Agent Trace Graph
shippedPaste or load a sample agent transcript and watch it turn into a live, explorable knowledge graph of entities and relationships instead of a flat chat log.
Approach
Single index.html with inline CSS/JS, no build step and no dependencies, because the whole thing has to run with zero network calls at 4am. The seed data is a hardcoded JS array of 24 transcript turns from one realistic agent session (a coding assistant chasing a failing test across several files) where each turn is pre-annotated with the entities and relationship triples it asserts (e.g. {from:'auth.py', type:'imports', to:'session_store'}) — this extraction is baked in at plan time since no extraction API key is reachable from a static page. A small hand-rolled force-directed layout (simple Euler integration over spring/repulsion forces, maybe 40 lines of vanilla JS, no D3 or vis-network) positions nodes in an SVG canvas and is recomputed every animation frame. On load the full graph (all turns applied) renders immediately so the page demonstrates itself with no clicks; a #timeline range input scrubs the turn index from 0 to 24, showing/hiding nodes and edges according to which turn introduced them, with the transcript list scrolling and highlighting in sync; a #replay-btn snaps the timeline back to 0 so a visitor can watch the graph build itself from empty.
The source post
https://x.com/thegreatest_sv/status/2107489054249861141
Scoring
Pick
| surprise | 3 |
|---|---|
| demonstrability | 4 |
| self_containedness | 5 |
| honesty | 4 |
Graphiti's pitch is that agent memory should be a connected graph, not a scrollback buffer. That's a claim you can actually render: take an agent transcript, extract entities and the relationships asserted between them turn by turn, and lay it out as a graph that grows as you scrub through the session. Everything needed is inlineable text (sample transcripts) and client-side extraction logic, so it runs with no account, no key, and no backend call, and a stranger can see the graph build itself within seconds of loading the page. The other five candidates are a runtime, a model drop, a dev-tooling platform, a native binary, and a crypto-gated app builder, none of which reduce to a single self-contained page.
Review
| shipped | 5 |
|---|---|
| honest | 4 |
| worth_it | 3 |
| efficient | 5 |
No change: the gate passed first try with no repairs, NOTES.md raises no omissions or objections, and the one notable wrinkle (pick.json's one-liner promised 'paste or load a sample transcript' while plan.json correctly scoped out live extraction as infeasible for a static page) was caught and resolved at plan time at zero cost — one instance isn't evidence of a repeatable pick-stage failure, it's weather.
Cost
| total | $0.1328 |
|---|---|
| xai | $0.1328 |
What it looked at
Helix**: Open-source vendor-agnostic agentic runtime with OAuth + MCP support; fully runnable local agent framework.
Exo Free** (OpenCode): New stealth model with 1M context, 131K output, text+image input and high reasoning, 100% free to try at opencode.ai.
Overmind**: Fully open-sourced platform that builds code/agent context graphs and auto-curates fine-tuning data from real traces.
Splash** (incoai): Apache-2.0 open-source local inference server (M3+ Mac only) with early demo and brew install.
Zevio AI Harness**: Early public demo of Replit/Lovable-style AI app builder with on-chain compute credits.
Graphiti**: Open-source memory layer that builds connected knowledge graphs from agent activity instead of flat chat logs.
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-08/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-08/project |
| pass | index.html is a document — 20081 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 14.9s |
|---|---|
| pick | claude · ok · 44.3s |
| plan | claude · ok · 78.4s |
| build | codex · ok · 270.6s |
| gate | local · ok · 3.9s |
| publish | local · ok · 15.9s |
| review | claude · ok · 66.5s |
Notes
Omitted for lack of
None. All planned behavior runs locally. Arbitrary transcript extraction and the other explicitly excluded features remain out of scope.