ThinWedge

ThinWedge is a local FP&A agent terminal for financial modeling, planning analysis, statistical-model workflows, and cost research. It runs as a terminal UI and CLI on your machine, keeps workspace state under your ThinWedge home directory, and connects to provider APIs with an OpenRouter-compatible API token.

Quick Start

Install ThinWedge globally from npm:

npm install -g @never2average-does-npm/cli

Authenticate before starting interactive mode. In a real terminal, thinwedge login prompts for the required OpenRouter-compatible API token and optional capability config such as Artificial Analysis, RunPod, AWS profile/region values, and Neon DB sandbox metadata:

thinwedge login
# Enter OpenRouter-compatible API key: ...
# ARTIFICIAL_ANALYSIS_API_KEY (...) [optional]: ...
# RUNPOD_API_KEY (...) [optional]: ...
# AWS_PROFILE (...) [optional]: ...
# AWS_REGION (...) [optional]: ...
# THINWEDGE_NEON_API_KEY (...) [optional]: ...
# THINWEDGE_NEON_PROJECT_ID (...) [optional]: ...

Or set the provider token in the environment first:

export OPENROUTER_API_KEY=...
thinwedge login

Then start the interactive terminal UI:

thinwedge

Run a one-shot CLI task without opening the TUI:

thinwedge exec "summarize this repository"

Or download a binary from the latest GitHub Release.

What It Is

ThinWedge is built for finance and operations teams that need agentic help inside a real repository or modeling workspace. The agent can inspect and edit files, run shell commands, maintain plans, manage long-running goals, coordinate side agents, and call domain tools for statistical modeling, training environments, LLM pricing, and cloud infrastructure cost analysis.

The product surface is intentionally local-first:

• thinwedge starts the interactive terminal UI.
• thinwedge exec runs a single non-interactive agent task.
• thinwedge login stores an OpenRouter-compatible API token locally.
• thinwedge db-sandbox configures and preflights DB sandbox providers.
• THINWEDGE_HOME controls where config, auth, logs, thread history, and local state live.

System Components

flowchart TB
    User[User terminal] --> CLI[ThinWedge CLI]
    CLI --> TUI[Interactive TUI]
    CLI --> Exec[One-shot exec mode]
    CLI --> Auth[Local auth and config]

    TUI --> Runtime[Agent runtime]
    Exec --> Runtime
    Auth --> Runtime

    Runtime --> Tools[Tool router]
    Runtime --> Store[Local state under THINWEDGE_HOME]
    Runtime --> Provider[OpenRouter-compatible provider API]

    Tools --> Shell[Shell and command execution]
    Tools --> Files[Filesystem read, edit, and patch tools]
    Tools --> MCP[MCP servers, plugins, and apps]
    Tools --> Finance[FP&A, model, and cost-analysis tools]

    Shell --> Sandbox[Sandbox and approval policy]
    Files --> Sandbox
    MCP --> Sandbox
    Finance --> Sandbox

Authentication

ThinWedge uses API-token authentication. Run thinwedge login before starting the interactive TUI so the agent can call the configured provider API. In a real terminal, thinwedge login behaves like an aws configure-style prompt: it asks for the required OpenRouter-compatible API token, then offers optional prompts for ARTIFICIAL_ANALYSIS_API_KEY, RUNPOD_API_KEY, AWS_PROFILE, AWS_REGION, THINWEDGE_NEON_API_KEY, and THINWEDGE_NEON_PROJECT_ID. The required provider token is stored in ThinWedge auth storage; optional capability values are written to THINWEDGE_HOME/.env, which ThinWedge loads on startup.

thinwedge login
# Enter OpenRouter-compatible API key: ...
# ARTIFICIAL_ANALYSIS_API_KEY (...) [optional]: ...
# RUNPOD_API_KEY (...) [optional]: ...
# AWS_PROFILE (...) [optional]: ...
# AWS_REGION (...) [optional]: ...
# THINWEDGE_NEON_API_KEY (...) [optional]: ...
# THINWEDGE_NEON_PROJECT_ID (...) [optional]: ...

You can also provide the provider token through the environment:

export OPENROUTER_API_KEY=...
thinwedge login

You can also pipe a token without leaving it in shell history:

printenv OPENROUTER_API_KEY | thinwedge login --with-api-key

The token is stored in local ThinWedge auth storage. Existing legacy managed-login credentials are not treated as a valid ThinWedge login by the TUI.

