Foxhound

Scale targets — 1B spans/day per customer · ingest p99 < 500 ms · trace query p95 < 2 s · dashboard query p95 < 300 ms. Architecture tour — 10-minute walk from SDK to dashboard. Scale readiness — live gap matrix · RFC index.

Source-available, self-hostable observability for AI agent fleets. Multi-tenant. OTel-compatible. Evaluators, Run Diff, Session Replay, cost budgets, SLA monitoring, behavior regression, CI quality gate.

Licensing notice This repository is publicly visible for reference and evaluation only. No permission is granted to use, copy, modify, distribute, sublicense, or sell this software without prior written permission from Caleb Love. Licensing inquiries: hello@caleb-love.com

See LICENSE for full terms.

What is Foxhound?

Foxhound is a source-available observability platform purpose-built for AI agent systems. Generic APM and logging tools do not model agent behavior well: tool calls, LLM invocations, branching workflows, replay, evaluation, and regression detection all get flattened into the wrong abstractions.

Foxhound gives you the missing layer:

• trace every agent run as a structured span tree
• compare executions to see exactly where behavior diverged
• evaluate responses with datasets, experiments, and LLM-as-judge workflows
• enforce budgets and SLAs before costs and latency drift silently
• audit and isolate tenant data for security-sensitive deployments

Not a generic logging product. AI agent observability.

Sandbox

The sandbox is a self-contained demo workspace that tells a realistic seven-day story: a returns resolution agent regresses after a prompt rollout, gets detected through traces and regressions, is investigated via run diff and session replay, and recovers through a dataset-backed experiment.

Run locally with no external dependencies:

pnpm install
pnpm dev:web:demo
# Open http://localhost:3001/sandbox

What you will see:

• Fleet Overview: platform posture, operator focus items, and trend charts (Recharts)
• Traces: 568 seeded traces with tree-structured span timelines, cost badges on LLM calls, and parent/child indentation
• Run Diff: side-by-side comparison showing exactly where two agent runs diverged
• Session Replay: step-through playback with play/pause, speed control, seek bar, and live state diff
• Regressions: behavior drift detection by agent version with linked investigation paths
• Experiments: dataset-backed evaluation with promotion-ready candidates
• Budgets and SLAs: per-agent cost budgets and latency/reliability targets
• Prompts: versioned prompt registry with label-based promotion and diff view
• Notifications: alert routing to Slack, webhook, GitHub, Linear, PagerDuty
• Settings: theme switcher, SDK ingestion simulator, CI quality gate demo
• Guided tour: first-visit walkthrough with keyboard navigation (auto-shows once, re-trigger from the Tour button)
• Command palette: Cmd+K for quick navigation across all surfaces
• Mobile responsive: hamburger menu with slide-over drawer on small screens

Core capabilities

Quickstart

1. Clone and install

git clone https://github.com/caleb-love/foxhound.git
cd foxhound
pnpm install

2. Start local infrastructure

docker compose -f docker-compose.dev.yml up -d
cp apps/api/.env.example apps/api/.env
pnpm --filter @foxhound/db db:migrate

3. Run the app

pnpm dev       # API
pnpm dev:web   # dashboard in another terminal

Default local endpoints:

• API: http://localhost:3000
• Web: http://localhost:3001

4. Send your first trace

Python

pip install foxhound-ai

from foxhound import FoxhoundClient

fox = FoxhoundClient(
    api_key="fh-...",
    endpoint="http://localhost:3000",
)

async with fox.trace(agent_id="support-agent") as tracer:
    span = tracer.start_span(name="tool:search", kind="tool_call")
    span.set_attribute("query", "refund policy")
    span.end()

TypeScript

npm install @foxhound-ai/sdk

import { FoxhoundClient } from "@foxhound-ai/sdk";

const fox = new FoxhoundClient({
  apiKey: process.env.FOXHOUND_API_KEY!,
  endpoint: "http://localhost:3000",
});

const trace = fox.trace({ agentId: "support-agent" });
const span = trace.startSpan({ name: "tool:search", kind: "tool_call" });
span.setAttribute("query", "refund policy");
span.end();
await trace.flush();

Tooling

Architecture

SDKs / OTLP  ->  API (Fastify)  ->  PostgreSQL
                  |
                  ->  Worker (BullMQ) -> Redis
                  ->  Web dashboard (Next.js)

Current monorepo layout:

apps/api/               Fastify REST API
apps/web/               Next.js dashboard
apps/worker/            BullMQ workers
packages/sdk/           TypeScript SDK
packages/sdk-py/        Python SDK
packages/cli/           CLI
packages/mcp-server/    MCP server
packages/api-client/    Typed API client
packages/db/            Drizzle schema + queries
packages/types/         Shared types
packages/billing/       Billing + entitlements
packages/notifications/ Notification delivery

Self-hosting

Foxhound is designed to run on your own infrastructure.

Minimum stack:

• PostgreSQL 16+
• Redis
• Node.js 20+
• pnpm 9+

Primary local/dev commands:

pnpm build
pnpm test
pnpm lint
pnpm typecheck

For API configuration, see:

• apps/api/.env.example

Security

Foxhound is built for security-sensitive, multi-tenant environments.

Current repo expectations include:

• API keys hashed at rest
• tenant-scoped data access via org_id
• JWT auth for user-facing operations
• audit logging for sensitive actions
• rate limiting and security headers on the API/web surfaces

If you discover a vulnerability, use GitHub security advisories or follow SECURITY.md if present.

Development

pnpm install
pnpm build
pnpm test
pnpm lint
pnpm typecheck
pnpm format

Useful dev commands:

pnpm dev       # API only
pnpm dev:web   # web only
pnpm dev:all   # API + web
pnpm --filter web verify   # canonical web verification lane

For the web preview surface, /sandbox is canonical and /demo is compatibility-only. See apps/web/README.md and docs/reference/sandbox-compatibility-retirement-checklist.md.

Contributing

Contributions are welcome.

• open issues for bugs and feature requests
• keep multi-tenant safety and security review standards high
• prefer small, verifiable changes over broad speculative refactors

See CONTRIBUTING.md for contribution conventions.

License

All rights reserved. This repository is public for reference and evaluation only. No permission is granted to use, copy, modify, distribute, sublicense, or sell this software without prior written permission. See LICENSE and contact hello@caleb-love.com for licensing inquiries.