Changelog

​

Upload an OnPrem license file

​

What’s new

• OnPrem license — An organization admin uploads a .license file under Settings → Organization → OnPrem License. One file covers the installation and replaces the current license. The screen shows who it is licensed to and when it expires
• Trial — With no file applied, the installation runs for 30 days. An expired or invalid license is shown on the same screen
• Downloaded workflow service — Set WORKFLOW_DURABILITY=local to keep each run in its own file on a mounted volume, so a restart continues from the last checkpoint. Leave it unset to run in this process. Do not set it together with TEMPORAL_ADDRESS

​

Improvements

• An agent API stays in the MCP app list when its spec has no LLM summary

​

Bring your own OpenAI-compatible model

​

What’s new

• OpenAI-compatible — Under Settings → Organization → Model Settings, add a provider with a display name, a base URL (the /v1 root), an API key, and the model ids it serves. Mark a model that does not accept tools. The API key is stored like the other provider keys and shown as the last 4 characters

​

Install a prepared spec without an AI key

​

What’s new

• Provisioned OpenAPI bundle — A spec file that already includes its summary and agentic metadata installs without an organization model key. The form names the file and asks for the application name
• Zip in chat — Drop or paste a .zip in Chat. Copilot unpacks the text and images inside and attaches those files. A zip larger than 25 MB is refused

​

Improvements

• Agent editor tabs use the full page width, so every tab fits without scrolling

​

Run a downloaded workflow service on Temporal

​

What’s new

• Temporal — The Docker service download includes an optional Temporal setup. Leave the lines blank to run the graph in the service container. For Temporal Cloud, set the address, namespace, API key, and a task queue used only by this service, then start a second process from the same image

​

Continue after a failed step

​

What’s new

• On error — Each step’s Advanced settings choose Stop the run or Continue on the error output. Continue sends the error to a failure output so a later step can handle it

​

Improvements

• A generative UI template opened fullscreen covers the chat composer
• A failed save in the workflow step editor stays in the dialog

​

Pause a workflow for up to 60 seconds

​

What’s new

• Wait — Add a Wait step that pauses the run for 0 to 60 seconds, then continues with the same data. Dry runs do not wait

​

Improvements

• An approval step shows its preview and field keys
• The graph zoom control is a floating minus, percent, and plus

​

Pick workflow steps from the agent

​

What’s new

• Action — Pick the application and operation from this agent. Parameters are limited to what that action accepts
• Call workflow — The step lists the called workflow’s inputs, and its output follows that workflow’s Result
• Pick a path — Add and remove cases, each with its own outlet, and rename a case in place
• Condition, Filter, and Switch — Add and remove conditions on the rule
• Show UI — The closed template picker shows the template id

​

Approve a paused workflow API

​

What’s new

• Test — A run that is waiting shows the approval title, message, and preview. Approve or Reject. A reason is optional. Edits is a JSON object merged into the run when you approve
• Run status — Callers can read a run by its id and see the status, result, and approval payload while it is waiting
• Resume — Callers continue that run with approve or reject, including a reason and edits

​

Workflow service and code-step fixes

​

Improvements

• Docker service — The downloaded service is built for linux/amd64, the platform the runtime image is published for, so the image starts
• Request body — A call to the standalone service sends the same JSON body as Test in Copilot
• API description — A published workflow API description can be up to 1000 characters
• Code — A Python date or datetime in result comes back as an ISO-8601 string. A failure names the script line. Raising Exception, ValueError, or TypeError fails the step
• Expressions — Workflow help is split into sections, and each rule has a worked example
• Workflow list — The agent table shows each workflow id. A cut-off name or id shows the full value on hover

​

Call another workflow

​

What’s new

• Sub-workflow — Pick a saved workflow on this agent. The step waits for it and passes its result forward. The called workflow has to finish in that step. One that stops for approval fails the step

​

Improvements

• Add step — Every step type is in one searchable menu, sorted by name
• Action parameters — On an Action step, add, rename, and remove parameters the same way as Input fields

