Changelog

​

Remote MCP servers

• Connect remote MCP servers with OAuth. Enable OAuth in the server’s Authentication section, let Overcut register a client automatically or use your own OAuth client with its secret stored in the Vault, then connect an account. Tokens are renewed during a run, so long runs keep access to the server. See Connect with OAuth.

​

Code review

• A new Resolve Review Thread tool lets agents resolve a review thread once its feedback is addressed, or reopen one that needs more work. It works on GitHub, GitLab, Azure DevOps, and Bitbucket. The tool is opt-in: select it under Additional tools in the agent’s settings. See Tools.

​

Single sign-on

• Sync team memberships from identity provider groups. Link a team to one or more IdP groups and turn on Sync team memberships from IdP groups for the provider; each SSO sign-in then adds users to matching teams and removes them from linked teams they no longer match, so new users land with the right project access on their first sign-in. See Sync teams from IdP groups.

​

Project repositories

• For each repository selected in a project, Listen to events controls whether that repository’s events reach the project. Turn it off to keep a repository available to agents without it starting workflows or orchestrations.
• Agent access set to All workflows gives agents the repository in every workflow of the project, not only when the workflow’s trigger or a clone brings it in. Scheduled and Slack-triggered workflows can now reach a repository they do not start from. See Configure Events and Agent access.
• Project repositories is now the first tab in project settings, and General is the last.

​

LLM models

• Max output tokens on a custom model is now honored, so Bedrock models no longer stop at 4,096 output tokens and Azure OpenAI models receive the limit in the format they expect. The field now sits under Provider configuration.
• Improved prompt caching for agent steps on Anthropic models, directly or through Bedrock, which lowers input-token usage on long runs.

​

App experience

• Refreshed app styling, including a new font, an updated dashboard, tinted run and step status badges, and calmer run-log messages that mute agent setup lines.
• Labels, titles, and navigation use sentence case throughout.
• Teams and roles are created from an Add button and dialog.

​

Fixes

• Claude engine steps on Opus 5.5 now run instead of failing with an unsupported-model error.
• Scheduled workflows no longer start a duplicate run for the same occurrence.
• A sub-agent turn cut off at the output token limit now fails the run instead of returning an empty result that the coordinator keeps retrying.
• Messages from classic Slack apps, such as Sentry or Opsgenie, are now recognized as bot messages, so trigger conditions that exclude bots skip them.
• The built-in Workflow Builders and Workflow Run Viewers teams now include the skill and MCP server permissions they need, so non-admins can view and add skills and MCP servers.
• Skills in a repository are no longer hidden when its marketplace.json cannot be resolved.
• MCP servers whose names contain __ now expose their tools correctly.
• overcut mcp now installs a working connection for Claude Desktop.

​

Custom agent images

• Register container images from your own registry and select one per repository, so agents clone, build, and test with the exact SDKs, compilers, and internal CLIs your code needs. Images are pinned to a digest, so runs stay reproducible until you choose to pick up a new build. Private registries are supported. Custom images are an Enterprise feature; the system images remain available on every plan. See Agent Images.

​

Build from your coding agent

• Connect Claude Code, Cursor, Codex, or any MCP-capable coding agent to Overcut with one command, npx overcut init. The agent can then connect your Git provider, register repositories, design and publish workflows, and check on runs, all in conversation and without switching to the web app. Everything it does runs as you, inside your workspace’s permissions and audit trail. See Connect Your Coding Agent.
• Sign in from a terminal, including SSH and headless sessions, and manage API tokens from the CLI for CI and scripts.

​

Single sign-on

• Improved SSO experience with built-in management: view and control your workspace’s SSO settings under Security > Single sign-on, including whether the provider is enabled and whether users on your email domains join automatically. SSO still requires initial setup by the Overcut team.

​

LLM models

• Custom LLM models now discover their own capabilities. Testing a model detects whether it supports tool calling, vision, PDF input, and prompt caching, and runs use those verified results instead of assumptions based on the model name. See Testing a Model.
• Prompt caching detection is improved for Anthropic models served through custom endpoints, such as Bedrock inference profiles, so caching is used wherever the endpoint supports it.

​

Messaging triggers

• Choose the emoji Overcut reacts with when it picks up a Slack message, so the acknowledgment matches the conventions your team already uses and does not collide with reactions that mean something else in your channels. You can also turn the reaction off entirely. See Acknowledgment reactions.

