DEPLOY-TIME RISK INTELLIGENCE

The check between
merge and outage.

One SQL query across six tools on every pull request, so you ship knowing exactly what is about to break.

github · sentry · datadog · pagerduty · slack · linear

acme/checkout-service · PR #2847187ms
87/ 100CRITICAL
PagerDuty2 SEV1 incidents
Datadogerror budget 12%
Sentry142 errors / 30d
Slack5 active threads
GitHubexpert: @lena-petrova
Linear1 linked ticket

Every outage was predictable. The signals were just in six different tabs.

Sentryknew the file had been throwing for a week.
PagerDutyhad three past incidents on the same path.
Datadogshowed the error budget already critical.
Slackhad an open thread about refactoring it.

No human checks all six tools on every pull request. Blast Radius does, in one query.

ONE QUERY, SIX SOURCES

Coral turns six tool integrations into one SQL statement.

coral sql
-- pr_risk_assessment.sql
SELECT
pr.changed_files, -- github
incidents.count AS sev_count, -- pagerduty
slo.error_budget_remaining, -- datadog
errors.total_30d, -- sentry
threads.open_count, -- slack
experts.top_committer -- github
FROM github.files AS pr
JOIN pagerduty.incidents AS incidents USING (service)
JOIN datadog.slo_metrics AS slo USING (service)
JOIN sentry.issues AS errors USING (culprit)
WHERE pr.repo = $1 AND pr.pull_number = $2;
1 row187 ms6 sources joinedauth · pagination · retries handled below deck

No six API clients, no pagination loops, no rate-limit retries. Coral runs the join locally and hands back a single tabular result the agent can reason over.

Same answer. A fraction of the tokens.

Stitching six MCP tools together means paginating every response into the context window. One Coral query returns just the columns that matter.

SIX MCP TOOLS, STITCHED
77ktokens
6 round trips · paginated payloads · glue prompts
ONE CORAL QUERY
4ktokens
1 statement · only the columns asked for
~19xfewer tokens per analysis, which is why an analysis costs a fraction of a cent.

A five-node agent runs on every pull request.

01deepseek-v4
Planner
Picks the Coral query for this PR.
02SQL
Coral
Joins six sources into one result.
03deterministic
Scorer
Risk 0–100 from fixed weights. No LLM.
04deepseek-v4
Narrator
Writes the plain-English PR comment.
05kimi-k2
Verifier
Checks every claim against source data.
THE OUTPUT

The agent posts a comment on the PR, pages on-call if it is critical, and streams to the dashboard. Every number is traced in Langfuse and verified against the source.

blast-radius-ci commented · now
87 / 100 CRITICALdo not merge

This rework touches a historically unstable path.

  • • SEV1 token-refresh outage in March on these files.
  • • Datadog error budget at 12%, burn rate critical.
  • • Suggested reviewer: @lena-petrova (211 commits here).

The six tools your incident already touched.

Connect them once. Coral queries them live on every PR. Nothing is copied or cached off your machine.

GitHubGitHub
SentrySentry
DatadogDatadog
PagerDutyPagerDuty
SlackSlack
LinearLinear

built on Coral · Fireworks AI · LangGraph · Langfuse · n8n