​

Date & time, crypto, data table, and expressions

​

What’s new

• Date & time — Format a UTC timestamp, or add, subtract, and diff days, hours, and minutes
• Crypto — Hash or HMAC with md5, sha256, sha384, or sha512, as hex or base64. Or generate a uuid, hex, base64, or ascii string
• Data table — Insert, get, upsert, or delete rows for later steps in the same process. The first insert creates the table. Rows are cleared when the process restarts, and a dry run does not write
• Expressions — A field that is only ={{ }} keeps the result’s type. Read $json, this run’s inputs, another step by id or by its name, and the current UTC time. Helpers include if, find, and filterBy
• If and code — An If step accepts a list of conditions. A JavaScript code step can return its value. A connection can use the step’s display name

​

Code step and version compare

​

What’s new

• Code — A workflow can run a Python or JavaScript script. Assign result, and that JSON value is the next step’s data. The editor highlights the language, shows syntax errors, offers a starter snippet, and keeps the last run’s output. Python provides json, math, re, date, datetime, timedelta, and print. JavaScript provides the language and console.log. Scripts cannot import packages. The timeout is 1 to 30 seconds, and the default is 5
• Compare versions — Version compare is a collapsible diff. It includes edits inside workflows, skills, templates, and contexts. Save HTML downloads the comparison

​

Improvements

• Example values are generated for tool inputs on every model

​

HTML app templates

​

What’s new

• HTML app — Author the document in a full-height code pane and preview it at desktop and mobile width. New templates start as an HTML app
• Tool calls — A tool call from the app re-renders the page with the tool result

​

Sign in from the downloaded workflow service

​

What’s new

• MCP OAuth — A workflow service downloaded from Versions can complete the MCP OAuth 2.1 sign-in on its own

​

Improvements

• Tool results — JSON text returned by an MCP tool is parsed into data for the next step
• runtime.env — The example env is filled in, so a local call runs immediately, and the download steps are easier to follow
• Design Agent · Extend — Opening Extend selects the agent you named

​

Try an agent from its card

​

What’s new

• Try Now — On a connected agent card, Try Now opens that agent in chat, ready to use

​

Security updates

​

What’s new

• MCP gateway — Fixed a critical IP-spoofing issue in request forwarding, and updated the MCP SDK and related HTTP libraries
• Copilot — Updated Next.js, image processing, and source-map handling
• Workflow runtime — Updated the image’s packaging tools
• API image — The compiler’s kernel headers are removed after the image is built

​

Download a workflow Docker service

​

What’s new

• Docker service — On Versions, Docker service downloads a zip of the published workflow APIs you select. The file name describes what it contains
• Run steps — After the download, copy three steps: set SERVICE_API_KEY in runtime.env (callers send it as a bearer token), build and run, then open http://localhost:8080/docs and authorize with that key

​

Workflow step timing

​

What’s new

• Duration — The step inspector shows how long each node ran
• Design Agent — Starting Design Agent from a chat pastes that agent’s handoff instruction into the composer

​

Improvements

• View and edit share one Input, Logic, and Output panel, so a step reads the same way in both modes
• Workflow help matches the canvas: it no longer lists an End marker

​

Publish a workflow as an API

​

What’s new

• Workflow API — On a workflow, open API (Configure OpenAPI). Publish it as an authenticated endpoint, set the HTTP method and path, and view or download the OpenAPI document. POST and PUT send inputs as a JSON body. GET and DELETE send them as query parameters. Test runs the endpoint from the dialog. The URL uses the agent slug. Unpublish sits beside close
• Microsoft Copilot Studio — On the agent Setup tab, choose Microsoft Copilot Studio. Enable generative orchestration, add an MCP tool, paste the MCP URL, then connect with OAuth 2.0 or an API key and publish the agent