Agents

ThinWedge has one root conversation and can create additional agent work streams for parallel investigation or execution. In the TUI, slash commands expose the main coordination model:

• /goal starts, resumes, pauses, and monitors durable multi-turn goals.
• /plan keeps a live task plan visible while work is in progress.
• /agent and /subagents select or manage agent identities.
• /side opens side conversations for scoped work.
• /review switches into code-review behavior.
• /model changes the active model.
• /copy exports prior transcript cells to a timestamped folder with transcript.xlsx and transcript.docx; markdown and pasted tables are split into real workbook sheets.
• /status, /diff, /permissions, /mcp, /skills, /apps, and /plugins expose runtime, workspace, and integration state.

DB Sandbox Setup

Finance agents should not run migrations or experiments directly against production state. ThinWedge models this as a provider-first setup: validate a source provider, then hand agents only disposable database state.

Neon is the default path:

thinwedge db-sandbox configure --enabled --provider neon --neon-project-id <project-id> --branch-backend none
thinwedge db-sandbox preflight --dry-run
thinwedge db-sandbox preflight --provider neon

Ardent is optional. Use it as the branch backend only after the Neon or Postgres source checks pass:

thinwedge db-sandbox configure --branch-backend ardent
thinwedge ardent status --dry-run

The bottom-up source of truth is still the probe scripts:

scripts/probes/check-db-sandbox-readiness.sh --source-provider neon --branch-backend none

Logical Tool Tree

ThinWedge organizes tools in layers so the agent can reason about local execution, workspace state, and finance-specific systems without mixing those responsibilities.

ThinWedge
|-- Interfaces
|   |-- TUI: interactive chat, slash commands, goal display, diffs, approvals
|   |-- CLI: exec, login, logout, status, sandbox helpers, release/runtime commands
|   `-- App server: thread, account, goal, filesystem, and event APIs
|-- Agent runtime
|   |-- Root session, side sessions, subagents, and agent identity registry
|   |-- Thread store, rollout trace, persisted state, and resume support
|   |-- Goal engine: create, update, resume, pause, and continuation prompts
|   `-- Planning and collaboration modes
|-- Core tools
|   |-- Shell execution, stdin streaming, local filesystem reads, and patch apply
|   |-- Plan updates, permission requests, user input requests, and image viewing
|   |-- Agent orchestration: spawn, send, wait, resume, close, and list agents
|   |-- MCP resources, dynamic plugin tools, app connectors, and tool discovery
|   `-- Web and media tools when enabled by the runtime
|-- FP&A tools
|   |-- Statistical model jobs: training and batch inference
|   |-- Training environments: launch, attach, and stop sandboxed environments
|   |-- LLM cost tools: list, inspect, and compare model pricing
|   `-- Infrastructure cost tools: AWS product search, VM pricing, BOQ estimates,
|       cost-and-usage queries, forecasts, anomaly checks, and billing views
|-- Sandboxes
|   |-- Local process sandboxing and approval policy enforcement
|   |-- Python runtime support for pandas, numpy, matplotlib, and related analysis
|   `-- Statistical-model sandbox support for wandb-backed experiment tracking
`-- Persistence
    |-- Config files, auth storage, logs, and state DB under THINWEDGE_HOME
    |-- Thread and goal history
    `-- Local mirrors of remote model, environment, and cost-analysis state

FP&A Workflows

ThinWedge includes finance-oriented tool surfaces for:

• Building and editing financial models in a repository.
• Running statistical-model training or batch inference jobs.
• Managing remote training environments for model experiments.
• Comparing LLM model costs before selecting a provider or model.
• Estimating AWS infrastructure cost from service dimensions, price-list data, usage history, billing views, forecasts, and anomaly signals.
• Keeping long-running analytical tasks in /goal so the agent can resume and report progress cleanly.

Useful Docs

• Rust workspace
• Install and build
• Configuration
• Authentication
• Contributing
• Project logic
• CI and release flow
• Governance
• Code of Conduct
• Open source fund

License and Attribution

This repository is licensed under the Apache-2.0 License. ThinWedge includes software derived from OpenAI Codex, which is also licensed under Apache-2.0.

Apache-2.0 allows use, modification, distribution, sublicensing, and publication of derivative works, including an open-source CLI distribution. ThinWedge preserves the required license and attribution notices in NOTICE and THIRD_PARTY_NOTICES.md, and modified files are part of the ThinWedge derivative work.

ThinWedge is not affiliated with or endorsed by OpenAI. The Apache-2.0 license does not grant OpenAI trademark rights; OpenAI and Codex are referenced only for factual attribution to the upstream project.