Get Orchid
Self-Hosted • Free for Developers

Every call your agents make. Recorded. Replayable.

The agent in your app stack went off the rails again. Stop digging through logs. Orchid captures the full run at the wire — every prompt, every tool call, every dollar. Inspect it step by step. Replay it offline.

Get OrchidSee How It Works

Spin up the local demo

API_KEY="orchid_demo_8675309" docker run --pull=always -d --name orchid-proxy-demo \ -p 4320:4320 -p 4321:4321 \ -e ORCHID_API_KEY=$API_KEY \ -e ORCHID_DEMO=1 \ -e ORCHID_RETENTION_DAYS=0 \ ghcr.io/mario-guerra/orchid-proxy

Once running, open http://localhost:4321 in your browser and enter the API key orchid_demo_8675309 to explore the pre-seeded demo trace.

localhost:4320/session/spacex_investment
provider:vertex status:2xx
#1 IntentSchema200
gemini-2.5-flash1.30s
#2 o3-mini200
o3-mini4.90s
#3 serpapi.com200
Google Search4.60s
#4 UncertaintyDecision200
gemini-2.5-flash2.70s
#5 serpapi.com200
Google Search17.90s
#6 UncertaintyDecision200
gemini-2.5-flash1.60s
Provider
serpapi
Status
200 OK
Latency
4.60s
Tokens
-- / --
{
"search_parameters": {
"engine": "google",
"q": "SpaceX IPO detailed investment analysis and risk factors expert opinions",
"location": "Austin, Texas",
"google_domain": "google.com"
},
"organic_results": [
{
"position": 1,
"title": "SpaceX Share Valuation & Investment Risks - CNBC",
"link": "https://www.cnbc.com/spacex-valuation-risks"
}
]
}
STREAMING ACTIVE

Everything your agent did.
No changes to your code.

Orchid is the capture layer for AI agent traffic. It records every API call for forensics, replay, and cost attribution — your tools and agents reason on top of it.

Zero Instrumentation

Drop the heavy SDK wrappers and decorators. Orchid hooks natively into your transport layer to capture all traffic automatically—no telemetry boilerplate required.

Local-First Storage

Keep your proprietary prompts and PII safe. Everything is stored in a local database inside your network. Zero data egress by design — verify it yourself with a single firewall rule.

Framework Agnostic

Works perfectly with raw HTTP clients, LangChain, or custom code. Switch LLM providers or orchestration frameworks tomorrow without touching your observability layer.

AI-Assisted Debugging

When a run fails, ask your coding assistant why. Orchid's built-in MCP server lets tools like Cursor and Claude Code query the recorded traffic directly to find the bug for you.

Real-Time Cost Tracking

Every exchange gets USD cost attribution from a pricing engine you control — per session, per task, per step. The cost denominator for agent ROI, captured where it can't be instrumented wrong.

Deterministic Replay

Blocks outbound network traffic and serves recorded session fixtures during local development and testing. 100% offline, deterministic, and free.

Debugging AI shouldn't feel like archaeology.

Your AI app makes three LLM calls in a row, and the final answer is a hallucination. Was it a bad prompt? Did the AI lose the context? Did a tool return garbage? You can't tell from the final output — but you can tell from the wire. Here's how:

1

Proxy Traffic

Add the thin Orchid SDK — one import, zero rewrites. It patches the HTTP transport layer so every LLM call from standard client libraries routes through the proxy in capture mode automatically. Or skip the SDK and set headers manually from any language.

2

Store & Inspect

Every LLM request/response is captured into a local thread-safe SQLite database. Inspection runs via the Web Visualizer or MCP queries.

3

Replay & Test

Export portable JSON fixtures. Run offline test suites with replay mode — zero API cost, fully deterministic.

Your App
Capture Mode
Orchid ProxyPort 4320
LLM Providers
SQLite
Request
Response
Stored to disk
Get Orchid

Deterministic Testing

Kill flaky tests with replay mode

In replay mode, the proxy blocks all outbound traffic. Requests are matched against recorded responses by semantic hash and served instantly. No API calls, no costs, fully deterministic.

Without Orchid

  • →Tests hit live APIs — slow, expensive, rate-limited
  • →Non-deterministic outputs cause flaky assertions
  • →Manual mock maintenance per model update
  • →CI costs scale linearly with test count

With Orchid

  • →Saved fixtures serve responses locally instantly
  • →Identical response every run, fully deterministic
  • →Re-record fixtures after making agent changes
  • →CI LLM spend: $0.00

See replay prove itself. Run a full multi-agent pipeline — OpenAI, Anthropic, and Vertex AI — entirely offline from a recorded fixture. No API keys, no cost, about 5 minutes.

Run the Replay Demo
Your App
Replay Mode
Orchid ProxyX-Orchid-Mode: replay
LLM ProvidersOFFLINE
SQLiteRecorded fixtures
$0.00
API Cost
100%
Deterministic
Request (to proxy)
Response (from storage)
Live APIs blocked

AI-Assisted Debugging

When a complex agentic workflow fails, digging through logs is a nightmare. Orchid comes with a built-in MCP server that exposes your recorded traces directly to your IDE.

  • Ask Cursor or Claude Code: "Why did the last run fail?"
  • Your agent queries the local trace DB automatically.
  • It finds the hallucination, explains it, and writes the fix.
Claude Code

> Querying Orchid for last session...

✓ Found session: 4 exchanges

> Analyzing payloads...

The agent failed because the search tool returned a 404 error on exchange #3. I will update the tool's URL endpoint.

gpt-4o$0.042
claude-3-5-sonnet$0.018
tool_call: web_search$0.000
Total Session Cost$0.060

Know What Every Agent Task Actually Costs

Provider invoices give you a monthly total. Orchid gives you the cost of each run, each task, each step — calculated at the wire, the one place attribution can't be instrumented wrong. When leadership asks what your agents cost per completed task, you'll have the number.

  • Per-Task Rollups: Know exactly which run cost you what — no surprise bills, no month-end archaeology.
  • Custom Pricing Engine: Bring your own JSON pricing map for any model or custom provider.
  • Local Only: All calculations happen inside your proxy. No SaaS subscriptions required.

Runs Local. Stays Local.

Orchid runs entirely inside your infrastructure. No telemetry, no external cloud.

The Architecture

Local-First

  • 100% Offline: Run via Docker in your environment.
  • Zero Data Leaks: Session recordings never leave your network.
  • No SaaS Tax: Stop paying per-event fees to monitor your own application.
Get Orchid

The Builder

Mario Guerra

  • 18 years shipping developer tools and AI platforms.
  • Drove TypeSpec adoption across 30% of Azure at Microsoft.
  • Onboarded 150+ customers to AI hardware at Qualcomm.
Learn More
For Teams in Regulated Environments

Need a system of record for your agents?

When an auditor — or your CFO — asks what your agents did and what it cost, Orchid is the record you answer from: every exchange captured at the wire, stored in your infrastructure, with per-task cost attribution built in. I'm working directly with a small number of design partners to shape what compliance-grade agent records look like. If this is a pain point for you, let's talk.

Talk to the Builder

See what your agents are actually doing.

Start recording and replaying your AI agent traffic locally in minutes.

Get Orchid