Runcell releases

Release 0.2.3

release date: 2026-08-11

Runcell 0.2.3 makes longer Agent projects easier to continue, pause, and recover. It also improves inline code completion and fixes clean-install problems related to image support.

You can now turn a larger request into a persistent Goal. Runcell keeps the objective visible above the chat input and lets the Agent continue the work across multiple turns instead of trying to finish everything in one response.

Goals can be started from the composer “+” menu or created by the Agent when a task needs more time. You can pause a Goal, resume it later, and restore its state after refreshing JupyterLab or reopening the conversation. Stopping an active Goal also preserves it in a paused state so the work is not lost.

More Reliable Recovery

Runcell now restores Agent progress more accurately after a refresh or temporary connection interruption. Interrupted tool calls are no longer mistaken for completed work, repeated tool results are filtered out, and completed multi-part results are preserved more reliably.

These changes reduce duplicated work and make it safer to continue a long task after reconnecting.

Smarter Inline Completion

Inline code completions now understand code blocks more accurately when replacing or rewriting existing code. Additional safety checks help prevent suggestions from overwriting the wrong range.

Model and Installation Updates

Claude Opus 5 is now available as a model option.

Clean installations now include the image-processing dependency Runcell needs. This fixes an issue where the extension could fail to start after a fresh installation, even though the installation itself appeared to succeed.

Release 0.2.2

release date: 2026-07-20

Runcell 0.2.2 adds collaborative Agent workflows, support for external tool connections, richer image understanding, and more dependable long-running tasks.

Sub-Agents for Parallel Work

The Agent can now delegate independent parts of a larger request to sub-agents and run several of them in parallel. Their progress and results appear in expandable cards inside the main conversation, so you can follow the work without switching between multiple windows.

For work that should continue independently, the Agent can also create a separate conversation and hand off an initial task to it.

External Tool Connections

Runcell can now use compatible MCP connections, including local tools and supported remote services. Connected tools participate in the normal approval flow, and sub-agents can reuse the connections already available to the parent task.

Better Image Understanding

Images produced or read by Agent tools can now be passed back to supported multimodal models. This helps the Agent reason directly about charts, plots, screenshots, and other visual results instead of relying only on a text description.

Chat Export and More Models

Synced conversations can now be downloaded as Markdown, making it easier to keep analysis notes, share results, or archive important Agent sessions.

This release also adds support for Claude Sonnet 5, GPT-5.6, and Grok 4.5.

Reliability Improvements

Improved recovery for long-running tasks after a connection interruption or JupyterLab refresh.
Added more reliable shell support on Windows.
Fixed notebook kernel recovery and several cases where notebook tools could fail after reconnecting.
Fixed cases where a sub-agent could block the parent Agent from continuing its own tools.
Improved sign-in, upgrade, credit display, and narrow-panel behavior.

Release 0.2.1

release date: 2026-06-25

Runcell 0.2.1 makes Agent code changes easier to review and introduces a complete interface for creating and managing reusable Agent Skills.

Review Every Agent Turn

After an Agent turn changes files, Runcell now shows a compact summary of the affected files and added or removed lines. You can expand the card to inspect the exact diff and revert the turn when needed.

Git Viewer

New Git panels let you inspect changed files, repository history, and diffs directly in JupyterLab. Improved navigation makes it easier to move between files and review what the Agent changed without leaving your workspace.

Agent Skills

You can now create, import, enable, disable, and delete reusable Agent Skills from the Runcell interface. Skills can apply to one project or be kept in your personal library, and slash commands let you explicitly invoke a skill from the chat input.

Notebook and File Reliability

Runcell now checks and recovers the notebook kernel before running Agent code.
File-tree updates and Git status refreshes are more stable.
Large file and search results use smaller, more focused windows so the Agent can find relevant content without filling the conversation context.
GLM 5.2 is now available through OpenRouter.

Release 0.2.0

release date: 2026-06-13

Runcell 0.2.0 introduces interactive shell sessions and gives you more control over how conversations are reused, copied, and recovered.

Interactive Shell Sessions

The Agent can now keep a terminal command running, read new output as it arrives, and send additional input to the same process. This is useful for installers, development servers, training jobs, and command-line programs that need confirmation or other interaction.