​

Improvements

• MCP OAuth authorize links keep query parameters that were already on the URL

​

OpenLegacy Integration import

​

Improvements

• Add application labels the Hub import OpenLegacy Integration
• Chat follow-up buttons that need a sign-in show Login to

​

Install SOAP services from a WSDL

​

What’s new

• From SOAP — On Applications, click Add application and choose From SOAP. Paste a WSDL URL or upload a .wsdl or .xml file. Apigene converts that service to OpenAPI and runs the usual install and index
• SOAP port — If the WSDL lists more than one port, pick the port (SOAP version and operations are shown) before Install. A WSDL with a single port installs that port
• Authentication — Optionally record HTTP Basic, Bearer token, or API key (with a header name). The secret is entered on the application Security tab after install

​

Improvements

• Follow-up suggestion chips are translated, and the arrow points forward in right-to-left languages

​

Agent editor tab bar

​

Subagents in chat

​

What’s new

• Subagents — In Live chat, open Select operation mode and turn on Subagents. The agent can delegate a self-contained task — for example a batch of large tool results — to a subagent with a fresh context. Several can run at once. Each run is a card with status, task, answer, and model. The switch is off while Manual approval is on, and Design Agent Create and Extend do not use subagents
• Fast model — A subagent can run on a cheaper model from the same provider (Claude to Haiku, GPT to mini, Gemini to Flash, Grok to mini). The card marks that run Fast
• Compare — When model comparison is on, the candidate has its own Subagents switch
• Chat Info — Token usage is split by model and includes subagent runs
• Model catalog — Built-in pickers add Claude Haiku 5.5

​

Connect an agent to AWS AgentCore

​

What’s new

• AWS AgentCore — On the agent Setup tab, choose AWS AgentCore under MCP Client. Install the AgentCore CLI, open or create a project, generate an API token, and run the agentcore add harness command. Then agentcore deploy and agentcore invoke. Claude Sonnet 5.5 must be enabled for the AWS account in Amazon Bedrock
• The project name and harness name together must stay within 40 characters

​

Improvements

• run_action_ui includes the template HTML, so ChatGPT can render any generative UI template

​

Classify workflow step and Kiro install

​

What’s new

• Classify — Workflows can include an apigene.classify step. It sends a payload and named questions (choice, score, or yes/no) to TypeSafe Jev and returns typed answers with probabilities. Branch on those answers with If or Switch. Inspect questions and answers on the graph; author the step from Copilot or MCP
• TypeSafe Jev — Admins add the key under Settings → Organization → Model Settings → TypeSafe Jev (Workflow Classify). Classify steps use that key. It is not a chat model. Dry run does not call TypeSafe
• AWS Kiro — On the agent Setup tab, choose AWS Kiro. Use Quick install with Kiro, or paste the MCP config into ~/.kiro/settings/mcp.json

​

Improvements

• MCP apps advertise picture-in-picture, so hosts that support it can offer that layout

​

OL Terminal in chat

​

What’s new

• OL Terminal — When a chat tool returns a host screen, OL Terminal opens as a floating window (it starts minimized). Drag it, resize from any corner, go full screen, or dock beside the chat. Reopen it from the OL Terminal pill
• Type on the screen — Click fields, move the 3270 cursor, and press keys. The next message tells the agent what you did by hand
• Ask OL Gene — Ask about the screen, a field, or a selected value without leaving the window
• Send to an app — Hand a value or the screen to another application on the agent. Copilot proposes the action and waits for your OK before anything is written
• Sign-on — Log in when the terminal app needs a session, and log out of the mainframe. A dropped session keeps the last screen and shows Logged out; Log in · new session starts again with your saved sign-on
• Model catalog — Built-in pickers add GPT-6.1 Sol. Claude Sonnet 5 is replaced by Sonnet 5.5

​

Improvements

• Host screens use OL Terminal colors
• The chat header overlays the thread so the conversation uses more of the window
• Mock Mode starter prompts use ids from the generated mock data, so Try in chat opens on a real record