​

Workspace

• The workspace sidebar is regrouped into Context, Agent Runtime, and Governance & Admin, so integrations, model and image settings, and administration are easier to find.

​

Fixes

• An agent session kept open for comments now completes when its time limit ends, instead of leaving the run in Terminating indefinitely.
• The runs list sorted newest first no longer starts with queued or skipped runs.
• Orchestration steps for a cancelled or completed instance are no longer re-triggered.

​

Workspace Library

• Every workspace now has one Workspace Library, opened from the project picker, for the agents, MCP servers, skills, and secrets that several projects share, and for workflow and orchestration templates. See Workspace Library.
• Projects use Library agents, MCP servers, skills, and secrets by reference, so an edit in the Library reaches every project. Library items carry a Library badge in pickers.
• Promote to library moves a project resource into the Library without breaking existing links. Install from Library copies a published template into a project as an independent draft.
• Chat can create and list Library resources, promote project resources, and install templates.

​

Context parameters

• Define a value once and reference it as {{params.<key>}} in step instructions and agent instructions. Each project, repository, workflow, orchestration, or agent can set its own value, and the most specific one wins at run time. See Context Parameters.
• Manage definitions on the new Context Parameters workspace page and the Context Parameters tab in Project Settings.
• Instruction editors offer completion for {{params. references.
• The Playground previews the value each key resolves to, and the existing Run Context panel now lists the resolved parameters under params.
• Chat can define parameters, set values, and preview resolution.

​

Orchestration discussions

• Pending orchestration decisions now have a dedicated in-app discussion page. Send the first message to start the discussion; decision details, case-file context, agent activity, and the full conversation history sit alongside the chat. See Approvals and Discussions.
• Pending decision cards and tracked-item questions link directly to their discussion.

​

Messaging channels

• Every workflow run now receives all active channels registered to its project, so agents can post updates to any project channel from a scheduled, manual, Git-triggered, or message-triggered run.
• When a project has one registered channel, agents use it automatically. With several channels, agents receive the list of destinations and choose one explicitly.
• Replies default to the triggering thread only in the channel where the trigger message arrived. Messages to other channels start at the channel root.
• Native Slack slash commands that arrive without a message timestamp now dispatch instead of silently failing.
• A project or channel lookup failure, or one invalid channel credential, no longer stops the run or blocks the project’s other channels.

​

Claude engine

• Claude-engine token usage is now counted in Token Usage Analytics for workflows, runs, steps, models, and agents. Billing and credits are unchanged.

​

LLM models

• Custom LLM models have a Supports Vision setting that controls whether image content is sent to the model. See LLM Models.

​

Playground and workflow drafts

• A draft missing a trigger or another required field no longer breaks its project’s workflow list or detail pages. Open, repair, export, or delete the draft; publishing still validates it.
• A malformed trigger on one workflow no longer prevents the project’s other workflows from loading or running.
• The Playground recovers when its selected workflow has been deleted: it shows a warning, keeps valid run history and current selections, and stops submitting the stale workflow. A failure to load a workflow’s parameters no longer replaces the whole page.
• A never-published workflow no longer shows preview data in the Playground unless the working draft is explicitly selected.

​

Other fixes

• Repository selections persist across searches and pages when adding repositories, and the Add Selected button shows how many are selected. After a partial bulk-add failure, the failed repositories remain selected for retry.
• Saving an LLM model no longer leaves a false unsaved-changes warning or stale configuration values in the form.
• Agents without a description can be edited and saved.
• Unknown tool names are removed from an agent’s tool list consistently, so agents no longer advertise tools that cannot run.

​

Claude engine models

• Claude-engine steps now use the models you configure in the workspace, each with its own credentials, instead of a single workspace-wide Claude API key. Anthropic and AWS Bedrock models are supported. See Claude Agent SDK.
• Set a workspace Default Claude Engine Model in Workspace Settings > AI. A step resolves its model in order: the explicit step or agent choice, then the workflow default, then the workspace Claude default, then the main workspace default.
• Pin a different model per sub-agent on the Claude engine. Pinning changes only which model answers: credentials, billing, region, and gateway always come from the session’s coordinator model, so a sub-agent model must be from the same provider as the session.
• Anthropic models accept a Base URL, so the Claude engine can route through a gateway that serves the Anthropic wire format (/v1/messages).
• Existing workspaces were migrated automatically: the legacy workspace “Claude API key” secret became a “Claude (migrated)” Anthropic model, set as the workspace Claude default. No action is needed. The Workspace Secrets tab was removed along with the legacy secret.
• Misconfigured models fail at dispatch with a message naming the model and the cause. There is no silent fallback to another key or model.

