<- all tokdocs

Graft (Now Trail-Code), OpenMontage, Codebase Memory MCP, Agency Agents

By , founder of Mavric and Apso

View on TikTok ->

A TikTok from Valeri Does AI (@valeridoesai) lists four GitHub repos that "went viral this week" and that "barely anyone's actually using": Graft, OpenMontage, Codebase Memory MCP and Agency Agents. All four are real and maintained. Graft has since been renamed Trail-Code, its own benchmark shows a 32% cost cut rather than the "roughly four times" claimed, and the four repos had between about 10,000 and 160,000 stars by October 2026.

What is Graft, and why is it now called Trail-Code?

Graft is a tool that gives coding agents a persistent map of your codebase so they spend fewer tokens searching. Its description reads "Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase." The video's transcript renders the name "Graphed," but the caption says Graft.

The repository is now trailhq/Trail-Code. The old trailhq/Graft path redirects there. The README says "Graft is now Trail-code. Graft lives on as the graft map, and every graft command still works." The rename landed in commits on 2026-09-25, after the video. The command changed from graft to trail, and the npm package from @nanonets/graft to @trailhq/trail-code, with the old names still working.

Searching GitHub for "Graft" now turns up h0rn3t/Graft, a separate 6-star Go project created 2026-09-21 that copies the same description. It is not the tool in the video.

Does Graft cut Claude Code costs by 4x?

No. The video says it "cuts [Claude] Code costs by roughly four times and runs about three times faster." The README's own benchmark, over 162 runs on two repos with 3 trials each, reports:

  • Tool calls: 4.2 to 2.3 (46% fewer)
  • Uncached input tokens: 8,070 to 4,650 (42% fewer)
  • Latency: 39.8 s to 15.8 s (60% less)

The README summarizes it as "32% lower cost, 42% fewer uncached input tokens, 60% less time." On PocketBase, "15 real tasks ran 21% cheaper and 14% faster." A 32% cost cut is about 1.5x. The only "4x" in the README is a diagram caption reading "up to 4x tokens saved."

The project's strongest result goes unmentioned. On SWE-bench Verified, graded with the official harness, Claude Sonnet 5 solved 33 of 50 issues with the graft map against 27 of 50 without it, using 23% fewer tokens.

Are people actually using these repos?

Yes. As of 2026-10-19 the GitHub API showed:

All four had been updated within four days. The npm downloads API shows @nanonets/graft at 7,980 downloads in the week the video posted, and 98,766 over the following 30-day window.

What is OpenMontage?

OpenMontage is an open-source system for producing video with AI coding agents. Its description now reads "World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files."

The video's card says "11 pipelines · 47 tools · 124 agent skills," which is out of date. The README said "11 production pipelines" only in early April 2026. By late August the repo held 13 pipeline definitions, 167 tool scripts and 157 skill files.

The README says it "Works with Claude Code, Cursor, Copilot, Windsurf, Codex, any AI coding assistant that can read files and run code." It needs FFmpeg installed and renders through Remotion or HyperFrames. It runs with no API keys using free stock footage, and paid keys add more tools. OpenMontage is AGPL-3.0, the only one of the four that is not MIT, which carries extra obligations if you build on it.

What is Codebase Memory MCP?

The video says it "loads your whole repo into a knowledge graph your agent can query instantly." That matches the README, which describes indexing "functions, classes, imports, HTTP routes, and their relationships across 162 languages," a full Linux kernel index in about 3 minutes, Cypher queries under 1 ms, and 17 MCP tools across 45 client surfaces, with no hosted service or API key. The maintainer is Martin Vogel.

The project's arXiv paper, dated 2026-03-28, reports a tradeoff the video skips. Across 31 repositories, it "achieves 83% answer quality versus 92% for a file-exploration agent, at ten times fewer tokens and 2.1 times fewer tool calls."

How many agents are in Agency Agents?

The video says "232 sub-agents, covering 16 domains." The repo, owned by Michael Sitarzewski, held 289 agent files across 19 top-level directories on 2026-08-26 and holds 298 today. The largest are engineering (65), specialized (59) and marketing (37).

The README has said "230+ Specialized Agents" since before the video, which is the likely source of 232. The per-domain counts shown on screen come from the separate Agency Agents desktop app and are lower than the repo's own directories. The README lists setup options for Claude Code and other tools including Copilot, Gemini CLI, OpenCode, Aider, Windsurf and Codex.

Bottom line

All four repos are real, active and widely used. Look for Graft under its new name, Trail-Code, and expect about a third off costs, not a quarter of the original bill. Codebase Memory MCP saves tokens at some cost to answer quality, and OpenMontage's AGPL license matters if you ship anything built on it.

Sources

Published August 31, 2026. Based on a TikTok by @valeridoesai. Full video notes (markdown)