilder_file only reads files from your own workspace, and says clearly when a file reference points somewhere it cannot reach.
  • The Word template server (@skilder-ai/ms-word-template-server): get-template-info works again, and validate-template now agrees with process-template. It accepts a loop closed with {/} and flags a loop that opens and closes in different places, such as inside and outside a table cell.

v0.13.0

Added

  • Your workspace now opens on a Home page. It has shortcuts to Create a role, Create a skill, Add a connector and Connect your agent, shown only when you are allowed to use them, and an Ask your agent band that suggests one task at a time to copy into your agent.
  • A new Skills tab in Activity ranks every skill by Calls or by Unpriced uses, 50 at a time. Members see their own calls, and admins can also filter by user.
  • On the Value tab of Activity, click a person in Top users by value to narrow every figure to them, and click Show all users to go back. Find a person searches past the top five.
  • Skill scripts now work with files. Your agent can pass a file it uploaded, or one of the skill's assets, as an argument, and gets back a reference to each file the script declares with output() from the Script SDK or prints as a path.
  • In beta, on request. Your agent can now work with Guardrails. An admin's agent can list, write and test a rule, and publishing it stays a step you take in the app. Any agent can tell you why a call was refused: the rule that decided it, whether it was blocking or observing, and its reason. Guardrails are off by default and there is no setting to turn them on yourself: ask us to enable them for your workspace.

Changed

  • Navigation moved from the top bar to a sidebar with four chapters: Capabilities (Roles, Skills, Tools), Governance (Guardrails, Dependencies, Evals), Agents (Connections, Studio) and Monitoring (Activity, Insights). Skills and Tools are rows of their own instead of tabs, and Insights is its own page instead of an Activity tab. Workspace settings is now in the workspace menu, and Personal settings and Appearance are in your user menu.
  • The dependency map moved out of Activity to its own Dependencies page, under Governance. Keep clicking nodes to follow a chain: earlier selections stay in view, clicking one again closes it, and everything off the trail is hidden.
  • The Roles page now opens on the Org chart, and the cards view is gone: switch to the list view instead. Teams start folded, with a chevron to show the roles inside, you can zoom in to 300%, and opening a role lands on its Overview.
  • A role's Graph tab opens in Columns, highlights a card's whole chain when you hover it, and has a search. Double-click a skill to open it.
  • The Add Role and Add Skill dialogs open on featured catalog entries you install in one click, with Browse all roles or Browse all skills for the rest and a Create from scratch card below. Add Skill also imports from Community, GitHub or a ZIP file, and no longer offers Studio.
  • The Guardrails list shows each rule's Coverage.

Fixed

  • Files your agent uploads reach your workspace again. Uploads were failing with a server error, which also stopped scripts that take a file.
  • Edits made while a skill is publishing are held until the publish finishes, instead of racing it. If the publish fails, they are saved.
  • A guardrail rule's figures now count only its own decisions. A rule that agreed with another rule's refusal used to count as refusing, and one that ran fine could show up in Needs attention as erroring.
  • A tool call that fails after a guardrail judged it keeps that verdict in Activity.
  • The guardrail rule editor no longer drops characters when you type fast. The rule page catches up when its draft is published, discarded or taken over elsewhere, and tells you which.

v0.12.5

Added

  • In beta, on request. Guardrails are workspace rules that inspect a tool call and decide whether it runs. You write a rule as a script, say which calls it covers, and start it in Observing mode to watch what it would have done. Promote it to Blocking when you trust it, and refused calls show up in Activity with the rule and its reason. Rules have drafts, versions and a tester like the rest of your capabilities. Guardrails are off by default and there is no setting to turn them on yourself: ask us to enable them for your workspace.
  • A script can import the Script SDK, now documented: the helpers a TypeScript skill script uses to call tools, delegate, run another script and return a result, and a second entry point a TypeScript or JavaScript guardrail rule imports to rule on a call instead of writing the JSON contract by hand. There is nothing to install and nothing to declare.
  • Your agent can now read your workspace activity itself, through a new analytics tool: call totals, value, sessions and individual calls. It answers at your own scope, so an admin gets the whole workspace and a member gets their own calls. See Ask your agent about activity.
  • The server catalog has a new Database server: a private SQL sandbox for your workspace, with nothing to install and no account to connect. The Mock Data Builder role designs, seeds and queries a sample dataset in it. See MCP servers.
  • A role now has a Graph tab, drawing the role, its skills, their tools and the servers behind them as one picture. See Create a role.

