synthpanel

Site: https://synthpanel.dev · Benchmark: https://synthbench.org

SynthPanel is the synthetic-population MCP server for AI agents.

When your agent needs to know what a representative slice of humans would say about a decision — pricing, naming, friction points, copy — it makes one tool call:

// MCP tool call
{
  "tool": "run_panel",
  "arguments": {
    "stimulus": "price at $49 or $79?",
    "decision_being_informed": "choosing launch tier price"
  }
}

// Response (same envelope from both `run_panel` and `panel run --output-format json`)
{
  "result_id": "result-20260510-abc123",
  "model": "claude-haiku-4-5",
  "synthesis": {
    "summary": "Cohort splits on $79; price is the dominant objection.",
    "themes": [...],
    "agreements": [...],
    "disagreements": [...],
    "surprises": [...],
    "recommendation": "..."
  },
  "rounds": [{ "name": "default", "results": [...], "synthesis": null }],
  "path": [],
  "warnings": [],
  "total_cost": "$0.0142",
  "total_usage": { "input_tokens": 4231, "output_tokens": 1102, ... },
  ...
}

You get a structured synthesis block — themes, agreements, disagreements, surprises, and a single recommendation line your agent can act on — plus the full per-panelist transcript under rounds[].results[] and per-turn cost telemetry. Bring your own LLM key — Claude, OpenAI, Gemini, or local. Drops into Claude Code, Cursor, Windsurf, LangChain, CrewAI, OpenAI Agents SDK.

Note (v1.0.5): the v1.0.0 panel_verdict envelope (headline, convergence, dissent_count, flags[], schema_version) is defined in schemas/v1.0.0.json and validated by the response gate, but is not yet emitted on the success path by either run_panel or panel run. Wiring is tracked separately; until it lands, consume the synthesis block above. Error responses already use the typed envelope (error_code, schema_version, retry_safe).

pip install synthpanel

Frozen contract: the v1.0.0 schema lives in the package at synthpanel/schemas/v1.0.0.json and is echoed in every response (schema_version: "1.0.0"). Field-by-field reference: docs/response-contract.md. Migrating from v0.12? docs/migration-v1.md. Methodology and inspectability: docs/methodology.md.

Zero-config inside any MCP host that speaks sampling (Claude Desktop, Claude Code, Cursor, Windsurf) — drop the config in and run a panel with no API key set. The host runs the model on your behalf, using its own subscription. Bring your own provider key when you want reproducibility, ensembles, or larger panels. Personas and instruments are plain YAML; every response is schema-validated with per-turn cost telemetry.

Why

Traditional focus groups cost $5,000-$15,000 and take weeks. Synthetic panels cost pennies and take seconds. They don't replace real user research, but they're excellent for:

Pre-screening survey instruments before spending budget on real participants
Rapid iteration on product names, copy, and positioning
Hypothesis generation across demographic segments
Concept testing at the speed of thought

Agent Quick Start

Wire the MCP server into your editor (see Use with Claude Code / Cursor / Windsurf / Zed below) and call the four research tools from agent code. Every panel-running call requires a decision_being_informed field (12–280 chars, single line) — the panel won't run without one.

// run_panel — full synthetic focus group
{
  "tool": "run_panel",
  "arguments": {
    "personas_pack": "general-consumer",
    "instrument_pack": "pricing-discovery",
    "decision_being_informed": "choosing launch tier price"
  }
}

// run_quick_poll — one question across personas
{
  "tool": "run_quick_poll",
  "arguments": {
    "question": "Which name feels most premium: Core, Plus, or Pro?",
    "personas_pack": "general-consumer",
    "decision_being_informed": "naming the paid tier"
  }
}

// extend_panel — append an ad-hoc follow-up round
{
  "tool": "extend_panel",
  "arguments": {
    "result_id": "result-20260503-abc123",
    "questions": ["What would you pay for this if it shipped tomorrow?"],
    "decision_being_informed": "validating the indie pricing ceiling"
  }
}

Read synthesis.recommendation for the headline call, synthesis.disagreements for dissent, and rounds[].results[] for the per-panelist transcript. See docs/response-contract.md for the full v1.0.0 envelope (note: panel_verdict fields like convergence and flags[] are defined but not yet emitted on the success path — see the note in the lede).

Human Operator Quick Start

Prefer a terminal? Same engine, CLI surface.

