Franklin — The AI Agent with a Wallet

Franklin is an autonomous AI agent that runs directly in VS Code. It doesn't just write text — it spends USDC from a user-funded wallet to execute real work: coding, trading signals, and content generation.

Install the latest version from the VS Code Marketplace to get all new features and bug fixes.

Features

• Chat interface — Side panel chat powered by the Franklin agent engine
• 55+ models — Claude Opus/Sonnet/Haiku, Gemini, GPT, Grok, Kimi, DeepSeek, and more, switchable mid-session
• Image & video generation — /image and /video slash commands; the agent picks the best model per prompt and shows a cost preview before spending
• Smart routing — Auto/Eco/Premium profiles pick the right model per task automatically
• Extended thinking — Watch the model reason step-by-step, collapsible per turn
• Workflow timeline — Visual timeline of every tool call and action taken
• Proactive market data — detects tickers in your messages and prefetches live prices before the model responds
• Session history — Persistent sessions with search and auto-resume
• Doctor panel — one-click environment health check (wallet, gateway, Node.js version, MCP config)
• Usage insights — 30-day spend and session analytics dashboard
• Trading dashboard — launches Franklin's web panel with real-time portfolio and market data
• Chain switcher — toggle between Base and Solana payment wallets from the toolbar
• MCP support — connect external tools via MCP servers

Requirements

• VS Code 1.85+
• Node.js 20+ (for Trading Dashboard auto-launch)
• A funded wallet on Base or Solana to use paid models (free models require no funding)

Getting Started

1. Install the extension from the VS Code Marketplace
2. Open the Franklin panel from the Activity Bar
3. Run franklin setup in a terminal to create your wallet, or fund the address shown in the panel with USDC
4. Start chatting — free models work immediately, paid models activate once your wallet has USDC

Changelog

0.5.1

• Fix: "Saved vs Opus" no longer goes negative — when meaningful spend hit ImageGen / VideoGen, the savings widget showed numbers like $-8.79 because chat-only Opus baseline was being compared against total spend (chat + media). Now compared apples-to-apples (media counts on both sides; saved is the chat-side delta, clamped to >= 0). Mirror of upstream PR #36.

0.5.0

• Major core sync — extension now rides on Franklin core v3.10.0, picking up two months of upstream work since 0.4.5 (no more cherry-pick lag):
  • Detached background tasks (v3.10.0) — Detach capability spawns long-running work as a separate process; franklin task list / tail / wait / cancel to drive them
  • Skills system (v3.9.0, #34) — SKILL.md loader + registry + bundled budget-grill skill, the start of a plugin layer
  • First-class Wallet tool (#35) — agent can introspect balance / address without going through Bash
  • Smart timeout recovery (v3.8.41, #26) — stream timeouts no longer replay expensive previous turns; auto-continue
  • Default per-turn spend cap raised — $0.25 → $1.00 (v3.8.42, #28) → $2.00 (v3.9.1) so reference-image edits and multi-stage tasks don't trip on legitimate workflows
  • Status bar shows active chain (v3.9.1)
  • Kimi K2.6 alignment (v3.9.2), /model picker trimmed 28 → 23 (v3.9.3)
  • Roleplayed JSON tool-calls + V4 Flash / Omni metadata (v3.9.4)
  • Nemotron Omni reasoning-prose stripping (v3.9.5)
  • Reasoning-model TTFB defaults + long-task guidance (v3.9.6)
  • Plus proxy fixes (#31), grounding hardening, viem direct-dep cleanup, and more

0.4.5

• New mascot artwork on the empty state — bigger, transparent-background AI+coin themed Franklin pixel-art portrait (no more dark rounded-rectangle frame, blends seamlessly into any theme background)
• Per-turn spend cap is now configurable — new ⚙️ settings field "Per-turn spend cap (USD)" lets you raise the default $0.25 limit (or set 0 to disable) without editing source; mirrors the new franklin config set max-turn-spend-usd <n> CLI key
• Image-to-image timeout fix — gpt-image-2 reference-image edits no longer abort after 60s (the old shared budget couldn't cover reasoning-driven edits + base64 upload + x402 retry); image-to-image now gets 180s, text-to-image keeps 60s
• Synced with Franklin core: PR #19 (i2i timeout), PR #20 (configurable spend cap), PR #21 (README VS Code section)

