Skip to content

Releases: i-am-bee/agentstack

v0.5.2

20 Jan 13:01
@tomkis tomkis
v0.5.2
cd5b5c0

Choose a tag to compare

View all tags
v0.5.2 Latest
Latest

🚀 Agent Stack version 0.5.2 has been released

This release brings a major TypeScript SDK restructuring, a new Canvas agent, comprehensive UI redesign, and significant improvements to authentication and CLI experience.

Major Changes

Breaking: TypeScript SDK Restructuring

The agentstack-sdk-ts has been completely refactored with a new modular architecture. The API client is now organized into dedicated modules (extensions, api, core) with proper Zod schemas and types for every extension and API method. A new buildApiClient core function with unwrapResult utility provides standardized response handling. Error handling is now structured with ApiErrorException and specific error types (Http, Network, Parse, Validation). All consumers of the TS SDK need to update imports and usage patterns.

New Canvas Agent

A new agent for multi-turn artifact editing is now available. Users can select and modify specific sections of text content, enabling precise iterative refinement of generated artifacts.

GUI Shell Redesign

The UI has been completely redesigned with a new navigation structure. New dedicated pages for agent details (showing description, docs, authors, contributors, tools) and agent-specific settings are now available.

SDK: User Approval Extension

New ApprovalExtensionServer and ApprovalExtensionClient enable explicit human-in-the-loop workflows. Agents can request user approval for critical actions using structured ApprovalRequest/ApprovalResponse models. The older ToolCallRequest and ToolCallExtensionServer are now deprecated.

Agent Authorization via Context Tokens

A significant security improvement to how agents authenticate and communicate:

  • Single token authentication: A2A clients can now use the platform context token directly in the Authorization header, eliminating the previous two-token approach (user OIDC token + context token in extension fulfillment)
  • Token exchange for agent-to-agent calls: When the A2A proxy forwards requests to agents, it performs token exchange - creating a new token with the same claims and expiration but with a specific aud (audience) claim targeting that
    agent's URL
  • Audience validation: Agents using PlatformAuthBackend validate that incoming tokens have the correct audience. This prevents tokens issued for one agent from being used to call other agents directly
  • Security fix: Closes a vulnerability where a malicious agent could bypass the proxy and directly call other agents, potentially accessing other users' data via task/list or task/get A2A methods
  • RSA signing: JWT signing upgraded from symmetric HS256 to asymmetric RS256. A new /.well-known/jwks endpoint exposes the public key for signature verification
  • Granular permissions: The a2a_proxy permission can now be scoped to specific provider UUIDs instead of just *

For SDK users: Add auth_backend=PlatformAuthBackend() when calling Server.serve() to enable platform authentication on managed agents. For Helm deployments: Configure jwt_private_key and jwt_public_key in auth settings
(auto-generated if not provided).

CLI Improvements

  • The agentstack remove command now features interactive multi-selection for agent deletion with a new --all flag for bulk removal.
  • Better error messages for 403 Forbidden responses guide users to contact administrators.
  • WSL2 mirrored networking mode is dropped in favor of NAT mode with dynamic host IP resolution and updated documentation for Ollama/local agent configuration.

API: Delete Messages Endpoint

New endpoint for deleting messages from a selected ID onward, enabling conversation history management.

What's changed

Assets 2
Loading

v0.5.0

07 Jan 14:27
@tomkis tomkis
v0.5.0
6a5ff8a

Choose a tag to compare

View all tags
v0.5.0

Major Features

Canvas Support: Introduced initial Canvas functionality to the UI, enabling agents to create and edit rich interactive content directly within chat conversations. Users can select text in Canvas artifacts and submit edit requests to agents, unlocking new use cases like collaborative document editing and code generation. The feature includes SDK updates for both Python and TypeScript with new examples for code and recipe agents. Note: This includes a breaking change in the TypeScript SDK where handleInputRequired was renamed to resolveUserMetadata and InputRequiredResponses to UserMetadataInputs. (#1667)

Permission Elevation System: Added comprehensive role management capabilities with a role versioning mechanism that immediately invalidates access tokens when user roles change. Administrators can now elevate user permissions via a new API endpoint (PUT /users/{user_id}/role), and the CLI gained a agentstack user command for listing users and managing their roles. This ensures proper security boundaries as agents are deployed to production. (#1740)

Interactive CLI Experience: The CLI now supports interactive modes for agentstack add, agentstack update, and agentstack run commands. When arguments are omitted, users are presented with fuzzy search interfaces to select agents or GitHub repository tags, making agent management significantly more intuitive for developers who prefer guided workflows. (#1682, #1687)

MCP Auto-Injection in UI: Connected MCP servers are now automatically injected and resolved when agents request them through the UI. This removes the manual textbox approach and consolidates MCP documentation, providing a seamless experience for agents that require Model Context Protocol integrations. A new GitHub MCP agent example demonstrates the capability. (#1695)