# Install from PyPI
pip install synthpanel

# For MCP server support (agent integration)
pip install synthpanel[mcp]

# Or install from source for the latest unreleased changes
pip install git+https://github.com/DataViking-Tech/SynthPanel.git@main

# Provide an API key (Claude, OpenAI, Gemini, xAI, or any OpenAI-compatible provider)
# — either export it in your shell:
export ANTHROPIC_API_KEY="sk-..."
# — or persist it once with `synthpanel login` (stored at
#   ~/.config/synthpanel/credentials.json, mode 0600):
synthpanel login --provider anthropic --api-key sk-...
synthpanel whoami   # see which providers have credentials available

# Run a single prompt
synthpanel prompt "What do you think of the name Traitprint for a career app?"

# Run a full panel
synthpanel panel run \
  --personas examples/personas.yaml \
  --instrument examples/survey.yaml

Works with

SynthPanel is MCP-native — it ships an MCP server, and every major agent framework now supports MCP as a first-class tool source. That means SynthPanel works out of the box with any framework that speaks MCP, with zero framework-specific wrapper packages to install. Runnable examples for each framework live in examples/integrations/.

Framework	Example	Bridge	One-line install
OpenAI Agents SDK	openai_agents.py	Built-in `MCPServerStdio`	`pip install openai-agents synthpanel[mcp]`
LlamaIndex	llamaindex_tool.py	`llama-index-tools-mcp`	`pip install llama-index-tools-mcp llama-index-llms-anthropic synthpanel[mcp]`
CrewAI	crewai_tool.py	`crewai-tools[mcp]`	`pip install "crewai-tools[mcp]" crewai synthpanel[mcp]`
LangChain	langchain_tool.py	`langchain-mcp-adapters`	`pip install langchain-mcp-adapters langchain-anthropic synthpanel[mcp]`
LangGraph	langchain_tool.py	`langchain-mcp-adapters`	`pip install langchain-mcp-adapters langgraph langchain-anthropic synthpanel[mcp]`
Microsoft Agent Framework 1.0	microsoft_agent.py	Built-in `MCPStdioTool`	`pip install agent-framework synthpanel[mcp]`
n8n	n8n_workflow.json	Built-in MCP Client tool	`pip install synthpanel[mcp]` on the n8n runner
LangChain via Composio	composio_langchain.py	`synth_panel.integrations.composio` (in-process, non-MCP)	`pip install composio composio_langchain langchain langchain-anthropic synthpanel`
CrewAI via Composio	composio_crewai.py	`synth_panel.integrations.composio` (in-process, non-MCP)	`pip install composio composio_crewai crewai synthpanel`

Also reaches Zapier MCP (30K+ actions), the VS Code AI Toolkit, Windsurf, Cursor, Zed, Claude Code, and Claude Desktop via the same MCP server — all clients in that list install SynthPanel with pip install synthpanel[mcp] and a one-line MCP config entry (see Use with Claude Code / Cursor / Windsurf / Zed).

Don't see your framework? MCP bridges are available for nearly every major agent framework. Start from examples/integrations/README.md — the pattern is identical in each case (point the client at synthpanel mcp-serve over stdio) — or file an issue so we can add a sibling example.

Run via Docker

A pre-built image is published to both GitHub Container Registry and Docker Hub on every tagged release. Use it for ephemeral or serverless invocation (Lambda, Cloud Run, GitHub Actions, n8n) where you'd rather spin up a container than pip-install.

# Pull (either registry works — same image, multi-arch: amd64 + arm64)
docker pull ghcr.io/dataviking-tech/synthpanel:latest
docker pull synthpanel/synthpanel:latest

# One-off prompt
docker run --rm \
  -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
  synthpanel/synthpanel \
  prompt "What makes a name feel trustworthy?"

# MCP server on stdio (default CMD — wire this into an agent's MCP config)
docker run --rm -i \
  -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
  synthpanel/synthpanel

# Panel run with a mounted instrument file
docker run --rm \
  -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
  -v "$PWD":/work -w /work \
  synthpanel/synthpanel \
  panel run --personas personas.yaml --instrument survey.yaml

The image's default CMD is mcp-serve, so omitting the command starts the MCP stdio server. Any synthpanel subcommand can be passed as arguments to override. Provider keys are read from environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY/GEMINI_API_KEY, XAI_API_KEY) — pass whichever your model requires.

Pin to a specific version (:0.11.0) in production rather than :latest.

Use as a Python Library

Everything the CLI and MCP server can do is also callable from Python. No subprocess, no extra install — just import and go.

from synth_panel import quick_poll, run_panel, run_prompt

# One-shot LLM call
reply = run_prompt("What makes a name feel trustworthy?")
print(reply.response, reply.cost)

# Ask a bundled persona pack a single question
poll = quick_poll(
    "Which pricing tier name feels most premium: Core, Plus, or Pro?",
    pack_id="general-consumer",
)
print(poll.synthesis["recommendation"])

# Run a full branching instrument against a bundled pack
panel = run_panel(
    pack_id="general-consumer",
    instrument_pack="pricing-discovery",
)
print(panel.path)         # e.g. ["discovery", "probe_pricing", "validation"]
print(panel.total_cost)

The package root exposes eight functions plus three typed return dataclasses — PromptResult, PollResult, PanelResult. Every result is dict-compatible (result["model"]) so code that used to consume the MCP JSON payload works unchanged.