0.4.3

• History replay shows generated media inline — closing and reopening a conversation now re-renders any images / videos as preview cards instead of dropping them
• "+" New chat truly resets the session — previously only the UI cleared while the agent kept the same sessionId, leaking tool guards (ImageGen disabled) and prior context into what looked like a new chat
• Auto-resume warm sessions — opening the panel within 24h of the last reply continues the same chat instead of fragmenting /history into pieces
• Workflow timeline polish — line and dots now blend like beads on a string (continuous track, hollow thinking dots filled with editor bg, soft pulsing shadow on active tools)
• Tool name shows model for ImageGen / VideoGen — e.g. ⚙ ImageGen gpt-image-2 so the picked model is visible at a glance; long prompts hidden to keep the timeline scannable
• Synced with Franklin core v3.8.36–v3.8.39: image-to-image (gpt-image-2) support, evaluator hardening for short user inputs, grounding retry forces tool use, longer LLM timeouts (45s/90s) for Sonnet / Opus reasoning

0.4.1

• Vision-capable models can now actually see images — Read on .png / .jpg / .gif / .webp returns the bytes inline as a tool_result content block; combined with the gateway-side fix (BlockRun gateway commit 6ac64da), Sonnet / Opus / GPT-4o / Gemini now describe user-provided images instead of hallucinating. (Closes #10 via core PR #11.)

0.4.0

• Settings popover — new ⚙️ button in the composer toolbar to toggle payment chain (Base / Solana) and set default image / video models in one place; Save dismisses the popover
• Inline edit diff cards — Edit / Write / MultiEdit show a green/red diff in the chat with Open and Revert buttons; Revert restores the file from an in-memory snapshot
• Local-path seed images for VideoGen — image_url on VideoGen now accepts local file paths (auto-converted to a data URL, capped at 4 MB)
• Routing-mode picker — when the active model is Auto / Eco / Premium the picker shows a single profile card with a toggle; click the toggle to exit routing mode and pick a specific model
• Preserve routing label across turns — Auto mode no longer flickers to the per-turn routed model name; the picker stays on "Auto" until you change it
• Model picker search + recent — fuzzy search bar at the top, 3 most recently used models pinned under "Recent"
• Inline media preview — generated images / videos from ImageGen / VideoGen appear as inline preview cards right below the tool result
• AskUser inline prompts — cost previews and multi-option questions from the agent render as clickable buttons instead of silently waiting
• Streaming caret — blinking ▍ at the end of the assistant's reply while it's being streamed
• Empty-state example prompts — three clickable starter prompts on first launch
• Inline franklin config commands — typing franklin config list / set / get / unset in the chat is handled locally without round-tripping to the LLM
• Default model is now blockrun/auto (was google/gemini-2.5-flash) — matches the CLI default so the Smart Router picks a model per request
• Various layout fixes (empty-state centering, model dropdown search clipping, settings panel positioning on narrow sidebars)
• Synced with Franklin core v3.8.35: prompt refinement for media, VideoGen async submit + polled settlement

0.3.0

• Added image & video generation — /image and /video slash commands prefill a natural-language prompt; the agent picks the right model, shows a cost preview with cheaper/premium alternatives, and only spends after you confirm
• Added in-chat confirmation prompts — when the agent needs a yes/no or a choice (e.g. media cost preview), a prompt card now appears inline with buttons instead of silently hanging
• Synced with Franklin core v3.8.31–v3.8.34: LLM-routed media model selection, model-choice-preserving status bar, exit/quit mid-turn, reliability pass (ESM, timeouts, retry budget)

0.2.1

• Fixed model switching bug — selected model was reverting to default (Gemini 2.5 Flash) after each turn, because baseModel wasn't being updated alongside model
• Fixed Trading Dashboard on Windows — PATH now split with platform delimiter, executables searched with .cmd/.exe suffixes, %APPDATA%\npm added to search paths

0.2.0

• Added chain switcher — toggle Base ↔ Solana payment chain from the toolbar badge
• Added prefetch status indicator — shows a live pulse "Fetching market data for BTC…" when the agent proactively pulls prices before responding
• Updated free model lineup: GLM-4.7 (new default), Qwen3 Coder 480B, Llama 4 Maverick, Qwen3 Next 80B Thinking (NEW)
• Synced with Franklin core v3.8.9–v3.8.30: unified turn analyzer, proactive prefetch, stocks via x402, panel Base ↔ Solana chain switcher

0.1.0

• Initial release
• Chat panel with model picker, wallet balance display, session history with search
• Doctor health panel, usage insights dashboard
• Trading dashboard icon (auto-launches franklin panel in the background)
• History dropdown, session auto-resume

Links

• Website
• GitHub
• Issues