Changed

  • The Map tab in Activity is now a role-centered network: roles in the middle, their skills around them, tools and servers at the edge. Click a node to isolate everything that depends on it. See Dependency map.
  • The Org chart view on the Roles page now drills down one level at a time. Each team shows how many roles sit under it and its 30-day activity, and a team with no role yet is drawn as an Open position you can fill on the spot. See Organize roles in teams.
  • Installing a role from the catalog now picks the only team for you, then offers to install the servers its skills recommend in one click. Servers that need configuration or a connection are left out of that step.
  • The setup callout on a shared skill leads with Review setup, and links every recommended tool at once with Link all. See Import and share.
  • Calls in Activity now name what they targeted, not only the tool that dispatched them: manage · create-skill instead of manage. See Activity.
  • Skill names in a role's Overview now open the skill.

Fixed

  • An action that fails now says so. Some buttons used to look like they had worked when the write had been refused.
  • Publishing, discarding or restoring a role no longer shows all of its skills as enabled until the next reload.
  • A call that two learned skills could both own is no longer credited to both. It stays unattributed instead.
  • The workspace list in the header scrolls once you have more workspaces than fit on screen.
  • Agents holding several connections at once no longer lose them when one of those connections refreshes its token. Clients that register without a client secret can refresh too. See Connect your agent.
  • The enable switch on a disabled role or skill is now offered only to members who are allowed to enable it.

v0.12.4

Added

  • The user menu now has a Notifications row for pending invitations and unread announcements, and a banner appears ahead of planned maintenance windows. The help menu also links to these release notes under What's new.

Changed

  • Connecting an agent to Skilder now always asks you to approve access. The consent screen shows which site will receive the authorization code, and flags any client Skilder doesn't recognize. See Connect your agent.
  • The role page has a new layout: an Overview tab summarizing skills, recent activity and instructions, separate Skills, Instructions, Activity and History tabs, 30-day call and value stats, and a Needs attention tab for open drafts or skills nobody has called recently. See Create a role.

Fixed

  • The token count shown in a call's detail view no longer doubles up. See Activity.
  • Tables in the markdown editor no longer render white when dark mode is on.
  • Buttons show the pointer cursor again when you hover them.
  • Calling a tool right after installing its MCP server no longer hangs or times out. See MCP servers.
  • Autocomplete and type checking for Node.js now work correctly in the script editor. See Build a skill.

v0.12.3

Added

  • In beta, on request. Your agent can now show you an interactive card instead of describing its work in prose: a recap of the skills, roles and MCP servers it just touched, or a skill's or role's draft and publish state with a Publish button you click yourself. It is off by default and there is no setting to turn it on yourself: ask us to enable it for your workspace. You also need an MCP client that renders cards.
  • The docs have a new Architecture section for anyone evaluating Skilder: what the platform is made of, how a tool call runs and stays isolated, and how your data is handled. See Architecture.

Changed

  • Your agent can now replace a skill asset's file in place. The asset keeps its link, so mentions of it elsewhere in the skill keep resolving. Replacing used to mean deleting and re-attaching, which quietly broke them. See Build a skill.

Fixed

  • Tool calls in Activity are no longer credited to the wrong skill when two skills equip the same tool. A call is now attributed only to skills the session actually learned, so a workspace with session tracking switched off gets no skill attribution on those calls at all. See Activity.

v0.12.2

Added

  • Your agent can now push a file into your workspace runtime and pass it straight to a tool, without you uploading it in the browser first.
  • Your agent can now attach a file to a skill as an asset, without opening the skill editor. See Build a skill.

