RelevanceAI · claude · Jul 31, 2026
diff --git a/changelog.mdx b/changelog.mdx
@@ -4,31 +4,47 @@
 mode: "wide"
 ---
 
+<Update label="July 31, 2026" description="Relevance MCP gains run-only mode and Chat role support">
+
+New Features
+
+## Relevance MCP gains run-only mode and Chat role support
+
+Two new access controls have been added to Relevance MCP, available to all users.
+
+In **run-only mode**, users generate MCP credentials that restrict the connected client to listing and executing agents only — the create, publish, and edit agent endpoints are blocked for that token. For **Chat role support**, users assigned the Chat role — previously unable to authenticate with MCP — can now connect and invoke any agent shared with them via the standard agent-sharing flow. The two controls are independent: a Chat-role user can authenticate with a run-only token, applying both restrictions simultaneously. Existing MCP credentials and full-access tokens are unaffected by these additions.
+
+**Example:** A team member with the Chat role connects their MCP client using a run-only token. They can list and invoke the "Summarise support ticket" agent shared with them, but cannot modify it, publish it, or create new agents.
+
+Select the **Run only** permission level when generating MCP credentials in your Relevance AI settings. Chat-role users follow the same MCP authentication flow. See the [MCP documentation](/integrations/mcp/mcp-server) for full setup instructions.
+
+</Update>
+
 <Update label="July 22, 2026" description="Concurrency visibility for project and organization admins">
 
 ## Concurrency visibility for project and organization admins

 ![Concurrency panel showing a seven-day usage chart with queued runs stacked in red above the limit line, 20 of 20 slots in use with 6 tasks waiting, and a live capacity by project bar with other projects anonymized](/images/changelog/changelog-concurrency-visibility.png)

 Concurrency monitoring on the Analytics page now covers queued runs, live slot counts, and a per-project capacity breakdown.

 Queued runs now stack in red above the limit line on the usage chart, live counters show slots in use and tasks waiting, and a per-project bar attributes running slots to the projects holding them. Between them you can see whether you're at the limit, how much is queued behind it, and which project is consuming the capacity. Project admins see the same charts with other projects' names and task details anonymized.

 Project admins find the chart on the Analytics page; organization admins can also view it on the Plan & Billing page.

 </Update>

 <Update label="June 29, 2026" description="Eval cost breakdowns and Workforce evaluations">

 ## Eval cost breakdowns and Workforce evaluations

 ![Eval run cost breakdown panel showing total credits and actions, with per-component rows for Scenario Runner, Agent Execution, and Checks broken down by AI Judge model and Tool Usage](/images/changelog/changelog-eval-cost-breakdown.png)

 Eval runs now show exactly where their cost goes, itemized per check, and cover full Workforce runs including sub-agent and tool calls.

 Each completed Eval run renders a cost breakdown panel that itemizes credit and action consumption per component. Checks account for 1 action per run across every check type, including LLM as Judge and Tool Usage, so the panel's totals reconcile to the exact source of every credit and action. Workforce runs that fan out to sub-agents or call tools record and evaluate each of those interactions, the same as individual Agent runs.

 To see the breakdown, open any completed Eval run; the panel appears alongside the results. See the [Evals documentation](/build/agents/build-your-agent/evals) for full details.

 </Update>

@@ -66,7 +82,7 @@

 ![Microsoft Teams trigger settings showing the new Outreach replies only toggle enabled](/images/changelog/changelog-teams-outreach.png)

 Microsoft Teams triggers now support an **Outreach replies only** mode that limits responses to direct replies on the Agent's own outbound DMs.

 The trigger distinguishes between inbound replies (replies to messages your Agent sent) and channel mentions. In outreach-only mode, only inbound replies route to the Agent — channel chatter doesn't trigger a run.

@@ -120,7 +136,7 @@

 You can now manage public sharing settings for all agents, workforces, and tools from one central location!

 Project-Level Public Agent Controls give admins the power to disable public cloning and embeddability across their entire workspace with a single setting. When you disable these features at the project level, they’re automatically turned off for all agents, workforces, and tools—and individual users can’t re-enable them.

 ➡️ Centralized control – Manage security settings for all assets from one place

@@ -168,7 +184,7 @@

 ![Agent Performance Observability](/images/changelog/changelog-observability.webp)

 You can now track how your agents perform in real-world scenarios with comprehensive production monitoring!  Just like software development has build → test → monitor stages, agent development now has the same complete lifecycle. After building your agents and testing them with Evals, you need visibility into production performance at scale.

 ➡️ **Automatic evaluation** – Set up LLM judges to continuously assess your agents on custom criteria like summary quality, hallucination detection, and workflow execution

@@ -180,7 +196,7 @@

 With Agent Performance Observability, you can catch performance degradations immediately—like when a model update changes behavior or agents start hallucinating—instead of discovering issues weeks later.

 🔐 **Permissions required:** Available exclusively for select Enterprise customers. Evals delivers the advanced testing capabilities needed for mission-critical AI deployments. This feature is not available on non-Enterprise plans.

 To access Agent Performance Observability, enterprise users can go to their agent’s dashboard and select the “Monitor” tab, then navigate to the “Test Suites” section. If you’re on an Enterprise plan and don’t yet have access, contact your Relevance AI sales representative to register interest.