Function	What it does
`run_prompt(prompt, *, model=...)`	Single LLM call — no personas
`quick_poll(question, pack_id=...)`	One question across a panel + synthesis
`run_panel(pack_id=..., instrument_pack=...)`	Full branching panel run
`extend_panel(result_id, questions)`	Append an ad-hoc follow-up round
`list_personas()` / `list_instruments()`	Discover installed packs
`list_panel_results()` / `get_panel_result(id)`	Reload saved results

Use this path when subprocess overhead hurts (Jupyter, serverless, CI) or when you want to wrap SynthPanel in a LangChain / LlamaIndex tool in three lines. See examples/sdk_usage.py for a runnable end-to-end walkthrough.

Typed extraction with Pydantic (1.0.4)

run_panel accepts a pydantic.BaseModel subclass directly as extract_schema=. SynthPanel generates the wire JSON Schema via model_json_schema() and runs model_validate on each panelist's extracted payload — validation failures surface per-response as extraction_validation_error (the panel still produces a usable result even when the LLM emits wire-valid JSON that violates a typed constraint, e.g. rating: 7 against a 1..5 Likert).

from pydantic import BaseModel, Field
from synth_panel import run_panel

class FeatureChoice(BaseModel):
    feature: str = Field(..., min_length=1)
    confidence: int = Field(..., ge=1, le=5)

result = run_panel(
    pack_id="developers",
    questions=[{"text": "Which feature should we ship first?"}],
    extract_schema=FeatureChoice,  # typed class accepted at the SDK boundary
)

for r in result.results:
    extracted = r["responses"][0].get("extraction")
    if extracted is not None:
        choice = FeatureChoice.model_validate(extracted)
        print(choice.feature, choice.confidence)

The same parameter still accepts a built-in name ("sentiment", "likert", "pick_one", …) or an inline JSON Schema dict — typed classes are an additional dispatch, not a replacement.

MCP Server (Agent Integration)

synthpanel ships an MCP server so AI agents can run synthetic focus groups as tool calls.

pip install synthpanel[mcp]
synthpanel mcp-serve

Add to your editor's MCP config (Claude Code, Cursor, Windsurf):

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Zero-config first run via sampling

No API key? No problem. When the invoking MCP client (Claude Desktop, Claude Code, Cursor, Windsurf) advertises the sampling capability, synthpanel falls back to asking the client to run the LLM completion on its behalf — using the client's own subscription. That means run_prompt and small run_quick_poll calls (up to 3 personas) work with zero env setup:

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"]
    }
  }
}

Sampling mode is great for first-touch UX and quick exploratory polls. For cross-provider ensembles, larger panels, and reproducible model versioning, set a provider key in env to graduate to BYOK. See docs/mcp.md#model-resolution-order for the full matrix of when sampling kicks in and which default model each provider key picks.

Tools (12)

Tool	Description
`run_prompt`	Send a single prompt to an LLM — no personas required
`run_panel`	Run a full synthetic focus group panel with parallel panelists and synthesis
`run_quick_poll`	Quick single-question poll across personas with synthesis
`extend_panel`	Append an ad-hoc follow-up round to a saved panel result
`list_persona_packs`	List all saved persona packs (bundled + user-saved)
`get_persona_pack`	Get a specific persona pack by ID
`save_persona_pack`	Save a persona pack for reuse
`list_instrument_packs`	List installed instrument packs (bundled + user-saved)
`get_instrument_pack`	Load an installed instrument pack by name
`save_instrument_pack`	Install an instrument pack with validation
`list_panel_results`	List all saved panel results
`get_panel_result`	Get a specific panel result with all rounds and synthesis

run_panel accepts an inline instrument dict or an instrument_pack name for v3 branching runs. extend_panel appends one ad-hoc round — it is not a re-entry into the v3 DAG. See docs/mcp.md for full tool schemas, resources, and prompt templates.

Use with Claude Code / Cursor / Windsurf / Zed

Copy the JSON snippet for your editor into the listed config file, set your API key, and restart the editor. synthpanel mcp-serve is launched on demand over stdio — no long-running process to manage.

Claude Code

Add to .mcp.json at your project root (or ~/.claude.json for all projects):

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Or install the bundled plugin (adds the /synthpanel-poll <question> slash command plus five skills — focus-group, name-test, concept-test, survey-prescreen, pricing-probe):

/plugin install synthpanel

Not using Claude Code, or prefer manual install? See docs/agent-skills.md for cp-based install steps and per-host guidance.