​

AWS Bedrock

• Bedrock models can now authenticate with a Bedrock API key as an alternative to an IAM access key pair. When both are set, the API key is used. See LLM models.
• Prompt caching now applies to Claude models routed through OpenRouter and AWS Bedrock, cutting token usage and latency on repeated context.
• Bedrock cache tokens are now counted in LLM usage.

​

Triggers

• Messaging trigger conditions now expose the full message context in the workflow builder’s condition selector: channel ID, channel name, channel provider and privacy, workspace ID, message text, and message and thread timestamps. Build conditions such as context.channel.id alongside the existing actor fields.

​

Orchestrations

• Human replies and decisions now resume an orchestration reliably: transient failures are retried automatically, and an instance that cannot recover is shown as stalled instead of failing silently.

​

Workflow builder and runs

• Run details and logs now show step names instead of step IDs, using the names from the run’s workflow snapshot even after the workflow changes.
• Model errors during runs are now classified and reported with distinct messages instead of all surfacing as model unavailable.
• Fixed parallel tool calls from OpenAI models being misrouted.
• Fixed workflow builder steps rendering stacked on top of each other.
• Fixed the Space key not registering in instruction editors inside the workflow builder.

​

Integrations

• Agents can now read ticket comments on Linear and ClickUp.

​

On-prem deployments

• All services now honor HTTP_PROXY, HTTPS_PROXY, and NO_PROXY, including spawned agent containers, so Overcut can run behind an egress proxy.

​

Other fixes

• Removed team members no longer appear in the team member list.

​

Orchestrations

• Bounded autonomy: agents pick the path, but only from the workflows you allowed, within the limits you set, and never past a gate without a human decision.
• Human judgment where it matters: gates pause the process for approval, and you can discuss a pending decision with the orchestration in chat before you decide.
• Supervised to completion: a supervisor keeps each instance moving, detects stalls, and judges the result against the goal, not just whether the steps ran.
• Accountable by default: every routing decision is recorded, and the conversation stays connected to the work item. Comments on the original issue or pull request flow into the process, and replies go back to it.
• Visible in operation: every in-flight item is tracked in one place, from intake to resolution, and any instance can be inspected or cancelled at any point.

​

Machine sizes for workflow runs

• Choose the machine size a workflow runs on with the new Machine Size setting in workflow settings: standard (default, 1 credit per run), large (2 credits), or xlarge (4 credits). Larger sizes provide more CPU and memory for workspace commands such as builds and tests.
• The size is frozen when the run starts and shown on run views, and the run log records the machine size used.

​

Slack thread reply trigger

• Workflows can now trigger on thread_reply: when someone replies in a watched Slack thread, the workflow fires with the reply as context. Configure it like any other trigger in the workflow builder.

​

File attachments

• Attached DOCX, XLSX, and PDF files are now parsed and their content made available to agents. PDFs are passed natively to Anthropic models.
• Fixed attachments with non-ASCII filenames (for example Hebrew) arriving garbled.

​

Models

• The GPT-5.6 family is available as system models: gpt-5.6-sol (flagship), gpt-5.6-terra (mid), and gpt-5.6-luna (small), all with vision enabled.
• Seven legacy system models were retired: o3, o3-mini, gpt-4.1, gpt-4.1-mini, gpt-4o-mini, gpt-5-codex, and gpt-5.2-codex.

​

Assistant and chat

• The workflow builder assistant now suggests a publish message when you publish workflow changes.
• The chat input suggests your next message as ghost text; press Tab to accept.

​

Azure DevOps

• The agent’s completion message is no longer overwritten by later status comments on pull requests.
• Fixed list_tickets keyword search failing due to invalid WIQL syntax.

​

Reliability & fixes

• Agent start failures (for example an image that cannot be pulled) now surface a clear error and fail fast instead of hanging.

