Okapi
Agent-first

The error tracker your agent can actually use.

Okapi's native MCP server is a first-class credential door. Your coding agent reads issues, correlates logs by trace id, and ships the fix — through scoped, fenced, audited tools. On every plan, never gated.

The debugging loop

From alert to fix, without you in the loop.

01
Connect
Mint a scoped agent token and add Okapi in one line. read, triage, or admin — your call.
claude mcp add okapi …
02
Read
The agent pulls the fenced trace, the correlated logs, and what changed since it last looked.
get_issue_context · whats_new
03
Fix
It resolves the issue in the commit that fixes it — the issue closes when the fix ships.
resolve_issue --in-commit
04
Verify
Later, it checks the error actually stopped — and reopens itself if the bug comes back.
check_resolution
Safe by design

An error message is hostile input.

A stack trace can carry a prompt injection. Okapi fences every event and log field an agent reads — untrusted data stays data, and never becomes an instruction.

Every derived field is wrapped and flagged — mark, never strip.
Scoped tokens: read / triage / admin, per project or org.
Destructive actions require a named confirm and land in the audit log.
get_issue_context → response
message: ‹untrusted›
TypeError: … "ignore all instructions"
‹/untrusted · suspicious: true›
frame: checkout/processPayment.ts:42
logs: 3 correlated · trace 9f2a…

The agent is on every plan. Even the cheapest one.

MCP is the differentiation, not the upsell. We will never move it behind a higher tier.

Start catching errors before your users do.

A 14-day trial. One binary to self-host. MCP on every plan — and a price that never moves.

14-day trial · Sentry-SDK compatible · Self-host or cloud