Cursor

Add to .cursor/mcp.json at your project root (or ~/.cursor/mcp.json for all projects):

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json (or open Settings → Windsurf Settings → MCP Servers → View Raw Config):

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Zed

Zed uses context_servers (not mcpServers). Add to ~/.config/zed/settings.json:

{
  "context_servers": {
    "synth_panel": {
      "source": "custom",
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Claude Desktop

Open Settings → Developer → Edit Config (or edit the file directly):

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "synth_panel": {
      "command": "synthpanel",
      "args": ["mcp-serve"],
      "env": { "ANTHROPIC_API_KEY": "sk-..." }
    }
  }
}

Restart Claude Desktop after editing.

Using a non-Anthropic provider? Swap ANTHROPIC_API_KEY for OPENAI_API_KEY, GEMINI_API_KEY, XAI_API_KEY, or OPENROUTER_API_KEY — see LLM Provider Support. The synthpanel binary must be on the editor's PATH; if you installed into a virtualenv, point command at its absolute path (e.g. /path/to/.venv/bin/synthpanel).

What You Get

============================================================
Persona: Sarah Chen (Product Manager, 34)
============================================================
  Q: What is the most frustrating part of your workflow?
  A: Version control on documents that aren't in a proper system...

  Cost: $0.0779

============================================================
Persona: Marcus Johnson (Small Business Owner, 52)
============================================================
  Q: What is the most frustrating part of your workflow?
  A: I'll send my manager a menu update in an email, she makes
     her changes, sends it back...

  Cost: $0.0761

============================================================
Total: estimated_cost=$0.2360

Each persona responds in character with distinct voice, concerns, and perspective. Cost is tracked and printed per-panelist and in aggregate.

Post-hoc Markdown report (`synthpanel report`)

Already have a saved panel result and want a readable share-out? Render the result to Markdown:

# Save a run first (prints a result ID you can reference later)
synthpanel panel run \
  --personas examples/personas.yaml \
  --instrument examples/survey.yaml \
  --save

# Render to stdout, either by result ID or path to a result JSON
synthpanel report <result-id>
synthpanel report path/to/result.json

# Write to a file
synthpanel report <result-id> -o report.md

Every rendered report opens with a mandatory synthetic-panel banner and closes with a matching footer so the output can't be mistaken for real-user research:

# Panel Report: <result-id>

> **Synthetic panel.** All responses below were generated by AI personas,
> not human respondents. Do not cite as user-research data.

...

_Generated by synthpanel report. Synthetic panel — AI-generated responses, not human data._

report is scoped to Markdown v1 — provenance, per-model rollup, persona summary, synthesis, and failure stats. HTML and chart rendering are deferred to v2. A report optional-deps extra exists (pip install synthpanel[report]) and installs cleanly, but is currently empty — it is a forward-compat slot for v2 HTML dependencies and adds nothing today.

Full spec: specs/sp-viz-layer/.

Defining Personas

# personas.yaml
personas:
  - name: Sarah Chen
    age: 34
    occupation: Product Manager
    background: >
      Works at a mid-size SaaS company. 8 years in tech,
      previously a software engineer. Manages a team of 5.
    personality_traits:
      - analytical
      - pragmatic
      - detail-oriented

  - name: Marcus Johnson
    age: 52
    occupation: Small Business Owner
    background: >
      Runs a family-owned restaurant chain with 3 locations.
      Not tech-savvy but recognizes the need for digital tools.
    personality_traits:
      - practical
      - skeptical of technology
      - values personal relationships

Merging Additional Personas (`--personas-merge`)

Layer extra personas onto a base file (or exported pack) without editing the original. --personas-merge is repeatable and appends in order; a later persona whose name matches an earlier one replaces it in place:

synthpanel panel run \
  --personas developer.yaml \
  --personas-merge contrarian.yaml \
  --personas-merge intern.yaml \
  --instrument pricing-discovery

Community packs

You can import persona packs straight from GitHub:

# Listed in the synthpanel registry — import by gh: URI
synthpanel pack import gh:dataviking-tech/example-pack

# Not yet in the registry — opt in explicitly
synthpanel pack import gh:alice/my-pack --unverified

The registry itself is an open, PR-based index at DataViking-Tech/synthpanel-registry. See docs/registry.md for the full reference — supported URL forms, cache + offline behavior, collision rules, and the flow for publishing your own pack.

Defining Instruments

# survey.yaml
instrument:
  questions:
    - text: >
        What is the most frustrating part of your current
        workflow when collaborating with others?
      response_schema:
        type: text
      follow_ups:
        - "Can you describe a specific recent example?"

    - text: >
        If you could fix one thing about how you work with
        technology daily, what would it be?
      response_schema:
        type: text

Adaptive Research (0.5.0): Branching Instruments

A v3 instrument is a small DAG of rounds. After each round, a routing predicate decides which round runs next based on the synthesizer's themes and recommendation. The panel chooses its own probe path — no human in the loop, no hand-coded conditional flows.