​

Agent zip export and import, and Design Agent Mock Mode

​

What’s new

• Agent bundle — On Versions, Export .zip downloads the draft (instructions, skills, contexts, OpenAPI apps, MCP servers, workflows, and UI templates — not credentials). A published version can be downloaded from its row menu. On Agents, Import uploads a bundle: missing apps are created and indexed; existing apps are reused as they are. If the agent name already exists, update the draft (owners only) or create a copy
• Mock Mode — In Design Agent, choose Mock Mode (alongside Create and Extend). Describe a system that does not exist yet; Copilot provisions a mock application, builds an agent on it, and offers Try in chat. The session is available when a mock service is configured
• Model catalog — Built-in pickers include GPT-6 (Astra, Sol, Luna), Claude Opus 5.5 / Fable 5.1 / Sonnet 5, Grok 4.7, and Gemini 3.8 Flash

​

Improvements

• Light-mode status and live/draft badge contrast
• Version compare reads older to newer
• Use case proposal cards fit a full idea on screen, with app icons on chips
• Tool calls show agent display names; RTL chevrons flip correctly
• Settings section cards match application tabs
• Marketplace Try Now waits for spec reindex before creating the agent
• Switching Live / Design refreshes home when the thread was started there
• Mock Mode setup shows elapsed time

​

Live and Design workspaces, and Try Now with demo apps

​

What’s new

• Live / Design — Toggle Live and Design in the sidebar. Live is chat, Tasks, and personal settings. Design groups Build (Agents, Applications, Skills, Context), Monitor (Actions, Workflows, Insights), and Admin. Members and phones stay in Live
• Try Now — On a marketplace plugin, Try Now installs (or reuses) the agent, can prepare recommended demo applications, connects the agent to Copilot, and opens chat in Live. First-time setup shows a progress sheet; click the spinning button to reopen it
• Generative UI in chat — Tool groups that render a UI stay expanded; workflow runs that show generative UI are not folded into the collapsed group

​

Improvements

• Light theme contrast
• On mobile, the composer stays in view and mention/model pickers open as sheets
• Multi-line paste no longer doubles newlines
• Organization members can use models configured in org settings
• Failed API calls keep request and response data in Actions (when recording is on)

​

Localization

​

Side-by-side model comparison

​

What’s new

• Model comparison — Admins enable it under Settings → Organization → Advanced → Model comparison
• Compare — In chat, pick a second model and send. Chat as many turns as you need, then conclude for a review. Each comparison doubles token spend and may execute tools twice

​

Agent Marketplace and sidebar navigation

​

What’s new

• Agent Marketplace — On Agents, browse and install agents from a Claude-plugin marketplace repo. Admins point the gallery at GitHub under Settings → Organization → Agent Marketplace
• Sidebar — Agents, Tasks, Applications, Skills, and Context are in the left navigation (not only the app menu)

​

Improvements

• Shared agents stay read-only for non-creators

​

Design Agent Create and Extend

​

What’s new

• Create — Define a new agent’s instructions, apps, and behavior
• Extend — Add a skill, workflow, or UI to the draft; Publish stays explicit
• Draft chip — Extend shows Draft as soon as a skill, workflow, or UI is added

​

Improvements

• Create workflows from Copilot or MCP. The New workflow form is no longer on the agent Workflows tab — open a workflow there to Edit the graph

​

Draft and live agent versions

​

What’s new

• Draft vs live — Save updates the draft. Chat and the runtime MCP URL keep using the published live version
• Publish — On the agent’s Versions tab, publish the draft (optional note) to create a numbered version and point live at it
• Version history — Compare draft to live, inspect a snapshot, restore a version into draft, or switch live to an earlier version (the MCP URL stays the same)
• Live and Draft chips — Agent cards and the editor show the live version (vN) and a Draft chip when unpublished changes exist
• Design MCP — https://app.apigene.ai/agent/{name}/design/mcp reads and writes the draft (skills, workflows, templates). The usual /agent/{name}/mcp URL stays on live
• Owner-only edits — Viewers of a shared agent can use the live agent but do not see design or mutation tools

