Your AI agent spent an afternoon patching a bug Friday's deploy already fixed. The code gave no sign the errors stopped. Our get-version MCP tool looks up a git SHA and shows when its errors first and last appeared, so the agent can check before patching. roll.bar/eu5ur8

Your AI agent can read the retry loop in your code, but it can't see how often that loop gives up in production. Our MCP server's list-occurrences tool gives the agent a timestamped record of each failure, so it can count them before editing the backoff. roll.bar/4LSWh5

LiteLLM's MCP endpoint accepted forged tokens, a flaw CISA lists as exploited (CVE-2026-59822). Upgrade to 1.84.0 or later. If you run our MCP server behind LiteLLM, a read-scope token means a forged session can read your errors but can't change them. roll.bar/XbmrtX

A three-line fix to a production bug can still eat an afternoon of finding the file and rerunning tests. Hand the next one to Resolve. It writes the fix, runs your tests in an isolated environment, and opens a PR for you to review. roll.bar/ad6rji

We're rolling out personal access tokens, so each teammate gets their own Rollbar API token. Each one carries that person's permissions and stops working when they're offboarded.

A Rollbar item groups repeated occurrences of one error. The list-occurrences tool in our MCP server returns each occurrence behind an item with its own timestamp, environment, and error detail, so your AI agent sees when and where they happened. roll.bar/Y5BiVF

ArcadeDB's MCP server let a user with read permission fetch the database's cluster token, and that token was enough to act as root (CVE-2026-67357, fixed in 26.7.3). Our MCP server's one write tool, update-item, fails with a read-scope token. roll.bar/6dqzV5

To diagnose a Rollbar item, your AI agent needs its latest occurrence. The get-item-details tool in our MCP server returns the item and that occurrence in one call, so the agent can start on the stack trace right away. roll.bar/IKeAQQ

When an error starts right after a release, your AI agent needs to know what shipped. The get-deployments tool in our MCP server lists a project's recent deploys, so the agent can line up each deploy against when errors started or stopped. roll.bar/GJTf67

mcp-atlassian checked its tool allowlist when listing tools but not when calling them, so a client limited to one tool could call any of its 73 (CVE-2026-77243). Eight of our MCP server's nine tools only read, and update-item fails on a read-scope token. roll.bar/h2e6lO

A stack trace shows your AI agent one error but not how often it happens. The get-top-items tool in our MCP server returns an environment's most frequent items from the last 24 hours, so your agent starts with what is breaking most. roll.bar/CN3yKQ

Rollbar can now run root cause analysis on an error before anyone opens it. Write one rule with an environment and a severity level. Rules check every new occurrence, so errors that were open before you wrote the rule get analyzed too. roll.bar/bnI1Tl

If your agent's permissions are enforced when it lists tools and not when it calls them, anything that knows a tool name can still reach the tool. Eight of our nine MCP tools only read. update-item writes, and a read-scope token fails there. roll.bar/YGrwOq

An Every Occurrence rule with no filter attached pages you for errors you already know about. The bill for that arrives during the next real incident, in the seconds everyone spends deciding whether this page is different. docs.rollbar.com/docs/notifi…

Before markets opened on Aug 1, 2012, Knight Capital's systems sent 97 automated emails about a router error. None were built to page anyone. The next 45 minutes cost $460M, per the SEC's order. In Rollbar you pick which errors page a human, and who. roll.bar/0B8mRl

Wiring an agent to real production errors is the easy half. Running the tests is where our Resolve beta still bites. Complex suites cannot always stand up in an ephemeral environment, and then you get a fix with no test evidence behind it. roll.bar/3GTRbt

An agent hands you a patch for a production error. Nothing has run it. A fix is only worth trusting once your test suite has executed against it and the results are on the pull request. That is what Rollbar Resolve ships. roll.bar/3GTRbt

Your agent gets the stack trace and guesses. Give it the 30 seconds before the throw instead: the click, the failed fetch, the retry, then the error. Most impossible production bugs stop being impossible once the sequence is in the prompt.