# The Show HN demo: ~$0.20, one command, the panel decides
# whether to dig into pain, pricing, or alternatives.
synthpanel panel run \
  --personas examples/personas.yaml \
  --instrument pricing-discovery

pricing-discovery is one of five bundled v3 packs (pricing-discovery, name-test, feature-prioritization, landing-page-comprehension, churn-diagnosis). List them with synthpanel instruments list.

The output now carries a path array recording the routing decisions that actually fired:

discovery -> probe[themes contains price] -> probe_pricing -> validation

Render the DAG of any instrument:

synthpanel instruments graph pricing-discovery --format mermaid

Predicate Reference

route_when is a list of clauses evaluated in order. The first matching clause wins; an else clause is mandatory as the last entry.

route_when:
  - if: { field: themes, op: contains, value: price }
    goto: probe_pricing
  - if: { field: recommendation, op: matches, value: "(?i)wait|delay" }
    goto: probe_objections
  - else: __end__

Field	Source
`themes`	`SynthesisResult.themes` (list, substring match)
`recommendation`	`SynthesisResult.recommendation` (string)
`disagreements`, `agreements`, `surprises`	`SynthesisResult` (lists)
`summary`	`SynthesisResult.summary` (string)

Op	Meaning
`contains`	Substring match against any list entry or the string
`equals`	Exact string match
`matches`	Python regex match (use `(?i)` for case-insensitive)

The reserved target __end__ terminates the run; the path so far feeds final synthesis.

Theme Matching: The R3 Caveat

Predicates match against the synthesizer's exact theme strings.

themes contains price only fires if the synthesizer actually emitted a theme containing the substring price. LLM synthesizers paraphrase — "cost concerns" or "sticker shock" will not match. The bundled packs mitigate this with a comment block at the top of the instrument that hints at the canonical theme tags the synthesizer should prefer:

# Synthesizer guidance: when emitting `themes`, prefer the short
# canonical tags below so route_when predicates match reliably:
#   - "pain"   (workflow pain, frustration, broken status quo)
#   - "price"  (cost concerns, perceived value, sticker shock)
#   - "alternative" (existing tools, workarounds, competitors)

When you author your own v3 packs, always add a similar tag-hint block. The synthesizer reads it and tends to use the canonical tags; your contains predicates then route reliably. If you skip this step, expect routes to silently fall through to else because the synthesizer's prose theme labels won't match your predicate values.

Prefer short, lowercase, single-token tags (price, pain, confusion) over long phrases. contains does substring matching, so price will also match pricing, priced, etc.

`instruments` Subcommand

synthpanel instruments list                       # bundled + installed packs
synthpanel instruments show pricing-discovery     # full YAML body
synthpanel instruments install ./my-pack.yaml     # add a local pack
synthpanel instruments graph pricing-discovery    # text DAG
synthpanel instruments graph pricing-discovery \
  --format mermaid                                 # mermaid flowchart

The unified instrument resolver (used by panel run --instrument) accepts either a YAML path or an installed pack name, so you can iterate on a local file and then install it once it's stable.

Examples

The examples/ directory ships a persona pack plus one instrument per format (v1 flat, v2 linear, v3 branching). Start from examples/README.md for the full index and annotated walkthroughs — including two v3 branching patterns (demographic segmentation and A/B concept testing) you can adapt to your own studies.

LLM Provider Support

synthpanel works with any LLM provider. Set the appropriate environment variable:

Provider	Environment Variable	Model Flag
Anthropic (Claude)	`ANTHROPIC_API_KEY`	`--model sonnet`
Google (Gemini)	`GOOGLE_API_KEY` or `GEMINI_API_KEY`	`--model gemini`
OpenAI	`OPENAI_API_KEY`	`--model gpt-4o`
OpenRouter	`OPENROUTER_API_KEY`	`--model openrouter/anthropic/claude-haiku-4-5`
xAI (Grok)	`XAI_API_KEY`	`--model grok`
Any OpenAI-compatible	`OPENAI_API_KEY` + `OPENAI_BASE_URL`	`--model <model-id>`

# Use Claude (default)
synthpanel panel run --personas p.yaml --instrument s.yaml

# Use GPT-4o
synthpanel panel run --personas p.yaml --instrument s.yaml --model gpt-4o

# Use a local model via Ollama
OPENAI_BASE_URL=http://localhost:11434/v1 \
synthpanel panel run --personas p.yaml --instrument s.yaml --model llama3

Model Aliases

synthpanel ships with short aliases (sonnet, opus, haiku, grok, gemini, gemini-pro) that map to canonical model identifiers. You can override or extend these without changing code:

Resolution order (highest priority wins):

SYNTHPANEL_MODEL_ALIASES env var — JSON string of alias→model pairs
~/.synthpanel/aliases.yaml — YAML file
Hardcoded defaults — built into the package

# Override via env var (JSON)
export SYNTHPANEL_MODEL_ALIASES='{"sonnet": "claude-sonnet-4-6-20250414", "fast": "claude-haiku-4-5-20251001"}'
synthpanel prompt "Hello" --model fast

# ~/.synthpanel/aliases.yaml
aliases:
  fast: claude-haiku-4-5-20251001
  smart: claude-opus-4-6
  sonnet: claude-sonnet-4-6-20250414

Env var entries override file entries, which override hardcoded defaults. Aliases from all tiers are merged, so you only need to specify the ones you want to add or change.