​

Improvements

• Agents list is no longer limited to the first 50 agents

​

Native MCP login/logout and OS-aware shortcuts

​

What’s new

• Login / Logout — On the MCP dashboard, server detail, or @mention picker, complete OAuth in a popup or log out to switch accounts. Tokens are per user and apply to Copilot and the MCP gateway
• OS-aware shortcuts — Shortcut labels show Ctrl on Windows and ⌘ on macOS

​

Improvements

• Native MCP tool lists follow pagination so servers with many tools are fully cached

​

Paste screenshots into chat

​

What’s new

• Paste screenshot — With the chat input focused, press ⌘V / Ctrl+V to attach a clipboard image (screen capture, copied image, or a file copied from Finder / Explorer)

​

Mention picker columns, skill zip upload, and workflow help

​

What’s new

• Mention picker columns — Type @ to browse Agents, MCP, and Tools in separate columns
• Skill folder zip — Upload Skill accepts a .zip of an Agent Skills folder (SKILL.md at the root or in a wrapper folder), not only a single SKILL.md
• Workflow help — Open in-app help from the graph or a run for node types, expressions, and limits
• Extra HTTP headers — Admins add per-provider request headers under Settings → Organization → Model Settings (for example anthropic-workspace-id)

​

Improvements

• Hide or show the workflow mini map; clearer node inspector
• Agent Builder shortcut on the Agents page appears when creating a new agent
• Chat Agents list updates when an agent is renamed or deleted

​

Visual workflow editing, URL-safe agent names, and host-screen Actions

​

What’s new

• Visual workflow editor — On an agent’s Workflows tab, open a workflow and click Edit to add, connect, and inspect steps on the graph. Save compiles and stores the draft; View leaves edit mode and discards unsaved changes
• New workflow — Create a named workflow from the list (starts with a Set values step). You can still author or update graphs from Copilot chat or MCP
• URL-safe agent names — New agent names are lowercase kebab-case (support-agent). Spaces and underscores become hyphens as you type
• Terminal response tab — When an action response includes a host/terminal screen (screenText), Actions and shared action views show a Terminal tab with the rendered screen

​

Improvements

• Agent and MCP pickers in chat show connection status, sort more consistently, and keep the MCP status panel above the tools list
• Mermaid diagrams in chat auto-fit to the message so large charts stay readable
• Postman collection import reports clearer conversion errors (including collections with comments or stray quotes)
• Insights user charts merge duplicate accounts that share the same email

​

Install Hub projects as apps, Hub auth options, and Builder Hub skills

​

What’s new

• Install from OpenLegacy Hub — On Applications → Install Application, choose OpenLegacy Hub, search projects, and install a project’s OpenAPI as an application (shows method count, last updated, and an Installed badge)
• Hub authentication — Under Settings → Organization → Open Legacy, choose API Key (x-api-key) or Bearer token (Authorization: Bearer)
• Test connection — Verify the Hub URL and credential before saving; unsaved form values are used, and last-4 credentials reuse the stored secret
• Hub skills in Agent Builder — When you accept a use case that includes Hub projects, Builder attaches the ol-hub app and provisions a project-scoped skill with contract, method, and flow IDs so the agent can run Hub test flows without a discovery loop

​

Improvements

• Hub project picker in Agent Builder includes contract methods as context for use-case proposals
• Clearer Hub connection errors when the server URL or credential cannot reach Hub

​

OpenLegacy Hub in Agent Builder, scoped apps, and chat history fixes

​

What’s new

• OpenLegacy Hub — Under Settings → Organization → Open Legacy, enable the Hub, set the Hub Server URL and API key, and optionally Use in Agent Builder
• Hub projects in Agent Builder — Hub projects appear in the Apps picker; selecting a project includes contract methods as read-only context for use-case proposals
• Builder app scope — Limit Agent Builder recommendations to the apps (and Hub projects) you selected for that thread
• Expandable use-case summaries — Open a proposal on the card to read the full summary without leaving chat