Connector Management: Added comprehensive connector UI and presets system. Users can browse, connect, and disconnect predefined connector presets without manually entering configuration details. The system intelligently creates new connectors on first connection and provides table-based management with search functionality. (#1702, #1718)

Rate Limiting: Implemented global and per-role rate limiting for the server. Administrators can configure limits based on user roles (user, developer, admin) for specific operations like OpenAI chat completion tokens, requests, and embedding inputs. Supports both Redis and in-memory backends with configurable strategies, protecting production deployments from abuse. (#1686, #1698)

SDK Improvements

Platform SDK in UI: Migrated the UI to use the TypeScript Client SDK for platform operations instead of direct API calls via openapi-fetch. The SDK now supports custom fetch implementations for injection of authenticated wrappers, establishing the UI as a reference implementation. (#1691)

Exception Handling: Improved Python SDK's error handling during extension initialization. Exceptions during dependency setup are now properly captured and grouped into ExceptionGroup for comprehensive reporting, preventing silent failures. (#1710)

Dependency Reduction: Removed the email-validator dependency from Python SDK by changing User.email from EmailStr to str, reducing external package requirements. (#1708)

JSON Text Extraction: Added support for JSON format output from the docling text extraction service via an extras parameter, enabling structured data extraction alongside the existing Markdown format. (#1652)

CLI Features

Server Version Display: The agentstack self version command now shows the remote server version when connected to an Agent Stack server, fetched directly from the server's OpenAPI schema rather than local Helm values. (#1728)

Agent Feedback Viewing: Added agentstack agent feedback list command to view user feedback for agents directly from the terminal, with support for filtering by agent and pagination. (#1732)

Platform Files Support: The --dump-files flag now properly supports platform files in addition to agent files. (#1760)

UI Enhancements

Agent Management Behind Feature Flag: Agent management UI has been moved behind the Providers feature flag to reduce confusion and keep the interface focused on end-user actions, with CLI remaining the primary method for agent management. (#1625)

Improved Mermaid Rendering: Stabilized mermaid diagram rendering during streaming to prevent visual flickering. (#1676)

Missing Model Error Page: Added dedicated error page when required embedding models are not configured. (#1701)

What's Changed

Loading

v0.4.3

05 Dec 10:16
@tomkis tomkis
v0.4.3
10ef5b8

Choose a tag to compare

View all tags
v0.4.3

What's Changed

Full Changelog: v0.4.2...v0.4.3

Contributors

JanPokorny, tomkis, and 4 other contributors
Loading

v0.4.2

01 Dec 15:02
@tomkis tomkis
v0.4.2
699536a

Choose a tag to compare

View all tags
v0.4.2

🚀 Major Changes

This release brings major improvements to the CLI experience and web UI, a redesigned form extension architecture, plus many platform and SDK enhancements.

CLI Improvements - The CLI has been redesigned with organized help text, improved agentstack list formatting with status indicators, and agentstack build now builds on the server by default - removing the need for Docker to be installed locally.

UI Enhancements - Agent execution steps (trajectory) are now grouped together for easier reading. Both agent responses and trajectory now display Mermaid diagrams, LaTeX math, and images as rendered content instead of raw markdown code.

Breaking Change: Form Extension - The Form Extension is now split into two separate extensions: FormServiceExtension for initial forms (shown before conversation starts) and FormRequestExtension for dynamic forms (requested during conversation). This prevents initial setup forms from being overridden by runtime forms. Agents using forms need to be updated (See docs).

What's Changed

Read more

Contributors

abn, JanPokorny, and 11 other contributors
Loading

v0.4.1

11 Nov 11:09
@tomkis tomkis
v0.4.1
9caf1f9

Choose a tag to compare

View all tags
v0.4.1

What's Changed

Contributors

tomkis, matoushavlena, and 6 other contributors
Loading

v0.4.0

04 Nov 08:45
@tomkis tomkis
v0.4.0
ec364d6
This commit was signed with the committer’s verified signature.
jezekra1 Radek Ježek
SSH Key Fingerprint: WY/4hPdpbLCS3enpAvW/slfDB27VN4vXs1Sbg7vrwHo
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.4.0

🚀 Major Changes

Providers build straight from GitHub

You can now build and deploy providers directly from GitHub repositories. This makes it easier to integrate, test, and iterate on your agent code with minimal setup.

New Trajectory UX

A major rework of the Trajectory visualization and UX. The new design improves clarity of agent reasoning and tool calls.

Connectors API

The new Connectors API introduces a standardized OAuth-based system for integrating third-party services; such as Box, Google Drive etc.

Provider Versioning & Self-Registration

Providers now support automatic versioning and self-registration, simplifying deployment pipelines and ensuring consistent compatibility tracking across environments.