SynthBench-driven model picker

--best-model-for consults the SynthBench leaderboard and uses the top-ranked model for a topic or dataset:

synthpanel panel run --personas p.yaml --instrument s.yaml \
    --best-model-for "Economy & Work"

The leaderboard is cached for 24 hours at ~/.synthpanel/synthbench-cache.json. See docs/recommended-models.md for the full rules, offline behaviour, and a use-case → top-model table.

Architecture

synthpanel is a research harness, not an LLM wrapper. It orchestrates the research workflow:

personas.yaml ──┐
                 ├──> Orchestrator ──> Panelist 1 ──> LLM ──> Response
instrument.yaml ─┘                 ├──> Panelist 2 ──> LLM ──> Response
                                   └──> Panelist N ──> LLM ──> Response
                                                                  │
                                              Aggregated Report <──┘

Components

Module	Purpose
`llm/`	Provider-agnostic LLM client (Anthropic, Google, OpenAI, xAI)
`runtime.py`	Agent session loop (turns, tool calls, compaction)
`orchestrator.py`	Parallel panelist execution with worker state tracking
`structured/`	Schema-validated responses via tool-use forcing
`cost.py`	Token tracking, model-specific pricing, budget enforcement
`persistence.py`	Session save/load/fork (JSON + JSONL)
`plugins/`	Manifest-based extension system with lifecycle hooks
`mcp/`	MCP server for agent-native invocation (stdio transport)
`cli/`	CLI framework with slash commands, output formatting

Design Principles

Minimal dependencies — Python 3.10+ with httpx for HTTP and pyyaml for YAML parsing. Optional: mcp for the MCP server
Agent-native — invoke from your terminal or from an AI agent's MCP tool call
Provider agnostic — swap LLMs without changing research definitions
Cost transparent — every API call is tracked and priced
Reproducible — same personas + same instrument = comparable output
Structured by default — responses conform to declared schemas

Output Formats

# Human-readable (default)
synthpanel panel run --personas p.yaml --instrument s.yaml

# JSON (pipe to jq, store in database)
synthpanel panel run --personas p.yaml --instrument s.yaml --output-format json

# NDJSON (streaming, one event per line)
synthpanel panel run --personas p.yaml --instrument s.yaml --output-format ndjson

Budget Control

# Set a dollar budget for the panel
synthpanel panel run --personas p.yaml --instrument s.yaml --config budget.yaml

The cost tracker enforces soft budget limits — the current panelist completes, but no new panelists start if the budget is exceeded.

Persona Prompt Template Variants

The templates/ directory contains four prompt template variants for benchmarking how persona prompt construction affects response quality:

Template	File	Fields	Purpose
Current	`templates/current.txt`	name, age, occupation, background, personality_traits	Control — documents the default prompt style
Demo	`templates/demo.txt`	name, age, occupation, education_level, income_bracket, urban_rural, political_leaning, background	Demographic-enriched — adds SubPOP/OpinionsQA stratification axes
Values	`templates/values.txt`	name, age, occupation, background, core_values, decision_style	Values-enriched — adds belief and decision-making context
Minimal	`templates/minimal.txt`	name, age, occupation	Ablation control — tests how much narrative matters

Usage:

synthpanel panel run --personas personas.yaml --instrument survey.yaml --prompt-template templates/demo.txt

Templates use Python format-string syntax ({field_name}). Missing persona fields are left as literal {field_name} in the output.

Reproducibility (`--seed`)

Pass --seed N to panel run for reproducible sampling on providers that honor the seed parameter (OpenAI, Gemini, xAI, OpenRouter):

synthpanel panel run --seed 42 --personas p.yaml --instrument s.yaml

What synthpanel can promise:

Forwards the seed to providers that support it.
Records the seed in the run's metadata.parameters.seed and in the checkpoint fingerprint, so a --resume run with a different seed fails loudly instead of silently mixing samples.

What synthpanel cannot promise:

Anthropic's Messages API has no seed parameter. When --seed is set on a Claude model, synthpanel logs a single warning per provider and proceeds without determinism. Use --temperature 0 for closer-to- deterministic Claude output, but expect drift across model versions.
Even on supporting providers, "seeded" sampling is best-effort: model serving infrastructure and silent server-side updates can still shift outputs between runs.

--seed is for new runs you want to be reproducible. To replay a previously-cached run exactly, use synthpanel panel run --resume <run-id> — that path serves cached responses verbatim and is independent of --seed. See docs/reproducibility.md for the full picture.

