ai-hist (TypeScript SDK and MCP server)

TypeScript reader for the ai-hist history database, plus an MCP server for searching local AI coding-agent history from Claude Code, Codex, Cursor, Grok, and Agent Relay.

The SDK uses sql.js, so it has no native build step. It reads the same SQLite file that ai-hist sync writes, or falls back to scanning local Claude/Codex/Cursor/Grok JSONL files and compacted trajectory JSON when the database is missing.

Install

npm install ai-hist

Quick start

import { openAiHist, resumeCommand } from 'ai-hist';

const hist = await openAiHist(); // uses $AI_HIST_DB or ~/.local/share/ai-hist/ai-history.db
try {
  const sessions = hist.listSessions({ limit: 20 });
  for (const s of sessions) {
    console.log(`[${s.source}] ${s.firstPrompt.slice(0, 60)} (${s.promptCount} prompts)`);
    console.log('  resume:', resumeCommand(s));
  }
} finally {
  hist.close();
}

To require the ai-hist SQLite database instead of JSONL fallback:

const hist = await openAiHist({ fallback: 'error' });

MCP server

After the package is published, run the local stdio MCP server with:

npx -y ai-hist-mcp

From a local checkout:

npm install
npm run build
node dist/mcp-server.js

The MCP server exposes tools for search, recent history, session lookup, temporal context, evidence packing, stats, trajectory search, and task WHY lookup over stdio. It runs on the user's machine and uses the same data-opening behavior as the SDK: SQLite first, then local fallback scanning.

To expose only one project through MCP, pass a project scope when launching the server. Exact project matches and child paths are included.

npx -y ai-hist-mcp --project .
npx -y ai-hist-mcp --project /path/to/project

Contract tools:

• search_history(query, limit?)
• recent_entries(limit?, project?)
• get_session(session_id)
• get_context(id)
• stats()
• search_trajectories(query, limit?)
• why_for_task(query)

API

openAiHist(opts?: {
  dbPath?: string;
  projectScope?: string;
  fallback?: 'jsonl' | 'error';
}): Promise<AiHist>

hist.close(): void
hist.dbPath: string
hist.sourceKind: 'sqlite' | 'jsonl'
hist.projectScope: string | undefined

hist.recent(opts?): HistoryEntry[]            // newest prompts first
hist.listSessions(opts?): SessionSummary[]    // grouped by session_id, last activity DESC
hist.getSession(sessionId): HistoryEntry[]    // all prompts in a session, oldest first
hist.getEntry(id): HistoryEntry | null
hist.getInTimeWindow(timestampMs, windowMs): HistoryEntry[]
hist.search(query, opts?): HistoryEntry[]     // literal substring search, recent matches first
hist.searchTrajectories(query, opts?): TrajectoryEntry[]
hist.whyForTask(query): TrajectoryEntry | null
hist.stats(): Stats                           // counts + date range

All list-style methods accept { source?, project?, limit?, beforeMs? }. beforeMs is the cursor for paginating older results.

resumeCommand(entry): string | null           // shell command per source; null for relay
defaultDbPath(): string                       // resolve env / OS default

Trajectories

ai-hist indexes compacted per-run trajectory JSON files as the decision WHY. Set TRAJECTORY_ROOT to an explicit root; the scanner reads:

$TRAJECTORY_ROOT/**/compacted/*.json

Without TRAJECTORY_ROOT, default discovery scans:

~/Projects/**/.trajectories/**/compacted/*.json

The runtime contract is one JSON file per completed run:

{
  "id": "run-id",
  "version": 1,
  "personaId": "planner",
  "projectId": "agent-workforce",
  "task": { "title": "Task title", "description": "Task description" },
  "status": "completed",
  "startedAt": "2026-06-06T10:00:00.000Z",
  "completedAt": "2026-06-06T10:05:00.000Z",
  "decisions": [
    {
      "question": "What should we do?",
      "chosen": "Chosen option",
      "reasoning": "Why this option won",
      "alternatives": ["Other option"]
    }
  ],
  "retrospective": {
    "summary": "What happened",
    "approach": "How the work was done",
    "learnings": ["What to carry forward"],
    "confidence": 0.8
  }
}

Schema

The canonical ai-hist SQLite schema is:

CREATE TABLE history (
  id INTEGER PRIMARY KEY AUTOINCREMENT,
  source TEXT NOT NULL,
  session_id TEXT,
  project TEXT,
  prompt TEXT NOT NULL,
  prompt_hash TEXT,
  timestamp_ms INTEGER NOT NULL,
  UNIQUE(source, timestamp_ms, prompt)
);

CREATE VIRTUAL TABLE history_fts USING fts5(prompt, project, content='history', content_rowid='id');

Trajectory sync also creates a structured trajectories table and inserts each per-run compact file into history with source='trajectory', so general history search and WHY-specific lookup both work.

License

MIT