​

Improvements

• Historic threads restore Copilot vs Agent Builder mode without a UI flash
• Sidebar chat history updates when a thread is renamed or the list changes
• Clearer agent delete confirmation (localized title and button)
• Organization settings sections use consistent icons and borders

​

Collapsed tool calls in chat

​

Improvements

• Collapsed tool calls — Consecutive tool invocations in a message collapse into a compact group; expand to inspect an individual call
• Same collapse behavior in public/shared chat views

​

Outbound proxy, OpenRouter/Bedrock model control, and application filtering

​

What’s new

• Outbound HTTP proxy — Admins set an org-wide proxy URL under Settings → Organization → Outbound network; all outbound API calls from integrations and actions use it (leave empty for direct egress)
• OpenRouter model visibility — Enable or disable individual OpenRouter models in org settings, including models fetched live from your OpenRouter account
• Bedrock model visibility — Same per-model toggles for Amazon Bedrock; the model list is loaded from your AWS account instead of a fixed catalog
• Metadata filters — On an application’s Metadata tab, filter agentic metadata by type and priority rating (1–10) in addition to search
• Operations filters — On Operations, filter by HTTP method, path prefix, and tags when managing large OpenAPI specs

​

Improvements

• Agent Builder validation — Creating an agent requires both a description and instructions; name (64), description (300), and instructions (4,000) character limits match the API
• Insights date ranges — Custom start/end dates in Insights include the full selected UTC days (fixes edge cases with absolute date windows)

​

Insights exports, chat drag-and-drop, and a cleaner sidebar

​

What’s new

• Custom date ranges in Insights — Choose an absolute start and end date (not only presets); Agents analytics respects the same window
• Chart export — Download Insights charts as CSV, PNG, or SVG, with summary stats on chart cards
• Drag-and-drop attachments — Drop files onto the chat input to attach them
• Collapsible sidebar — Collapse the nav to icon mode; chat thread groups can expand/collapse and remember your preference
• Richer image generation — Generate images with OpenAI GPT image models and DALL·E options (model, size, and quality)

​

Improvements

• Built-in tools cleanup — Unused default tools (NanoBanana image, 3D/animation, HTTP, web content) are hidden from @ mentions and workflows
• Cleaner chat display — Mention prefixes are stripped from user messages when viewing and copying
• Smoother navigation — Light transition when moving between pages; consistent page headers across the app
• Deep-linkable tabs — Agent editor and Insights tabs stay in sync with the URL
• Workflow inspector — Clearer node summaries without duplicated metadata

​

Per-user credentials, unified Insights, and Agent Builder in the nav

​

What’s new

• Per-user app credentials — On an application’s Security tab, set Credential scope to Per-user (Basic / Bearer / API key). Each person connects their own secret via Connect / Clear credentials (or the connect-credentials link from chat)
• Org-shared vs per-user — Keep org-wide shared secrets when you want one key for everyone; switch to per-user when each member must use their own
• Agents tab in Insights — Org rollup of agent performance (volume, success rate, duration, errors) plus drill-down into a single agent
• Agent Builder in navigation — Open Agent Builder from the main sidebar and user menu (next to New Chat)

​

Improvements

• Insights replaces Dashboard — Conversations, Workflows, Actions, and Agents analytics live under Insights (the separate Dashboard page is gone)
• RTL-aware chat — Hebrew and Arabic messages, chips, and input align correctly (auto LTR/RTL)
• Reasoning duration — Elapsed time shown while the model is thinking
• Chat polish — Message actions appear on hover; clearer model-selector loading; Chat / Agent Builder mode control stays in a consistent place