Changed

  • The /get-this-done-right prompt is now /do-this-task-with-skilder, and /get-started takes a job argument instead of role. See Author with AI.
  • Your agent now asks before it creates, updates or deletes anything in your workspace. Reading roles, skills and tools stays unprompted.
  • Signing in from an invitation link now returns you to the invitation instead of the app.
  • The session timeline in Activity refreshes every 10 seconds while you watch a session. Slow it down or switch it off with the refresh control in the timeline header. It stops refreshing once you click Load more.

Fixed

  • Connecting your own OAuth integration no longer signs other members out of the same MCP server. See Personal integrations.
  • Refreshing the page while you name your first workspace no longer drops you into the app under the generated name.
  • An asset upload that fails partway no longer leaves an empty file attached to the skill.

v0.12.1

Added

  • New workspaces start with an Onboarding Concierge role that walks you through a first real task. It replaces the Hello World placeholder. See Getting started.
  • Your MCP client now lists the five Skilder prompts as slash commands, fetched over the protocol.
  • A skill can carry a value per use, shown in the Value section of the skill settings panel.
  • Activity has a new Value tab that rolls those numbers up for the whole workspace.
  • The Drafts page lists every open draft in the workspace, yours and other people's, and publishes several at once. See Publish and version.

Changed

  • Reviewing a draft before publishing now happens in one full-screen Changes surface. It replaces the inline diff in version history.
  • Connecting with a workspace named in the connection URL now grants that workspace directly, with no picker. If you are not a member of it, there is nothing to approve.
  • Agent setups still connecting over stdio are now refused at the handshake instead of accepted.
  • Replacing an asset file now asks for confirmation and names both files.
  • The terms and privacy links now open www.skilder.ai instead of in-app pages.

Fixed

  • Total Calls and the figures beside it no longer stop at 1000 once a workspace passes that many calls.
  • The Reconnect button on an MCP server with an expired token now works. See MCP servers.
  • An MCP server's status no longer goes blank after a reconnect attempt and a page refresh.
  • Discarding a skill draft and editing again no longer shows the discarded description.
  • A long skill Markdown file now scrolls to its end in the read view.
  • Importing a community skill no longer shows a success message when the import failed. See Import and share.
  • Skill content that mentions a file belonging to another skill now resolves it. The mention is highlighted in the editor, and the file is listed when an agent learns the skill.

v0.12.0

Added

  • You can now leave a workspace yourself. See Profile and security.
  • Sign-in got a magic-link fallback: if the 6-digit code doesn't arrive, click the link in the same email instead.

Changed

  • Signing in and creating an account now share one flow at /signin. There's no separate password-based /register anymore.
  • The Library section is now called Capabilities.
  • Agents now connect over Streamable HTTP only. The stdio and SSE transports are no longer supported. See Connect your agent.
  • Disabled skills and roles are now visibly dimmed in the list instead of looking active.
  • Deleting a skill or role now has one clear place to do it, and the confirmation says it's permanent. Removing a skill from a role is separate and easier to find.
  • Creating a role, workspace or team now warns you if the name is already taken.

v0.11.1

Changed

  • Workspace creation moved from the command palette to the header. The command palette has been removed.
  • The runtime and its keys are now provisioned by Skilder for you, based on your plan. They are no longer managed from workspace settings.

v0.11.0

Added

  • Skills and roles are now available through a public REST API at /api/v1: full CRUD, publish, and zip export.
  • API keys can be scoped to specific permissions instead of granting full access.
  • Members can now see their own activity in Activity. Admins keep the workspace-wide view. See Activity.

Changed

  • A workspace now runs on a single runtime for every MCP server and script, chosen once in workspace settings.
  • New API keys require an expiry and use a new format with a checksum.
  • A key's value is shown once, at creation. There is no way to reveal it again: losing it means revoking it and creating a new one.
  • New keys use the SKA prefix. Existing USK keys keep working but are marked deprecated.

Fixed

  • Importing or exporting a skill no longer drops its allowed-tools list.
  • Testing a script from the editor no longer times out early.
  • MCP connections no longer drop from a single missed ping.

On this page