Copy and Fork Conversations

Individual chat messages can now be copied with one click. You can also fork a conversation from an earlier point to explore a different approach while keeping the original conversation intact.

Fewer Permission Interruptions

Selected low-risk tools are allowed by default, reducing repeated approval prompts during routine work. Commands and actions that need confirmation still use the normal permission controls.

More Model Options

This release adds Claude Fable 5, Kimi K2.7 Code, MiniMax M3, and Xiaomi MiMo v2.5 Pro.

Reliability Improvements

Long-running Agent jobs now reconnect automatically when their event stream is interrupted.
Permission requests and completed tool results are preserved more reliably during recovery.
Failed turns can be retried from the point of failure instead of regenerating the entire response.
Authentication, conversation restore, and model refusal messages now explain the next step more clearly.
Fixed missing Plotly snapshots in Agent tool results and several cases where a disconnected task was incorrectly shown as stopped.

Release 0.1.25

Runcell AI 0.1.25 focuses on Agent memory, smoother long-running work, and fixes for several issues reported by users in recent Agent sessions.

In particular, this release addresses problems some users saw during long tasks where tool calls could appear to produce no output, responses could be cut off, or reasoning-model sessions could fail to continue cleanly. These issues were most visible with models such as Sonnet 4.6, Opus 4.7, and Opus 4.8.

Agent Memory

Runcell Agent can now remember useful information across sessions, such as your preferences, project conventions, repository context, and important notebook or dataset notes.

This helps the Agent pick up work with less repeated setup, especially when you return to the same project later. You can also ask the Agent what it remembers or ask it to forget outdated information.

Better Long-Running Agent Reliability

Long-running Agent tasks are more resilient when the notebook connection briefly drops or the JupyterLab tab refreshes.

Runcell now recovers more gracefully from short connection interruptions, reducing cases where an Agent task gets stuck because the notebook bridge temporarily disconnected.

Continue Long Responses

When a model response reaches its output limit, Runcell now shows that the response was incomplete and lets you continue from where it stopped.

You can send "continue" to resume the answer instead of restarting the request or losing the unfinished response.

Smoother Chat Workflow

You can now queue a follow-up message while the Agent is still responding. This makes it easier to steer long-running conversations without waiting for the current response to finish.

User messages are also easier to revise: edited messages now support image attachments, including pasted or uploaded images.

Cleaner Tool Output

Long command outputs are now automatically collapsed in chat, making conversations easier to scan while still keeping detailed output available when needed.

Notebook and command outputs are also cleaned up more consistently, including better handling of terminal formatting characters.

More Model Options

This release adds support for more models, including Claude Opus 4.8 and Gemini 3.5 Flash.

Image understanding and content-reading features also use improved default model routing for better quality and reliability.

Bug Fixes

Fixed long-task issues where tool calls could appear to have no output, incomplete output, or truncated responses.
Fixed continuation and reasoning compatibility issues seen with several advanced reasoning models, including Sonnet 4.6, Opus 4.7, and Opus 4.8.
Fixed several issues that could interrupt long-running Agent tasks after notebook reconnects or temporary frontend disconnections.
Fixed cases where stopped server-side tool calls were not saved correctly in chat history.
Fixed notebook execution timeout handling so timed-out runs are reported more clearly and preserve available output.
Fixed cases where the Agent could fail when trying to run or edit a non-code notebook cell as code.
Fixed file reading for some non-UTF-8 encoded files.
Fixed authentication state handling when login is cancelled.
Improved error handling so user-facing messages are clearer while preserving enough detail for the Agent to recover.

Release 0.1.24

Runcell AI 0.1.24 adds more control, clearer credit usage, and a smoother notebook experience.

Custom Reasoning Effort

Users can now choose the reasoning effort for supported models: Low, Medium, High, or Extra High.

This gives more flexibility when balancing speed, cost, and quality. Use lower effort for quick tasks, and higher effort for complex analysis, debugging, or coding work.

Token-Based Credits

Ask/Chat and Agent Credits are now calculated based on actual token usage instead of a fixed per-message cost.

