AI Agent Overview

The NeuralRepo AI agent is a conversational assistant embedded directly in the web app. It can search your ideas, create new ones, discover connections, brainstorm, summarize, and perform advanced analysis — all through natural language, with results rendered as visual artifacts.

Two Interfaces

The agent is available in two places:

Agent view

A full split-pane view with a chat panel on the left and an artifact canvas on the right. Open it from Agent in the sidebar. Best for deep exploration sessions where you want to see multiple artifacts side-by-side.

Mind Map Chat Panel

A floating, draggable chat panel that hovers over the mind map. Tool results animate directly onto the graph — nodes appear, edges draw, and clusters form in real time. Best for exploring your idea graph while conversing with the agent.

Model Registry

Five models are available. The three Anthropic models require a BYOK key; the two Workers AI models need no key at all.

Key	Label	Provider	BYOK	Context	Tool steps	Notes
`gpt-oss-120b`	GPT-OSS 120B	Workers AI	No	128K	12	The default. Runs in code mode
`qwen-30b`	Qwen 30B	Workers AI	No	32K	6	Compact prompt; also the creative sub-agent
`sonnet`	Sonnet	Anthropic	Yes	200K	10	Balanced reasoning
`haiku`	Haiku	Anthropic	Yes	200K	8	Faster, lower cost
`opus`	Opus	Anthropic	Yes	200K	10	Most capable

Every model is capped at 8,192 output tokens except GPT-OSS 120B, which gets 16,384. “Tool steps” is the hard stop on how many tool calls the model may chain in one turn — a request that needs more than its budget stops mid-way rather than looping.

Selecting an Anthropic model without a key configured returns a message telling you to pick a Workers AI model or add a key; it does not silently fall back. Your last-used model is remembered in localStorage.

All Anthropic requests are routed through Cloudflare AI Gateway — providing logging, caching, and rate limiting without exposing your key to additional parties.

Code Mode

GPT-OSS 120B runs in code mode by default: instead of calling tools one at a time, it receives a single codemode tool and writes JavaScript that calls the tools as functions, executed in a sandboxed Worker. Multi-step work — search, filter, then link — becomes one round trip.

You can force code mode on or off for any model with the code_mode user setting:

curl -X PATCH https://neuralrepo.com/api/v1/user/me \
  -H "X-API-Key: nrp_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"settings_json": "{\"code_mode\": true}"}'

The tools available are identical either way — code mode only changes how the model reaches them.

How It Works

The agent is a Cloudflare Durable Object (NeuralRepoAgent) built on the Agents SDK AIChatAgent class. It persists conversation history per user across sessions.

Messages stream back over the WebSocket connection. The frontend uses the Agents SDK client to manage the connection and reconstruct streaming state.

Thinking Blocks

Anthropic models run with extended thinking enabled and a 10,000-token thinking budget — internal reasoning the model performs before responding. In the chat UI, a collapsible “Thought process” block appears above the response when thinking is active. While the model is still generating, an animated neural icon shows it is working. The Workers AI models do not produce thinking blocks.

System Prompt

Each conversation begins with a system prompt that includes:

The user’s total idea count (active ideas only — archived ones are excluded)
5 most recent ideas, with number, title, and status
Up to 20 tags with their idea counts, and a note of how many more exist

Qwen 30B gets a compact version of that prompt — 3 recent ideas, 10 tags, and terser rules — because of its smaller context window.

In Graph Mode (mind map chat panel), the system prompt also includes a note that the graph is the primary output, encouraging the agent to use tools aggressively and keep prose brief.

Conversation History

The Durable Object keeps your conversation across sessions. For models with a context window below 200K, history is trimmed from the oldest message forward until it fits the model’s budget, so a long-running conversation quietly loses its beginning on Qwen 30B and GPT-OSS 120B while the Anthropic models keep everything.

Graph Mode

When using the chat panel embedded in the mind map, tool results animate directly onto the graph:

New ideas — nodes appear with a pulse animation
New connections — edges draw with an animation along the path
Search results — matching nodes highlight
Semantic layout — nodes re-cluster with a physics re-heat

Every graph mutation is added to an undo stack. The undo button in the chat panel header reverses the last agent action (deleting created relations, removing created ideas, etc.).

Working Memory Bar

At the top of the Agent view’s chat, a working memory bar shows a summary of what the agent has loaded in the current session:

Number of unique ideas referenced in artifacts
Number of connections made
Last search query

This gives you a quick sense of the agent’s context without scrolling through the conversation.

Prompt Palette

Both interfaces show a prompt palette — a set of contextual quick-action chips that update based on what you have selected or what the agent last did:

Default chips — “What am I working on?”, “Find missing connections”, “Weekly insights”, “Surprise me”
Single selection — Expand, Find related, Brainstorm, Suggest tags, Summarize
Multi-selection — Connect these, Compare, Summarize cluster, Visualize
Post-action — context-specific follow-ups after brainstorm, search, scoring, etc.

Click any chip to send that prompt immediately.

Selection Awareness

When you have nodes selected in the mind map, the agent automatically receives a list of selected idea IDs and titles as additional context. This allows prompts like “compare these” or “link these ideas” to work without you naming the ideas explicitly.