​

LLM cost dashboard, model visibility, and Exa web search

​

What’s new

• LLM cost on Dashboard — Conversations tab shows estimated spend over time, by model, by provider, and by user (toggle cost vs tokens)
• Workflow analytics on Dashboard — Workflows tab with KPIs, runs over time, success rate, duration, failures by step, and breakdowns by agent / user / workflow
• Date range presets — Quick ranges on the Dashboard date picker
• Model visibility — Per-provider Models menu in org settings to show or hide models in the Copilot selector
• Custom model IDs — Add provider-specific model IDs so they appear alongside built-in options
• Exa (Web Search) — Configure an Exa API key under organization model providers for web search / content tools
• OnPrem license apply — Self-hosted admins paste a signed license token in Settings → Organization (POST /api/org/apply_license)

​

Improvements

• Clearer trial and license messaging for self-hosted installs

​

Map loops, richer expressions, and workflow polish

​

What’s new

• apigene.map loop / done — Loop body on the first output; done arm receives the collected items array (n8n-style Loop Over Items)
• Nullish coalescing (??) — Prefer defaults only when a value is null (unlike ||, which also skips 0 / "" / false)
• find and filter_by — Look up or filter array objects by field in workflow expressions
• Flatter set assignments — Assignment lists accept a simpler flat array shape when authoring workflows

​

Improvements

• Clearer map edge labels and inspector hints in the workflow graph
• Light-theme consistency across Copilot UI
• More reliable simulation step indexing and feedback

​

Workflow History and run simulation replay

​

What’s new

• Workflows history — New Workflows nav item with org-wide run history (admins see all; members see their own)
• Run detail — Inputs, step traces, and result with copy/share; default Simulation opens the graph replay for the recorded run
• Activity timeline — Last-7-days histogram by run status with brush-to-filter (same pattern as Actions)
• Workflow Recording — Organization setting: Full, Headers only, or Disabled (HITL approval checkpoints still persist when disabled so resume works)
• Share links — Share a run via /workflow-history/share/{run_id} with the same detail + simulation UI

​

Improvements

• Real non-dry-run executions are persisted when a workflow id is present (not only approval waits)
• Display names for agent and workflow in the history table and detail view
• Dry runs are never written to history

​

Workflow dry-run, outputs, and richer run insights

​

What’s new

• Dry-run / simulation — Run a workflow without side effects to validate logic, see step status on the graph, and review inputs and outputs in the inspector
• Output node (apigene.output) — Define the final structured result of a workflow run
• On-error handling — Configure how workflows respond when a step fails
• Richer expressions — Inline object and array literals and improved nesting in workflow expressions
• MCP — run_workflow supports dry_run and optional step data in results

​

Improvements

• Copy and download workflow run results as JSON
• Step tracing, elapsed time, and result size in the workflows UI
• Resizable workflow inspector and clearer run status visualization

​

Approvals, LLM steps, and parallel joins

​

What’s new

• Human-in-the-loop approvals — Pause a run for review, then approve or reject and resume from Copilot or MCP (resume_workflow)
• LLM node (apigene.llm) — Call a language model as a first-class workflow step
• Join node (apigene.join) — Wait for parallel branches and continue with combined results
• Richer conditions — More string operators for if / expression logic

​

Improvements

• Clearer graph UI for approved, rejected, and waiting nodes
• Better workflow compile and lint error messages

​

Agent workflows and chat pinning

​

What’s new

• Workflows on agents — Create and manage workflows from a dedicated Workflows tab
• Workflow graph editor — Design steps visually with improved layout and node tooling
• Workflow linting — Catch compile issues early with actionable hints
• MCP workflow tools — Author, run, and manage workflows from MCP clients
• Pin chat threads — Pin important conversations in the sidebar
• Dashboard analytics — Conversation analytics on the dashboard

​

Improvements

• Clearer message timestamps in chat
• More reliable interactive / streaming app handling in chat

​

Chat layout fix

​

Fixes

• Empty chat is centered again (layout no longer stretches when there are no messages)