​

Run Script workflow action

• New script.run action runs an inline bash script as a deterministic workflow step, with no AI agent or model tokens involved.
• Scripts receive trigger data and previous step outputs through environment variables, and can return structured JSON to later steps by writing to $OC_OUTPUT_FILE.
• Secret values are redacted from script results, and a configurable timeout stops the script and every process it spawned.
• The script field uses a code editor with syntax highlighting, and the workflow builder assistant can add and configure the step for you.
• Read the docs

​

Run Context panel

• Workflow run details now include a Run Context panel: select Context to see the actual trigger data and outputs JSON available to the run’s steps and templates. Step outputs update while the run executes, so you can verify what template expressions like {{outputs.<stepId>.<field>}} resolve to.

​

LLM gateway support

• Custom LLM models on OpenAI, OpenRouter, and Azure OpenAI (Responses API) can now route through a proxy or API gateway: override the base URL, choose how the credential is sent (bearer Authorization, api-key header, or a custom header name), and add static extra headers for gateway routing. Standard Azure OpenAI also supports extra static headers. Existing models keep their current behavior.
• Fixed deactivated models becoming unrecoverable: inactive models stay visible on the LLM Models page, where you can review, edit, or reactivate them, while model pickers only offer active models.
• Read the docs

​

Azure DevOps

• Code review results are now submitted as reviewer votes with comment threads on Azure DevOps pull requests, and retries no longer create duplicate threads.
• Fixed automated replies from the integration’s service account re-triggering workflows.

​

Skills

• Add multiple skills at once from the Add Skill page.

​

Slack

• Channel pickers load faster: channel lists are cached server-side with debounced search, and repeated scans no longer trip Slack rate limits.

​

UI refresh

• Workflow builder nodes use a new tab-card style with a floating action toolbar on the canvas.
• Fixed text overflow on agent cards.

​

Reliability & fixes

• Fixed agent connections dropping mid-run due to premature socket closes and connection registration races.
• Fixed several tool-call handling errors that could stall LLM conversations or break retries.
• Agent status reporting now times out safely during transient outages instead of crashing the run.
• Repository connections no longer show duplicates across providers, and abandoned connection drafts are cleaned up automatically.
• Repository identification tolerates extra text around JSON output instead of failing the workflow.
• Fixed ClickUp requests failing on some content-type headers.
• The universal dev container image (version 6) now ships Node 22.

​

Token usage analytics

• New token usage analytics on the dashboard, with charts and metrics that break down LLM consumption over time and per workflow, so you can see where tokens are going.
• Workflow run consumption can be filtered at the workspace level, and workflow names in the usage charts link straight to their runs.
• Cached input tokens are now tracked and reported as part of usage.

​

Audit trail

• Full audit trail across every workspace-scoped resource, recording who changed what for security review, compliance, and incident investigation.
• Browse and filter the audit log in the UI, defaulting to the last 7 days, with field-level diffs of each change.
• Read the docs

​

Models

• Added support for GPT-5.5 and GPT-5.4-mini.

​

Prompt caching

• Anthropic prompt caching is now applied to the system prompt, reducing cost and latency on repeated runs.
• Prompt-cache tokens are reported consistently across the Anthropic and OpenAI Responses API paths.

​

Dev containers

• Added a new universal image (version 6) and a .NET 10 variant.
• Agent image resolution now spans multiple repositories, with precedence and clear reporting when images conflict.

​

Preparation steps

• Workflow run details now show preparation steps, such as setting up the workspace and provisioning the environment, as their own distinct steps. You can follow each one as it runs and see exactly where time is spent before your workflow begins.

​

Reliability & fixes

• Chat sessions are more stable and stay connected through infrastructure disruptions instead of hanging.
• Fixed LLM retries failing because trailing system messages violated Anthropic’s message ordering requirements.
• Custom model configuration errors are now surfaced, and stale configuration no longer lingers after editing a model.
• Fixed API tokens expiring prematurely instead of honoring the rolling inactivity window.
• Workflow failures now surface git token errors in the run status instead of failing silently.

​

Triggers

• Issue triggers now expose classification fields - priority, severity, components, areaPath, and iterationPath - across Jira, Azure DevOps, and Linear, so you can filter and branch workflows on them.
• Issue triggers now provide granular labelsAdded and labelsRemoved arrays, so you can react to exactly which labels changed rather than the full label set.
• Linear issues with no priority now map to an undefined priority instead of a “No priority” label, so priority filters behave consistently.

