Agent Access · MCP

Give your AI agents production context.

An MCP server with eight tools — errors, incidents, likely causes, proposed fixes — behind scoped keys, six-category exposure controls, and regex redaction applied before anything reaches an agent.

Start free trial →Read the docs

agent session · Claude Codescoped key · redaction on

agent

Checkout errors spiked after the last deploy. Finding the likely cause.

◆ find_likely_cause incident: "INC-241"

errorgap · mcp

{
  "cause":      "deploy v2.14.1",
  "confidence": 0.78,
  "signals": [
    "backtrace_overlap: charge.rb",
    "time_correlation: 4m",
    "similar_incident: INC-187"
  ],
  "top_frame":  "app/services/checkout/charge.rb:81"
}

agent

Deploy v2.14.1 touched charge.rb — the top frame. Proposing a fix.

◆ propose_fix error: "grp_9f2c41"

How it works

Connect. Query. Control. Redact.

Your agent gets the same evidence your engineers get — through a controlled surface, not a scraped dashboard.

01 · Connect

One config block

Point Claude Code, Cursor, or any MCP client at the errorgap server with a scoped key. Revocable from the Agent Access page.

02 · Query

Eight tools

list_errors, get_error, list_incidents, get_incident, explain_incident, find_likely_cause, propose_fix, uptime_status — structured results, not screenshots.

03 · Control

Six exposure categories

Per-project toggles over six categories of data, with org-level defaults and locks. What’s off stays off, project by project.

04 · Redact

Before the agent sees it

Regex redaction rules with a live preview run on every response — emails, tokens, whatever you define — before data leaves errorgap.

Exposure controls

Org defaults. Project overrides. Locks.

Admins set the default exposure across the org and lock the categories that must never reach an agent. Projects can tighten further — never loosen past a lock.

Six data categories, individually toggled
Org-level locks override project settings

agent exposure · ox-coffeeorg defaults + locks

Error groups & backtracesproject

Incidents & likely causeproject

Log excerptsproject

Source contextproject

Uptime statusproject

User & request metadata🔒 org-locked

Redaction

Write the rule. Watch it work.

Redaction rules are regex with a live preview against real recent data — you see exactly what masks before you save. Rules apply to every MCP response, no exceptions.

redaction rules · live previewapplied before any agent sees data

rule /[\w.+-]+@[\w-]+\.[\w.]+/ → [email]

rule /tok_[A-Za-z0-9]+/ → [token]

before charge failed for sam@oxcoffee.com · tok_1OqX84jE

after charge failed for [email] · [token]

Deterministic first, AI on demand

The agent reads the same evidence.

find_likely_cause returns the deterministic correlation panel — the same signal-weighted confidence your engineers see, as structured data. Your agent reasons from evidence, not vibes.

Structured JSON results with signals and weights
propose_fix grounded in backtrace + deploy diff

likely cause · deterministicno AI credits used

▸deploy v2.14.1 · confidence 78%

backtrace overlap with deploy diff

+42

time correlation · 4 min gap

+24

similar prior incident · INC-187

+12

AI summary runs only when you ask

Connect

One block in your client config.

Scoped key, explicit scopes. Revoke it anytime from Agent Access.

// .mcp.json
{
  "mcpServers": {
    "errorgap": {
      "url": "https://app.errorgap.com/mcp",
      "headers": {
        "Authorization": "Bearer egmcp_…"
      },
      "scopes": ["errors:read", "incidents:read", "uptime:read"]
    }
  }
}

Related capabilities

Give your AI agents production context.

Connect. Query. Control. Redact.

One config block

Eight tools

Six exposure categories

Before the agent sees it

Org defaults. Project overrides. Locks.

Write the rule. Watch it work.

The agent reads the same evidence.

One block in your client config.

It all feeds the same incident.

Close the errorgap.

Give your AI agents production context.

Connect. Query. Control. Redact.

One config block

Eight tools

Six exposure categories

Before the agent sees it

Org defaults. Project overrides. Locks.

Write the rule. Watch it work.

The agent reads the same evidence.

One block in your client config.

It all feeds the same incident.

One object. All the evidence.

The line. The frame. The fix.

Wired into your tools.

Close the errorgap.