chat

Query the keri.host knowledge base for spec-grounded answers about KERI, CESR, and ACDC.

When to Activate

Reviewing KERI protocol implementations for spec compliance
Evaluating architecture decisions involving KERI/CESR/ACDC
Answering questions about KERI specification details
Stress-testing designs against the specification

How to Query

Use the ask_keri_chat MCP tool. It maintains conversation history automatically across calls.

ask_keri_chat({ question: "What is pre-rotation in KERI?" })

For follow-up questions, just call again — history is preserved:

ask_keri_chat({ question: "How does that differ from delegation?" })

To start a fresh topic, reset the conversation:

ask_keri_chat({ question: "Explain CESR encoding", reset_history: true })

Curl fallback

If the MCP server is unavailable, query the API directly:

QUESTION='<your question>'
RESULT=$(curl -sN https://chat.keri.host/api/chat \
  -H 'Content-Type: application/json' \
  -d "{\"message\":\"$QUESTION\",\"history\":[]}" \
  | grep '^data: ' | sed 's/^data: //' \
  | jq -s '{
      answer: [.[] | select(.type=="chunk") | .text] | join(""),
      citations: ([.[] | select(.type=="citations") | .data] | add // [])
    }')
echo "$RESULT"

Attachments (curl only)

The request body accepts an optional attachments array for including file content with a question:

{
  "message": "Summarize this document",
  "history": [],
  "attachments": [
    {"name": "file.txt", "type": "text/plain", "content": "<base64-encoded content>"}
  ]
}

Each attachment has name (filename), type (MIME type), and content (base64-encoded).

SSE Event Types

The backend emits newline-delimited SSE events (data: {JSON}\n\n):

Type	Shape	Description
`chunk`	`{type: "chunk", text: "..."}`	Incremental answer text
`citations`	`{type: "citations", data: [{number, content, source}, ...]}`	Sent after answer completes
`done`	`{type: "done"}`	Stream end signal
`error`	`{type: "error", error, code, detail}`	Error (see codes below)

Error codes

BAD_REQUEST — missing message field
MODEL_ACCESS_REQUIRED — Anthropic use case details not yet submitted in Bedrock console
MARKETPLACE_SUBSCRIPTION — Bedrock Marketplace subscription still provisioning
THROTTLED — rate-limited, retry after a moment
INTERNAL_ERROR — unexpected server error

Query Log

Every query and response is appended to ~/.claude/keri-chat-log.jsonl. Each line is a JSON object:

{"timestamp":"2026-02-16T20:30:00Z","question":"What is pre-rotation?","answer":"Pre-rotation is...","citations":[...]}

Review the log:

# Last 5 queries
tail -5 ~/.claude/keri-chat-log.jsonl | jq .

# Search by keyword
grep -i 'delegation' ~/.claude/keri-chat-log.jsonl | jq .

Personas

KERI Architect Reviewer

When reviewing code or designs:

Before making any spec claim, verify it via ask_keri_chat
Cite specific specification sections in your review
Identify deviations from the spec with exact references
Suggest corrections that align with spec language

Devil's Advocate

When stress-testing designs or assumptions:

Query ask_keri_chat to find edge cases in the spec
Challenge assumptions by citing counter-examples from the specification
Ask "what if" questions grounded in spec-defined failure modes
Verify that error handling covers all spec-defined error conditions

Hosted MCP Endpoint

Every KeriChat deployment includes a hosted MCP endpoint at /mcp. Users with a deployment can configure Claude Code to use it directly — no local MCP server install needed:

{
  "mcpServers": {
    "keri-chat": {
      "type": "url",
      "url": "https://<your-domain>/mcp"
    }
  }
}

The hosted endpoint exposes the same ask_keri_chat tool. Since it's stateless (Lambda), conversation history is passed explicitly via the history parameter rather than auto-accumulated.

Guidelines

Always query the API before making authoritative spec claims
Include citation numbers from the API response in your analysis
If the API returns "I don't have information about that", say so explicitly
Prefer specification language over paraphrasing when precision matters
Cross-reference multiple queries for complex topics