Full Changelog

Full Changelog: v0.3.7...v0.4.0

Contributors

JanPokorny, tomkis, and 9 other contributors
Loading

v0.3.7

15 Oct 10:44
@tomkis tomkis
v0.3.7
e04cbda
This commit was signed with the committer’s verified signature.
jezekra1 Radek Ježek
SSH Key Fingerprint: WY/4hPdpbLCS3enpAvW/slfDB27VN4vXs1Sbg7vrwHo
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.3.7

Major Changes

Session History

Conversations can now be persisted with full context. View past sessions in the left nav, resume conversations, and delete sessions as needed.

Chat Agent Token-by-Token Streaming

Chat Agent responses now stream incrementally for a more responsive experience, check out the implementation to see how you can easily achieve per-token streaming in your agents.

Live Reloading

Edit your agent code and see changes instantly in the UI - no server restart required. Run: uv run watchfiles beeai_agents.agent.run

Complete Changelog

Full Changelog: v0.3.5...v0.3.7

Contributors

JanPokorny, tomkis, and 6 other contributors
Loading

v0.3.5

03 Oct 07:13
@tomkis tomkis
v0.3.5
e395307
This commit was signed with the committer’s verified signature.
jezekra1 Radek Ježek
SSH Key Fingerprint: WY/4hPdpbLCS3enpAvW/slfDB27VN4vXs1Sbg7vrwHo
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.3.5

Major Features

🔑 Secrets Management

You can now securely manage Secrets directly in the UI and store them in the API.
This allows delegating the responsibility of secrets provisioning to the user.

Checkout the docs for some examples.

🏗️ Server-Side Builds (experimental)

We’ve added server-side image builds:

  • You can now build agents as Github repos
  • Support for building images in external clusters

These changes are needed to enable dynamically adding github agents via GUI, which will come in next releases.

❗️Breaking Change

Agents history is no longer persisted automatically, you need to store it explicitely in your agent code. We've introduced this change to enable more flexible persistance which is always agent specific.

@server.agent()
async def my_agent(input: Message, context: RunContext):
    await context.store(input)       # Store incoming message
    
    response = AgentMessage(text="...")
    yield response
    
    await context.store(response)    # Store outgoing message

What's Changed

Full Changelog: v0.3.4...v0.3.5

Contributors

JanPokorny, tomkis, and 6 other contributors
Loading

v0.3.4

19 Sep 09:25
@tomkis tomkis
v0.3.4
aa96907

Choose a tag to compare

View all tags
v0.3.4

Major Features

🛠 Agent Settings

We’ve introduced Agent Settings, a new way to configure and customize agent behavior directly in the UI. This feature allows developers to define adjustable parameters for their agents such as toggles or dropdows.

Agent settings are making agents more flexible and easier to adapt without code changes. Alongside the UI, we’ve added documentation and simplified examples to help you get started quickly with building agents that support configurable settings.

🔁 Persistent Context History

Added support for storing and accessing context history so conversations retain continuity.

Context can be used over multiple messages/sessions, improving relevance of agent responses.

What's Changed

Full Changelog: v0.3.3...v0.3.4

Contributors

JanPokorny, tomkis, and 8 other contributors
Loading

v0.3.3

09 Sep 07:36
@tomkis tomkis
v0.3.3
1eefb09
This commit was signed with the committer’s verified signature.
jezekra1 Radek Ježek
SSH Key Fingerprint: WY/4hPdpbLCS3enpAvW/slfDB27VN4vXs1Sbg7vrwHo
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.3.3

🚀 Release Notes – v0.3.3

Key Highlights

Reworked Providers & Variables

We’ve redefined how LLM configuration is managed in BeeAI:

  • New beeai model command – add and manage multiple LLM providers.
  • Interactive setup improved – replaced beeai env setup with beeai model setup.
  • Model and provider listing – easily view all available LLMs and providers.
  • Smarter agent behavior; When an agent uses the LLM extension, the platform automatically selects the most optimal LLM based on your code’s suggestions.

UI Form Extension

Thanks to support from the Form extension, agents can now elicit structured user input via dialogs. Whether at the start of a session or dynamically mid-chat

  • Initial Form Rendering - You define a form layout (with FormRender) in your agent’s definition. Then, when a user begins interaction, the form appears immediately before any messages.
  • Dynamic Form Requests - Within the conversation flow, agents can prompt users with additional forms as needed

Noteworthy changes

  • Dependency Bumps: beeai-framework, A2A SDK, and updated Next.js.
  • Documentation updates
  • Layout tweaks such as session button placement, improved code block rendering, streamlined provider dropdowns, and consistent agent card URLs.

Changelog

New Contributors

Full Changelog: v0.3.2...v0.3.3

Contributors

JanPokorny, tomkis, and 11 other contributors
Loading