This makes Ask/Chat and Agent usage easier to understand: longer conversations, larger context, model choice, cached tokens, and output length now affect credit usage more directly. Other AI features, such as completions and utility actions, continue to use their standard feature-based credit rules.

New features:

Parallel Agent conversations — The headline feature of 0.1.22. You can now run multiple Agent-mode conversations in parallel, so a second conversation can work on task B while the first one is still busy with task A. Running several Agents side by side meaningfully improves throughput when you have multiple things going on at once. A couple of practical notes:
- We recommend avoiding having multiple Agents operate on the same notebook at the same time.
- Agent conversations are isolated from each other to prevent cross-task conflicts.
Relaxed default Shell permissions — Once you've approved a permission, similar Shell calls will no longer keep asking for it. In practice this means far fewer interruptions when the AI uses Shell Execute frequently, especially for basic read-only commands, which are now open to the Agent by default.
Offline / detached Agent tasks — Once an Agent task has started, it will keep running even if the notebook is closed or refreshed. This prevents long-running tasks from being interrupted just because a tab was reloaded or closed.
URL fetch tool — You can now paste URLs directly into the conversation and the Agent will read and understand the page content. This is great for pointing the Agent at specific documentation pages or other reference websites.

Bug fixes:

Fixed high latency in several tool calls.
Fixed an issue where the Agent would occasionally insert cells at the wrong position.
Fixed missing "currently streaming" prompt indicators in the Chat UI in certain situations.
Fixed the model list not updating automatically.
(0.1.23) Fixed a chat-merging issue between parallel Agent conversations that could occur depending on the timing of when conversations were created.

Release 0.1.21

release date: 2026-04-03

Added Compact for long-running conversations. When a model's context gets close to its limit, Runcell can now compress the conversation context to keep the session moving instead of forcing you to restart or lose momentum. We currently support both Manual Compact and Auto Compact, so you can either decide when to compress or let Runcell handle it automatically.
Added parallel tool calling. This significantly improves how fast the Agent can read and understand large, complex codebases by working across multiple tools at the same time, which makes deep repository analysis much more responsive in real-world engineering workflows.
Improved Agent mode performance, with better overall responsiveness, lower latency, and more stable network behavior during longer sessions.
Fixed missing History backup timing, making backups more reliable and reducing the chance of losing important session history.
Fixed a large-file write issue, so write-file operations are now more stable when working with bigger files.
Polished a number of UI details for a smoother overall experience.

Release 0.1.20

release date: 2026-03-08

Fixed an issue where text content was being truncated when file editing and read/write tools processed excessively large files.
Resolved an "Authentication Error" (specifically occurring under certain unique conditions).
Added support for Shell Execute, providing finer-grained control over command-line tool instructions.
Added support for new models, including GPT-5.4 and GPT Codex 5.3.
Runcell now can read and understand images/visual content on local devices/in notebook.
Better Agent context and better tools for agent to understand notebook(.ipynb) content

Release 0.1.19

release date: 2026-02-19

Fix compatibility with interactive charts libraries, like Plotly, Vega Altair or even ipywidgets.
Runcell Agent now can understand visual elements from plotly, vega altair and lots of ipywidgets instead of regarding them as html.
New upgrade tool, which allow user can one click to finish upgrade + restart jupyter
New model support - Sonnet 4.6, Gemini 3.1 Pro, GLM 5, MiniMax 2.5 are now available.
Fix credits remain display bug
Fix file system tools result with large content waste the context window of AI agent.

Release 0.1.18

release date: 2026-02-10

Pause/Resume for agents — The agent can now be paused and resumed, which means it can handle long-running tools (model training, large database queries, scripts that take hours) without hitting timeout issues. This was a known bug in earlier versions.
Permission controls for tool calls — runcell now asks for your approval before executing actions like running code, bash commands, or installing dependencies. This gives you more control over what the agent does. It's fully configurable — you can choose between manual approval or letting it act directly.
New model support — Opus 4.6, Codex 5.2, and Kimi 2.5 are now available.

Release 0.1.17

release date: 2026-01-12

Gemini 3 models support
Use nano banana to optimize visualizations

Release 0.1.16

release date: 2025-12-28

Runcell releases

On this page