Methodology Notes

Synthetic research is useful for exploration, hypothesis generation, and rapid iteration. It is not a replacement for talking to real humans.

Known limitations:

Synthetic responses tend to cluster around means
LLMs exhibit sycophancy (tendency to please)
Cultural and demographic representation has blind spots
Higher-order correlations between variables are poorly replicated

Use synthpanel to pre-screen and iterate, then validate with real participants.

Multi-Model Ensemble (0.7.0)

Run the same panel through multiple models and blend their response distributions for higher-fidelity results. SynthBench experiments show 3-model ensembles improve human-parity scores by +5-7 points over any single model.

# Run 3 models with equal weights and blend distributions
synthpanel panel run \
  --models haiku:0.33,gemini:0.33,gpt-4o-mini:0.34 \
  --blend \
  --personas personas.yaml \
  --instrument survey.yaml

# Each persona is interviewed by all 3 models independently.
# The --blend flag averages response distributions across models,
# producing more representative synthetic survey data.

The blended output includes per-model distributions and the weighted ensemble distribution, letting you inspect both individual model perspectives and the consensus view.

Running at Scale

For panels of 500 to 10,000+ panelists, synthpanel can track response-distribution convergence live via Jensen-Shannon divergence and optionally auto-stop once every bounded (Likert / yes-no / pick-one / enum) question has stabilized. The post-run JSON gains a top-level convergence section showing the smallest n at which each question converged, so you can confidently run smaller next time.

synthpanel panel run \
  --personas large-panel.yaml \
  --instrument pricing-discovery \
  --convergence-check-every 20 \
  --auto-stop \
  --output-format json > result.json

jq '.convergence.overall_converged_at, .convergence.auto_stopped' result.json
# 473
# true

See docs/convergence.md for methodology, tuning, and the optional --convergence-baseline flag that compares your run against a real-human baseline from SynthBench (install via pip install 'synthpanel[convergence]').

Submitting calibrated runs to SynthBench

A calibrated panel run (one made with --calibrate-against DATASET:QUESTION) produces a per-question JSD against a known human distribution — the same score the SynthBench leaderboard tracks. Add --submit-to-synthbench to upload the result automatically when the run completes:

export SYNTHBENCH_API_KEY=sk_synthbench_...    # mint at synthbench.org/account
synthpanel panel run \
  --personas examples/personas.yaml \
  --instrument happiness-probe \
  --calibrate-against gss:HAPPY \
  --convergence-check-every 20 \
  --submit-to-synthbench

First use shows a one-screen privacy notice (recorded at ~/.synthpanel/synthbench-consent.json so subsequent runs don't re-prompt; pass --yes for CI). Submission failures are warned-but-non-fatal so a slow SynthBench cannot fail your panel run. See docs/synthbench-integration.md for the full privacy model, what does and does not get uploaded, and the failure modes.

