Getting Started
Start simple. Begin with 2–3 toolkits based on your primary workflows. Don’t try to set everything up at once — add tools as you discover you need them.
Choose the right client. Different clients suit different use cases:
| Client | Best for |
|---|
| Civic Chat | Setting up your account, configuring toolkits, getting started fast — no setup required |
| Claude.ai | General use with a familiar interface (MCP support requires Pro plan or above) |
| Claude Desktop | Power users who want offline capabilities |
| Cursor / Windsurf | Developers working inside their IDE |
Start with Civic Chat to test and configure tools before committing to a desktop client setup.
Be explicit in prompts. Vague prompts cause the AI to explore multiple paths, wasting tokens. Specify which tools, which data, and which criteria you want.
❌ "Check my stuff"
✅ "Get my GitHub notifications from the last 24 hours"
✅ "Show all Jira tickets assigned to me in 'In Progress' status"
Token Optimization
Every tool listed in a toolkit appears in every LLM prompt. Fewer tools = fewer tokens = faster, more accurate responses.
Use post-processors
Post-processors transform tool responses before they reach the AI, often cutting token usage dramatically:
| Transform | Typical reduction |
|---|
| HTML → Markdown | 40–70% |
| JSON → CSV | 30–50% |
| Remove unused fields | 10–30% |
| Truncate long text | Up to 80% |
Ask the AI: “What post-processors are available? Add HTML-to-Markdown for my Confluence server.”
Pre-populate parameters
If a tool always runs against the same repository, database, or calendar, preset the parameter so the LLM never has to specify it:
“Always use repository civic-mcp for GitHub tools in this toolkit.”
Create named clones of generic tools pre-configured for specific tasks. For example, clone list_commits → get-dev-branch-sha with branch and limit preset. The AI will call the right tool directly without prompting.
Filter ruthlessly
Keep toolkits to 3–5 tools maximum. Ask the AI to filter out tools that weren’t used in recent sessions.
Security Best Practices
What NOT to do
- Share OAuth tokens or API keys via email, Slack, or other messaging platforms
- Store credentials in code repositories or documentation
- Grant broader permissions than necessary — use read-only access where possible
- Skip OAuth flows by manually entering tokens
What counts as sensitive data
Be extra cautious with:
- PII — names, addresses, emails, phone numbers, SSNs
- Financial data — credit cards, bank accounts, transactions
- Credentials — passwords, API keys, OAuth tokens, session IDs
- Proprietary data — trade secrets, unreleased product plans
Set up account-level guardrails first
Apply broad policies at the account level before adding toolkit-specific rules:
- PII redaction for all tools
- Maximum result limits to prevent accidentally pulling millions of records
- Sensitive field filtering (passwords, tokens, secrets)
Then layer on stricter toolkit-level guardrails for sensitive workflows.
Token rotation
- OAuth tokens — configure auto-refresh; review every 30–90 days
- API keys — rotate manually every 90 days minimum
- If compromised — rotate immediately
Monthly security review checklist
- Review which services are still connected; revoke unused authorizations
- Check who has admin access in your organization
- Review active guardrails and verify they’re working
- Scan audit logs for unusual activity or failed authorization attempts
Authorization Management
OAuth vs API keys
Prefer OAuth whenever the service supports it:
| OAuth | API Keys |
|---|
| Security | Higher — scoped, refreshable | Lower — static, manual rotation |
| Setup | Requires browser flow | Simple |
| Recommended | Yes | Only when OAuth unavailable |
Upgrade authorization levels as you scale
- User-level — for individual testing
- Profile-level — for sharing credentials across a team toolkit
- Account-level — for organization-wide policies
Use the Civic Manager MCP to upgrade authorizations without re-authenticating.
Troubleshooting Prevention
Before starting a session, verify:
- Is your OAuth token still valid?
- Are you on the correct toolkit?
- Have you connected the tools you need?
- Are any guardrails blocking the operation you want?
Corporate networks — proxy servers often intercept HTTPS and TLS inspection causes failures. Work with IT to whitelist *.civic.com domains if you hit connection issues.
Microsoft changes — Azure/Entra permission changes can take time to sync. Wait a few minutes after updating permissions before retesting.
- Batch large operations — break into smaller chunks and use pagination
- Use cursors over page numbers — more efficient for databases, prevents duplicates
- Reduce scope when timing out — narrow the date range or add filters before retrying
- Refresh long-running sessions — Civic Chat can accumulate memory over time; reload the page for long work sessions
Quick Reference: Common Commands
Toolkit management
"Create a new toolkit called 'marketing'"
"Switch to the development toolkit"
"Show all my toolkits and their tools"
"Filter out tools I haven't used in this session"
Tool configuration
"Install the Slack server into this toolkit"
"Preset the GitHub repo parameter to 'my-project'"
"Clone the list-events tool as 'list-my-events' with my primary calendar preset"
"Alias the execute_query tool as 'query'"
Guardrails
"What guardrails are available?"
"Add PII redaction to all responses from the CRM server"
"Add HTML-to-Markdown post-processor for web scraping tools"
"Show me all active guardrails in this toolkit"
Authorization
"Show my active authorizations"
"Revoke GitHub authorization"
"Upgrade my Slack auth to profile level"
"Configure auto-refresh for my Google tokens"