​

Interactive apps in chat and safer secrets

​

What’s new

• Interactive apps in chat — Richer in-conversation app experiences alongside assistant replies

​

Improvements

• Interaction history masks sensitive values
• Application / spec APIs no longer return stored security secrets in list and update responses

​

Tasks, chat search, and app secrets

​

What’s new

• Tasks — Create, schedule, and run tasks from the product navigation and task management UI
• Model per task — Choose which chat model a task uses
• Thread search and filters — Search and filter conversations in the sidebar
• Provider secret management — Clearer handling of secrets for application / provider configuration

​

Improvements

• Faster chat loading and paginated thread history
• “Run now” for tasks no longer times out behind long-running proxies
• Task names must be unique within the organization
• MCP lists can include tool details when needed

​

Agent skills download and clearer errors

​

What’s new

• Download agent skills — Export an agent’s skill harness as a downloadable package from the agent editor

​

Improvements

• Friendlier error screens with Try again and Go home instead of the default crash page

​

Chat file handling and prompt input

​

Improvements

• More reliable file handling in chat and the prompt input
• More responsive prompt input layout on smaller screens

​

Use-case proposals and app server URL

​

What’s new

• Use-case proposals in chat — Prompt input can surface suggested use cases to get started faster
• Server URL selection — Choose and validate the application server URL in app settings

​

Improvements

• Smoother loading behavior on the audit logs page

​

Chat commands, follow-ups, and audit coverage

​

What’s new

• Chat commands (prompt presets) — Save, pick, and reuse prompts from the chat input, including agent selection
• Follow-up suggestions — Toggle suggested follow-up prompts in chat settings
• Personal settings snapshots — Manage settings snapshots from personal settings
• Broader audit logging — Audit entries for API token generation and organization settings changes

​

Improvements

• Interaction / action views handle redacted request bodies safely
• Organizations can control when interaction bodies are recorded
• Prompt commands are sorted by name for easier browsing

​

Chat message queue, skills download, and clearer model errors

​

What’s new

• Message queue — Queue follow-up messages while a chat response is still running
• Download SKILL.md — Export skill documentation from the skills table
• More file types in chat — Broader file support in the prompt input
• LLM checks on install — Installing an application checks that a model provider is configured

​

Improvements

• Clearer Bedrock chat errors and filtering of unavailable / end-of-life models
• Agent icons display correctly in the tool picker and selected-agent list
• Better reindexing status feedback on application metadata

​

Keycloak authentication

​

What’s new

• Keycloak auth provider — Sign in with Keycloak for deployments that use enterprise IdP authentication

​

Reasoning support and new models

​

What’s new

• Model reasoning — Copilot supports generic model reasoning with configurable reasoning effort
• Fable 5 — Fable 5 available as a chat model option

​

Mock execution for API actions

​

What’s new

• Mock execution mode — Generate mock API responses for interactions when mock mode is enabled
• Optional LLM fallback for mocks — Prefer schema-based mocks by default; optionally fall back to an LLM when examples are insufficient
• Force reindex — Force an application reindex when refreshing agentic metadata

​

Skills upload UX and branding polish

​

Improvements

• Tabbed skill upload dialog with clearer file mapping
• More consistent branding and localization across Copilot

​

Apigene 5.0 — org models, skill registry, and self-serve auth

​

What’s new

• Org-scoped model credentials — LLM keys come from Settings → Models per organization (no shared global OPENAI_API_KEY on the API)
• Amazon Bedrock & Microsoft Foundry — Configure Bedrock (including API-key auth) and Azure / Microsoft Foundry providers with clearer validation
• Skill registry — Find and install skills from the skill registry in Agent Builder
• Apigene authentication — Native Apigene sign-in adapter and sign-in UI (alongside existing providers)
• Self-service organizations — Admins can create an organization during signup
• GPT-5.5 — GPT-5.5 available in the model list

​

Improvements

• Searchable model picker with collapsible providers
• Clearer organization settings layout with sticky save for unsaved changes