Baxter

Autonomous financial research agent. Ask any question about a public company and get a sourced, cited answer backed by real data.

$ bun start
> Is NVDA overvalued? Do a DCF analysis.

Baxter classifies your query, builds a research plan, gathers data from
SEC EDGAR and the web, runs a 2-round bull/bear debate, validates with
a reflexion loop, and synthesizes a comprehensive answer with tables
and citations.

                              Query
                                |
                                v
                    +---------------------+
                    |    Orchestrator      |
                    |  (fast model classifies
                    |   complexity + skills)
                    +---------------------+
                       |       |       |
              simple   |       |       |   complex
                       |       |       |
               +-------+   medium  +---+--------+
               |               |                 |
               v               v                 v
          Researcher      Planner            Planner
               |               |                 |
               |          task graph         task graph
               |               |                 |
               |          Researcher         Researcher
               |        (parallel waves)   (parallel waves)
               |               |                 |
               |          Analyst            Analyst
               |               |          (2-round bull/bear
               |               |              debate)
               |               |                 |
               |               |           Validator
               |               |        (reflexion loop)
               |               |                 |
               v               v                 v
                    +---------------------+
                    |    Synthesizer       |
                    |  (streaming answer   |
                    |   with citations)    |
                    +---------------------+
                                |
                                v
                            Answer

---

Table of Contents

1. Overview
2. Prerequisites
3. How to Install
4. How to Run
5. Agents
6. Tools
7. Skills
8. How to Evaluate
9. How to Debug
10. Configuration
11. How to Contribute
12. License

---

Overview

Most financial AI tools are a single ReAct loop with a "research" label slapped on top. Baxter is a genuine multi-agent system: 6 specialized agents with distinct roles, a dependency-aware task graph, and dynamic routing that skips unnecessary work on simple queries while deploying the full pipeline on complex ones.

The interesting parts happen after the data is gathered. On complex queries, a bull analyst and a bear analyst independently build their cases, then each reads the other's argument and writes a rebuttal. That is two rounds of adversarial debate before the synthesizer ever sees the results. When the validator detects quality issues -- a data quality score below 0.7 or an outright error -- a reflexion loop kicks in. A fast model generates corrective guidance, only the affected research tasks re-run, the analyst re-processes, and the validator checks again. This adds zero cost when the data is already clean.

Everything streams through a terminal UI. You see which agent is running, watch facts accumulate in real time, and get a final answer with inline citations and formatted tables. Follow-up questions work naturally -- "What about their margins?" resolves to the company you were just discussing -- and facts persist across sessions in a local SQLite database.

Baxter works out of the box with zero paid data keys. SEC EDGAR provides free financial data for every public US company. Add optional API keys to unlock richer data sources and web search, but the core pipeline runs on a single LLM API key and nothing else.

Prerequisites

• Bun runtime (v1.0+)
• At least one LLM API key (e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY)

That's it. Financial data is available immediately via free SEC EDGAR. Web search and premium financial data are optional.

How to Install

Install Bun if you don't have it:

# macOS / Linux
curl -fsSL https://bun.sh/install | bash

# Windows
powershell -c "irm bun.sh/install.ps1 | iex"

# Homebrew
brew install oven-sh/bun/bun

Then clone and install dependencies:

git clone <repo-url> && cd baxter
bun install

Set at least one LLM API key:

export ANTHROPIC_API_KEY=sk-ant-...

How to Run

Launch the interactive TUI:

bun start

Or pass a query directly:

bun start "What is AAPL's PE ratio?"

The TUI supports these commands:

Command	Description
/help	Show available commands
/cost	Show session cost summary
/history	Show recent queries
/skills	List available research skills
/debug	Toggle workspace debug panel (or Ctrl+D)
/clear	Clear the conversation

Agents

Agent	Role	Model Tier
Orchestrator	Classify query complexity, match skills, route pipeline	Fast
Planner	Decompose query into research tasks with dependency graph	Primary
Researcher	Execute tools to gather data, runs tasks in parallel waves	Fast
Analyst	Financial analysis; 2-round iterative bull/bear debate on complex queries	Primary
Validator	Cross-check facts, flag inconsistencies; triggers reflexion if quality is low	Fast
Synthesizer	Generate final answer with citations and tables	Primary

The orchestrator uses the fast model to classify queries into three complexity tiers. Each tier maps to a fixed pipeline:

• Simple -- single data point lookups skip straight to the researcher and synthesizer.
• Medium -- the planner creates a task graph, the researcher executes it in parallel waves, and the analyst provides a neutral assessment.
• Complex -- the full pipeline runs: task graph, parallel research, 2-round bull/bear debate, validator with reflexion, and synthesis.

Iterative Debate

When bull/bear debate is enabled on complex queries:

1. Round 1 -- Bull and bear analysts run independently in parallel.
2. Round 2 -- Each analyst reads the opponent's Round 1 output and produces a targeted rebuttal.
3. The synthesizer receives all 4 perspectives for a balanced assessment.

Reflexion Loop

When the validator finds significant issues (data quality score below 0.7 or severity "error"):

1. A fast model generates corrective guidance identifying which research tasks need re-running.
2. Only the affected tasks re-execute with reflection context.
3. The analyst re-runs with updated facts.
4. The validator checks again (up to 3 rounds, configurable).

