Put real-world failures in front of every design decision
CollectSEA works where design happens: inside AI coding assistants over MCP, in pull requests, in scripts through the REST API, and as rule files your agents already read.
- Step 1Describe the systemYour assistant submits the design or selected app context through MCP or the API.
- Step 2Retrieve the precedentsCollectSEA matches behavior to source-linked incidents and explains the shared failure patterns.
- Step 3Carry the safeguards into codeUse the returned lessons or export agent instructions, then verify the controls in your app.
MCP server for AI assistants
A remote, read-only Model Context Protocol server (Streamable HTTP, no sign-in). Your assistant can check a design against real incidents while it plans, without you copying anything.
Run once in your terminal:
claude mcp add --transport http collectsea https://collectsea.com/mcpTools
analyze_existing_appReview observed app behavior and text files from a repository your assistant has inspected; return precedents and safeguards to verify.analyze_designProfile a proposed feature or agent workflow and return matched incidents, failure patterns and safeguards.search_incidentsKeyword search across documented incidents and advisories, with pagination.get_incidentFull record for one incident: facts, source, pattern, safeguards.list_failure_patternsThe 14 failure patterns with triggers, warning signs and engineering principles.get_safeguardsAGENTS.md-ready rules for a proposal, citing the incidents behind each rule.
Try it: ask your assistant "Before we build the refund agent, check CollectSEA for similar failures."
AGENTS.md and rule files
Every analysis exports to AGENTS.md, CLAUDE.md, Cursor rules or Copilot instructions. To make agents check designs on their own, add this block to your repository's AGENTS.md.
## Design checks (CollectSEA)
Before implementing a feature that moves money, deletes or migrates data,
changes permissions, runs generated code, pushes configuration widely, or lets
an AI agent act on content from outside the team, call the CollectSEA MCP tool
`analyze_design` with a two- or three-sentence description of the feature.
- Follow the safeguards it returns, or explain in the pull request why one
does not apply.
- Link the most relevant precedent (https://collectsea.com/incidents/...) in
the pull request description.
- If the MCP server is not available, POST the description to
https://collectsea.com/api/analyze instead.To export safeguards for a specific feature, run an analysis and choose a format in the export panel.
Pull-request design check
A GitHub Actions workflow that comments relevant incidents and safeguards when a pull request is labelled design-review. The description is sent as data; nothing is stored.
name: CollectSEA design check
on:
pull_request:
types: [labeled]
permissions:
contents: read
pull-requests: write
jobs:
precedents:
if: github.event.label.name == 'design-review'
runs-on: ubuntu-latest
steps:
- name: Find real incidents like this change
env:
PR_TITLE: ${{ github.event.pull_request.title }}
PR_BODY: ${{ github.event.pull_request.body }}
run: |
jq -n --arg p "$PR_TITLE. $PR_BODY" '{proposal: ($p | .[0:20000]), max_results: 5}' \
| curl -sSf https://collectsea.com/api/analyze \
-H 'Content-Type: application/json' \
-H 'X-CollectSEA-Client: github-action' -d @- > analysis.json
{
echo "### Real incidents like this change (CollectSEA)"
jq -r '.matches[] | "- [\(.incident.display_title)](https://collectsea.com/incidents/\(.incident.incident_id)) — \(.incident.display_summary)"' analysis.json
echo
echo "**Safeguards to check**"
jq -r '.top_lessons[] | "- [ ] **\(.title).** \(.statement)"' analysis.json
} > comment.md
- name: Comment on the pull request
env:
GH_TOKEN: ${{ github.token }}
run: gh pr comment ${{ github.event.pull_request.number }} --repo ${{ github.repository }} --body-file comment.mdREST API
JSON over HTTPS, no key required, 60 analyses per minute per IP. Full schema at /openapi.json.
curl -s https://collectsea.com/api/analyze \
-H 'Content-Type: application/json' \
-d '{"proposal": "An AI agent that reads support emails and issues refunds through Stripe", "max_results": 5}'See the API reference for every endpoint and field.
Custom GPTs and other agents
Any agent framework that can call HTTP tools can use CollectSEA.
- Custom GPT: in the GPT editor choose Actions → Import from URL and enter
https://collectsea.com/openapi.json. - Agent frameworks (LangGraph, OpenAI Agents SDK, Claude Agent SDK and others): register the MCP server URL above, or wrap
POST /api/analyzeas a tool. - LLM-friendly summary of the whole site: /llms.txt.