​

Azure DevOps

• Connecting an Azure DevOps organization now uses a guided, multi-step OAuth flow.
• Added support for legacy visualstudio.com URL formats.
• Azure DevOps comments are now fully converted from HTML to plain text, so comment content reads cleanly in workflow runs.
• Clearer error handling when prefetching Azure DevOps data fails during setup.

​

Repository connections

• Connecting a repository now surfaces webhook creation errors immediately, instead of silently leaving the repository connected with no webhook.
• Unhealthy repositories now show their state with a retry option in the integrations UI.
• Soft-deleted repositories are excluded from project lookups, preventing stale settings validation errors.

​

Reliability & fixes

• Fixed agent log scrolling so a tall message no longer hijacks the page scroll until it is focused.
• Fixed dropped tool responses when several tool calls run in one batch. Synthetic attachment messages are now deferred until all sibling tool calls complete.
• Fixed orphaned tool calls by limiting internal agent delegation to one per turn.

​

Custom Events

• Incident response automation across your full stack - connect Datadog, Grafana, PagerDuty, internal alerting systems, or custom observability pipelines directly into remediation workflows.
• Deployment and CI orchestration beyond built-in integrations - Jenkins, Buildkite, internal deployment tooling, release scripts, or platform APIs can trigger workflows directly.
• Enterprise-wide workflow automation - connect internal portals, security systems, vendor platforms, legacy infrastructure, or any custom service into your SDLC automation layer.
• Unified orchestration across all systems - the entire operational ecosystem around your engineering organization, not just the tools Overcut ships with out of the box.

• Flexible authentication (bearer token, custom header, or URL-only) to support different sender capabilities.
• Token and secret rotation without downtime.
• Concurrency control via lockKey, allowing workflows that touch the same resource to execute safely and sequentially.
• Project-level scoping to control which teams can consume specific event types.
• Structured payload fields that keep workflow runs readable in the UI while remaining fully filterable in workflow logic.

​

Meet your new Overcut assistant

• A new home screen welcomes you with a focused chat experience. Start a workflow, ask a question, or kick off a task from a single input.
• Persistent header chat stays with you across the app, expanding and collapsing as you work.
• Conversation history with renaming, deletion, and auto-generated titles, so you can pick up exactly where you left off.
• Real-time updates across the UI. When the assistant creates or edits something like a workflow, permission, or trigger, the UI updates instantly without a refresh.
• Smarter and more capable. The assistant can now perform almost anything available in the UI. Ask it to connect an MCP server, inspect logs, manage integrations, or execute workflows.

​

Multi-project workspaces

• Global project picker in the workspace header for fast switching.
• Per-project repository access controls. Choose whether a project can access all connected repositories or only a selected subset.
• Project-level permissions with team and role assignment for finer-grained access control.
• Project colors for quick visual identification.
• Project-scoped runs, dashboards, and consumption summaries that automatically respect the active project.
• Restructured sidebar with workspace ↔ project mode switching and nested settings.

​

Slack integration

• Connect Slack to your workspace, manage channels, and send messages directly from workflows and the assistant.
• Channel message triggers that kick off workflows when Slack messages arrive, with filtering by channel, user, and message content.
• Expanded OAuth scopes for message history and reactions support.

​

Workflow builder & orchestration

• Syntax-highlighted instruction editor with live preview in a full code editing experience.
• New agent tools for linking and unlinking tickets across Jira, Linear, and other ticketing systems, including support for relationship types.

​

UI polish

• Brand-new icon system for crisper and more consistent visuals across the platform.
• Restyled sidebar, menus, and profile areas with refined design tokens.
• Centralized breadcrumbs for clearer navigation context across every page.
• Automatic selection of your latest execution or default repository when opening the playground.
• Smarter workflow run defaults that jump directly to the most relevant log step based on run status.

​

Reliability & fixes

• Improved handling for large agent outputs with automatic management of oversized tool results.
• Cleaner branch checkouts. Cached working trees now discard stray uncommitted changes before switching branches.
• Fixed an issue where deleting a comment containing a slash command could accidentally retrigger it.
• Fixed PR approvals via submit_review failing with “no pending review”.