Versions

Version	Highlights
0.11.0	`sp-i2ub` scaled-orchestration epic: panelist-level checkpointing with `--resume <run-id>` and auto-checkpoint on SIGINT/SIGTERM (every K=25 panelists), `--max-cost <USD>` mid-run projected-total cost gate that halts gracefully with valid partial JSON, and valid-partial-JSON discipline on every abort path (rate-exhaustion, SIGINT, cost-gate, panelist failure) with `run_invalid: true` + specific `abort_reason` and exit code 2; 6-bug loudness sweep turning silent failures loud across alias parse, synthesis partial payload, MCP `extend_panel`, condition evaluator, orchestrator follow-up exceptions, and the `test_aliases` fixture; auto-tag now fails loudly on unlabeled release PRs; `pip-audit` ignores CVE-2026-3219 in pip 26.0.1
0.10.0	`synthpanel report` post-hoc Markdown renderer for saved panel results (behind `[report]` extra); inline SynthBench calibration via `panel run --calibrate-against DATASET:QUESTION` with auto-derived `pick_one` schema and `per_question[key].calibration` sub-object wire format; decentralized pack registry — `pack import gh:<user>/<repo>` with `--unverified`, `pack search`, `pack list --registry`, 24h cache + offline fallback; optional `version:` field on persona packs with opt-in shadow warning
0.9.9	`--synthesis-strategy=auto` now routes to map-reduce on context overflow instead of hard-failing; OpenRouter alias resolution tightened for sub-1¢ local-table sanity checks; `--personas-merge` warns (or errors via `--personas-merge-on-collision`) on name collisions with bundled packs; version single-sourced from `src/synth_panel/__version__.py` with templated site render
0.9.8	Fail-loud synthesis (context-overflow pre-flight + structured `synthesis_error`), per-question map-reduce synthesis (`--synthesis-strategy=single\|map-reduce\|auto`), response-schema validation with deterministic distributions for bounded question types, rate-limit-aware client (`--max-concurrent`, `--rate-limit-rps`), live convergence telemetry + `--auto-stop`, 4 new bundled persona packs (`job-seekers`, `recruiters-talent`, `product-research`, `ai-eval-buyers`) raising shipped personas 24 → 84, `/synthpanel-poll` slash command
0.9.7	Provider-reported cost is authoritative — when a provider returns `usage.cost` (e.g. OpenRouter), that value is recorded verbatim instead of being recomputed from a local pricing table; `pricing_fallback` warning surfaced when a model falls through to `DEFAULT_PRICING`; ensemble rounding no longer silently drops low-weight models
0.9.5	Fail-fast on unsubstituted `{placeholder}` variables in instruments/personas, `--personas-merge PATH` for layered persona packs, `--dry-run` pre-run preview, `run_invalid` flag on degenerate runs, MCP BYOK detection routes through the credentials store
0.9.4	`synthpanel login` / `logout` / `whoami` credential-store CLI; MCP recognises `OPENROUTER_API_KEY` as BYOK and picks a sensible default; Docker images on GHCR + Docker Hub multi-arch; MCP sampling fallback for `run_prompt` and `run_quick_poll`
0.9.0	First release post-public-flip. Repo renamed to `SynthPanel` (PyPI name `synthpanel` unchanged)
0.8.0	`lookup_pricing_by_provider` public helper for synthbench-format provider strings; multi-question CLI cost shape symmetry (`total_cost` / `panelist_cost` / `total_usage` / `panelist_usage`)
0.7.0	Multi-model ensemble blending (`--blend`), OpenRouter provider support, temperature/top_p controls, prompt template customization
0.6.0	`--models` weighted model spec, `--temperature`/`--top_p` flags, persona prompt templates, pack generation, domain templates, MCP improvements
0.5.0	v3 branching instruments, router predicates, 5 bundled instrument packs, `instruments` subcommand (list/show/install/graph), MCP `*_instrument_pack` tools, rounds-shaped panel output, `extend_panel` ad-hoc round tool
0.4.0	`--var KEY=VALUE` and `--vars-file` for instrument templates, fail-loud on all-provider errors, default `--model` respects available credentials, `pack show <id>` alias, publish workflow fix
0.3.0	Structured output via tool-use forcing, cost tracking, MCP server (stdio), persona-pack persistence

See CHANGELOG.md for detailed release notes.

Contributing

See CONTRIBUTING.md for development setup, testing, and how to submit changes.

Benchmarked on SynthBench

synthpanel's ability to produce representative synthetic respondents is independently measured by SynthBench, an open benchmark for synthetic survey quality.

Want proof it works? Browse the leaderboard — ensemble blending of 3 models hits SPS 0.90 (90% human parity).
Got a great configuration? Submit your scores and compare against baselines.
Contributing an adapter? Heavy PRs with substantial behavior changes benefit from SynthBench results — reviewers can evaluate empirical quality, not just code. See docs/adapter-guide.md for the full adapter workflow.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 523 Commits
.claude-plugin		.claude-plugin
.devcontainer		.devcontainer
.github		.github
commands		commands
docs		docs
examples		examples
scripts		scripts
site		site
skills		skills
specs		specs
src/synth_panel		src/synth_panel
templates		templates
tests		tests
.dockerignore		.dockerignore
.env.example		.env.example
.gc-merge-strategy		.gc-merge-strategy
.gitignore		.gitignore
.pre-commit-config.yaml		.pre-commit-config.yaml
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
SPEC.md		SPEC.md
glama.json		glama.json
pyproject.toml		pyproject.toml
renovate.json		renovate.json
server.json		server.json

Folders and files

Latest commit

History

Repository files navigation

synthpanel

Why

Agent Quick Start

Human Operator Quick Start

Works with

Run via Docker

Use as a Python Library

Typed extraction with Pydantic (1.0.4)

MCP Server (Agent Integration)

Zero-config first run via sampling

Tools (12)

Use with Claude Code / Cursor / Windsurf / Zed

What You Get

Post-hoc Markdown report (synthpanel report)

Defining Personas

Merging Additional Personas (--personas-merge)

Community packs

Defining Instruments

Adaptive Research (0.5.0): Branching Instruments

Predicate Reference

Theme Matching: The R3 Caveat

instruments Subcommand

Examples

LLM Provider Support

Model Aliases

SynthBench-driven model picker

Architecture

Components

Design Principles

Output Formats

Budget Control

Persona Prompt Template Variants

Reproducibility (--seed)

Methodology Notes

Multi-Model Ensemble (0.7.0)

Running at Scale

Submitting calibrated runs to SynthBench

Versions

Contributing

Benchmarked on SynthBench

License

About

Topics

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 22

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

Post-hoc Markdown report (`synthpanel report`)

Merging Additional Personas (`--personas-merge`)

`instruments` Subcommand

Reproducibility (`--seed`)

Packages