This adds zero overhead when results are already good -- the validator simply passes through.

Tools

The LLM sees 7 tools. An agentic router dispatches financial_data requests to 14 sub-tools internally, keeping the model's decision space small and focused.

Tool	Description	API Key Required
financial_data	Any financial data. Routes to 14 sub-tools: income statements, balance sheets, cash flows, prices, key metrics, SEC filings, insider trades, institutional holdings, analyst estimates, segment data, and 3 EDGAR endpoints.	No (EDGAR is free)
web_research	Search the web or scrape a URL. Supports Firecrawl, Exa, Perplexity, and Tavily backends (first available key wins).	Optional
web_fetch	Fetch and extract content from any URL using Readability. No API key required.	No
calculate_financial_ratios	PE, PB, ROE, ROA, margins, liquidity, and leverage ratios from raw data.	No
calculate_growth_rates	CAGR, YoY growth, and sequential growth rates.	No
calculate_statistics	Mean, median, standard deviation, and percentiles.	No
calculate_dcf	Full DCF valuation with terminal value and sensitivity analysis.	No

The financial_data tool uses a fast-model LLM call to route natural language like "AAPL income statements last 3 years" to the correct sub-tool. When no FINANCIAL_DATASETS_API_KEY is set, it falls back to free SEC EDGAR data automatically.

Skills

7 built-in research skills activate automatically based on trigger keywords in your query. Skills inject specialized prompts into the researcher and analyst, guiding tool usage and analytical frameworks.

Skill	Triggers
DCF Valuation	"dcf", "discounted cash flow", "intrinsic value", "fair value"
Earnings Analysis	"earnings", "quarterly results", "eps"
Comparable Analysis	"comparable", "comps", "peer comparison"
Portfolio Review	"portfolio", "holdings", "diversification"
Risk Assessment	"risk", "risk factors", "downside"
SEC Filing Analysis	"10-K", "10-Q", "SEC filing"
Sector Analysis	"sector", "industry analysis"

How to Evaluate

Baxter includes an evaluation suite with 20 financial Q&A pairs scored by an LLM judge:

# Run the full suite
bun run eval

# Run a single eval by ID
bun run eval simple-pe

# Run evals by category
bun run eval lookup

How to Debug

Workspace panel -- press Ctrl+D or type /debug to toggle a live view of the workspace: current facts, matched skills, task graph status, and validation issues.

Log levels -- set LOG_LEVEL to control verbosity:

LOG_LEVEL=debug bun start    # See routing decisions, tool calls, agent handoffs
LOG_LEVEL=trace bun start    # Everything, including raw LLM inputs/outputs

All logging is structured via Pino, so you can pipe output through pino-pretty or ship it to any log aggregator.

OpenTelemetry -- set OTEL_EXPORTER_OTLP_ENDPOINT to export traces covering the full pipeline, individual agent runs, and tool executions.

Configuration

LLM Providers

Set at least one API key. Baxter supports 8 providers:

Provider	Environment Variable	Example Models
Anthropic	ANTHROPIC_API_KEY	Claude Sonnet, Haiku, Opus
OpenAI	OPENAI_API_KEY	GPT-4o, GPT-4.1, o3-mini
Google	GOOGLE_GENERATIVE_AI_API_KEY	Gemini 2.5 Flash/Pro
xAI	XAI_API_KEY	Grok 3, Grok 4
DeepSeek	DEEPSEEK_API_KEY	DeepSeek Chat/Reasoner
Moonshot	MOONSHOT_API_KEY	Kimi K2
OpenRouter	OPENROUTER_API_KEY	Any model via OpenRouter
Ollama	OLLAMA_BASE_URL	Local models

Model Selection

Baxter uses two model tiers. The primary model handles reasoning-heavy tasks (planning, analysis, synthesis). The fast model handles classification, routing, validation, and tool dispatch.

PRIMARY_MODEL=anthropic:claude-sonnet-4-20250514
FAST_MODEL=anthropic:claude-haiku-4-5-20251001

Optional Settings

Variable	Default	Description
FINANCIAL_DATASETS_API_KEY	--	Unlock prices, insider trades, analyst estimates
FIRECRAWL_API_KEY	--	Web search + scrape (highest priority)
EXASEARCH_API_KEY	--	Exa neural search
PERPLEXITY_API_KEY	--	Perplexity Sonar search
TAVILY_API_KEY	--	Tavily search
BULL_BEAR_ENABLED	false	Enable 2-round iterative bull/bear debate
REFLEXION_ENABLED	true	Enable reflexion loop when validator finds issues
MAX_REFLEXION_ROUNDS	1	Max re-execution rounds (0-3)
OTEL_EXPORTER_OTLP_ENDPOINT	--	OpenTelemetry trace export URL
LOG_LEVEL	info	Log level (trace / debug / info / warn / error)
CACHE_TTL_SECONDS	3600	Tool result cache TTL
MAX_TOOL_CONCURRENCY	5	Max parallel tool executions

How to Contribute

1. Fork the repository and create a feature branch.
2. Make your changes. Run bun test and bun run lint before submitting.
3. Open a pull request with a clear description of what changed and why.

Development commands:

bun dev              # Run with --watch
bun test             # Run all tests
bun run lint         # Check with Biome
bun run lint:fix     # Auto-fix lint issues
bun run typecheck    # TypeScript type checking

License

MIT