@@ -239,13 +255,13 @@

 </Update>

 <Update label="March 10, 2026" description="Monitor organization concurrency usage with detailed analytics and timeseries data">

 ![Monitor organization concurrency usage with detailed analytics and timeseries data](/images/changelog/changelog-concurrency.webp)

 **Org Concurrency Timeseries: Track Your Organization’s Performance Analytics**

 You can now monitor your organization’s concurrency usage and limits with detailed timeseries data!

 View comprehensive analytics directly from your organization settings page under the Plan & Billing section, giving you complete visibility into how your team uses Relevance AI’s resources.

@@ -255,9 +271,9 @@

 ➡️ Usage limit monitoring – Keep track of your current usage against your plan limits to avoid unexpected throttling

 ➡️ Visual timeseries charts – Easy-to-read graphs show your concurrency patterns, helping you identify optimization opportunities

 With Org Concurrency Timeseries, you can make data-driven decisions about your AI operations, ensuring optimal performance and resource allocation across your entire organization.

 🔐 Permissions required: Organization Admin access to view billing and usage analytics

@@ -493,11 +509,11 @@

 ![Claude Opus 4.6 model now available in Relevance AI](/images/changelog/68faec2d3ffbd34e9c9295b9-changelog-claude.webp)

 **Claude Opus 4.6: Anthropic's Most Advanced AI Model Now Available**

 You can now access Anthropic's flagship Claude Opus 4.6 model directly in the platform, with an optional 1M context window variant for handling massive documents!

 Claude Opus 4.6 delivers exceptional performance on complex reasoning, coding, and advanced agentic tasks. This latest version brings Anthropic's most powerful AI capabilities to your workflows, with both standard and extended context options to match your specific needs.

 ➡️ Standard 200K token context window – Process lengthy documents and conversations with ease

@@ -505,7 +521,7 @@

 ➡️ Adaptive thinking support – Tackle complex reasoning and multi-step problem solving

 ➡️ Advanced image processing – Work with jpg, jpeg, gif, webp, and png files

 ➡️ Exceptional output quality – Benefit from improved reasoning and response generation

@@ -549,15 +565,15 @@

 </Update>

 <Update label="February 9, 2026" description="Test and validate AI agents with Evals: Enterprise-grade agent testing">

 ![Test and validate AI agents with Evals](/images/changelog/698a6f9d99180607d34d642c-changelog-evals.webp)

 **Evals: Test and Validate Your AI Agents with Confidence**

 You can now thoroughly test your AI agents with our powerful Evals feature, ensuring they perform exactly as expected in real-world scenarios!

 Evals provides a structured testing framework for your AI agents, allowing you to create test suites with multiple scenarios to validate functionality. Create golden test examples, simulate user interactions, and evaluate agent performance without triggering actual tool executions or external actions.

 ➡️ Create organized test suites – Group related tests together for better organization and management

@@ -569,9 +585,9 @@

 ➡️ Coming soon: Tool simulations – Test workflows with sensitive or expensive tools without triggering actual executions

 With Evals, you can confidently deploy agents knowing they've been thoroughly tested against a variety of scenarios, ensuring reliability and consistent performance for your users.

 **Available exclusively for select Enterprise customers**, Evals delivers the advanced testing capabilities needed for mission-critical AI deployments. To access Evals, enterprise users can go to their agent's dashboard and select the "Monitor" tab, then navigate to the "Test Suites" section. If you’re on an Enterprise plan and don’t yet have access, contact your Relevance AI sales representative to register interest. This feature is not available on non-Enterprise plans.

 Start building more reliable AI agents with comprehensive enterprise-grade testing today!

@@ -651,11 +667,11 @@

 Workforce Trigger Controls give you granular management over your automation workflows. Quickly pause specific triggers when issues arise, cancel problematic workflows, or pause an entire workforce at once—then resume operations when you're ready.

 ➡️ Pause individual triggers – Temporarily stop specific automations without affecting others

 ➡️ Cancel active workflows – Immediately terminate problematic processes

 ➡️ Resume paused triggers – Restart automations when issues are resolved

 ➡️ Pause entire workforces – One-click solution to halt all triggers during maintenance

@@ -769,11 +785,11 @@

 ![Control your meeting agent in real time with chat commands](/images/changelog/695a95a2a931a05c425b7b72-changelog-meeting-bot-real-time-commands.webp)

 **Meeting Agent Realtime Commands: Control Your AI Agent During Live Meetings**

 You can now trigger your AI meeting agent with custom commands during live meetings without interrupting the flow of conversation!

 Meeting Agent Realtime Commands lets you interact with your meeting bot during meetings by simply typing trigger words in the chat. Your agent springs into action when it detects these keywords, performing tasks like searching information, controlling recordings, or creating meeting artifacts—all while your meeting continues uninterrupted.

 ➡️ Use custom trigger words – Define specific phrases that activate your agent during meetings

@@ -785,7 +801,7 @@

 ➡️ Extend with any Relevance AI capability – Connect to Notion, Slack, or other tools for seamless workflows

 With Meeting Agent Realtime Commands, you can enhance your meeting productivity by leveraging AI assistance exactly when you need it, without disrupting the conversation flow.

 To set up your own meeting agent with custom commands, create a new agent from scratch, add the "Relevance AI Meeting Bot" under Triggers, add and configure your desired meeting bot tools (such as 'Pause Recording' and 'Resume Recording') and more general tools (such as 'Google Search' and 'Slide Builder') and trigger words.