​

Orchestration Map

​

Skills

• Reuse skills from your existing ecosystem and marketplaces.
• Apply the same governance and management layer across all skills.
• Control which agents can use which skills to keep execution focused and avoid prompt bloat.
• Ensure consistency while still allowing flexibility at runtime.

​

Playbook Catalog

​

Workflow Visibility

​

Change Accountability

​

Automatic Workflow Self-Improvement

• Investigate - analyzes logs, tool usage, and outcomes across sampled runs. Identifies errors, inefficiencies, and recurring patterns, including whether existing memories helped or hurt.
• Process tentative memories - validates previously tentative memories. Repeated patterns are promoted to active; contradictions are removed.
• Create memories - generates new memories from cross-run patterns. Patterns seen in 2 or more runs become active; strong single-run signals stay tentative.
• Update weights - adjusts memory relevance (0.0-1.0) based on impact. Helpful memories are strengthened; ineffective ones decay. Memories below 0.05 are archived.
• Summarize - produces a structured report with findings, memory updates, and recommendations.

• Run threshold - number of runs before triggering a retrospective.
• Sample size - percentage of runs analyzed.

​

Agent Tool Visibility

• New “Tools Available” message in each agent thread.
• Captured for both main agents and sub-agents.
• Included in thread summaries for quick reference.
• Infrastructure logs excluded from step counts to keep things clean.

​

Cancel Queued Workflow Runs

• Cancel button with confirmation dialog.
• Transitions the run to Failed with a clear status message.

​

GPT-5.3 Codex as Default Model

​

Workflow Version Tracking

​

ClickUp Integration

• Create comments.
• Update statuses.
• Track attachments.

​

Agent Memory

• Dedicated UI to view, edit, and manage memory.
• Memory scoped per workflow step for precision.
• Full visibility into which memories are used during execution.
• Tentative memory system where agents suggest new memories for approval.

​

Retrospective Workflows

• Analyze agent logs and tool usage.
• Evaluate outcomes.
• Generate a consolidated summary with actionable insights.

​

Workflow Dashboard

• Recent runs.
• Average duration.
• Quick access to run details.

​

GPT-5.4 Model Support

​

Custom Agents

​

Additional Improvements

• Use case organization - cleaner tab-based navigation.
• Workflow output display - properly formatted single-message outputs.
• Scrollable chat - improved experience for long agent conversations.

​

MCP Catalog & One-Click Installation

• Configuring required secrets.
• Selecting which tools to expose.
• Previewing the final configuration.
• Creating the server.

​

MCP Server Secrets

​

Scratchpad Tools for Multi-Agent Workflows

​

GPT-5.3-Codex Model Support

​

New Code Review Playbook + Major Improvements

• The new scratchpad tools.
• Improved reasoning flow.
• Parallel execution.

​

Context Map

• Visually connect ticket repos to code repos with simple drag and drop.
• Instantly see which features are enabled per repo (caching, indexing, hints, custom instructions).
• Filter by org or configuration.
• Focus on a specific repo and highlight only its related connections.
• Edit repo configuration inline from the same screen.

​

CI Triggers & Tools

​

Fix CI Playbook

​

Build Your Own CI Workflows

​

A note on rollout

​

Vault

• Store API keys and credentials securely.
• Control which secrets are available per use case or per agent.
• Ask the agent to execute commands that reference secrets via environment variables.

​

MCP Support (Beta - Controlled Rollout)

​

Persistent Agent Memory

​

Dedicated Explore Agent

​

Parallel Multi-Agent Execution

​

New Skills in overcut-playbooks

• Pull the latest version of the repository.
• Use your preferred local agent.
• Leverage the new skills to generate or refine structured playbooks.

​

Workflow Versioning

• Draft vs Published - edits are saved as drafts. Only published versions run, so you can safely work without affecting live workflows.
• Publish with messages - when you publish, add a short message describing what changed, just like a commit.
• Full history - see every version of a workflow, who published it, and when.
• Safe rollbacks - restore any previous version with one click.
• Unpublished changes view - a new page shows all workflows with pending changes across the project, including visual diffs.
• Discard changes - reset your draft back to the last published version.

• Edit your workflow as usual; it auto-saves as a draft.
• When ready, click Publish and add a message.
• View history anytime from the workflow page.