---
title: "Graft (Now Trail-Code), OpenMontage, Codebase Memory MCP, Agency Agents"
date: "2026-08-31"
video_id: "7680294929351937302"
tiktok_url: "https://www.tiktok.com/@valeridoesai/video/7680294929351937302"
creator: "@valeridoesai"
category: "AI"
tags: ["trail-code", "graft", "openmontage", "codebase-memory-mcp", "agency-agents"]
excerpt: "Four AI coding repos checked: Graft is now Trail-Code with 32% lower cost (not 4x), and the repos have thousands of stars, not \"barely anyone\" using them."
---
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](https://github.com/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`](https://registry.npmjs.org/@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:

- [msitarzewski/agency-agents](https://github.com/msitarzewski/agency-agents): 161,743 stars, 26,173 forks, MIT
- [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage): 69,917 stars, 9,057 forks, AGPL-3.0
- [DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp): 46,948 stars, 3,884 forks, MIT
- [trailhq/Trail-Code](https://github.com/trailhq/Trail-Code): 10,455 stars, 918 forks, MIT

All four had been updated within four days. The [npm downloads API](https://api.npmjs.org/downloads/point/2026-08-25:2026-08-31/@nanonets/graft) 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](https://github.com/calesthio/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](https://github.com/DeusData/codebase-memory-mcp), 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](https://github.com/DeusData).

The project's [arXiv paper](https://arxiv.org/abs/2603.27277), 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](https://github.com/msitarzewski), 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

- [trailhq/Trail-Code on GitHub](https://github.com/trailhq/Trail-Code) - rename, license and benchmark results
- [calesthio/OpenMontage on GitHub](https://github.com/calesthio/OpenMontage) - license, description and requirements
- [DeusData/codebase-memory-mcp on GitHub](https://github.com/DeusData/codebase-memory-mcp) - features and performance figures
- [msitarzewski/agency-agents on GitHub](https://github.com/msitarzewski/agency-agents) - agent files and README
- [Michael Sitarzewski on GitHub](https://github.com/msitarzewski) - agency-agents owner
- [DeusData on GitHub](https://github.com/DeusData) - codebase-memory-mcp owner
- [arXiv 2603.27277](https://arxiv.org/abs/2603.27277) - Codebase-Memory quality and token results
- [@nanonets/graft on npm](https://registry.npmjs.org/@nanonets/graft) - package history
- [npm downloads for @nanonets/graft](https://api.npmjs.org/downloads/point/2026-08-25:2026-08-31/@nanonets/graft) - weekly downloads


## Video description and verification notes

**A 37.589 s vertical clip, 576x1024, delivered as an MP4 (`mov,mp4,m4a` container, `major_brand=isom`, `encoder=Lavf58.76.100`, TikTok format id `bytevc1_540p_399293-1`) carrying an H.265/HEVC Main profile level 93 video stream at 23.976 fps (`r_frame_rate=2997/125`, 901 frames, 362,013 bps) and an AAC HE-AACv2 stereo 44.1 kHz audio stream at 32,186 bps, 399,304 bps overall, 1,876,180 bytes on disk; posted at epoch `1788208017`, which is 2026-08-31T20:26:57 UTC, by handle `valeridoesai` (uploader id 7633381658390414339, nickname "Valeri Does AI"), audio credited as "original sound - Valeri Does AI", with a 748-character description; at capture time (metadata.json mtime 2026-10-19 06:24:21 local) the clip showed 66,700 views, 3,420 likes, 281 comments, 1,002 reposts and 4,631 saves. I read all 19 extracted frames (frame_001 through frame_019, 2 s apart) and the full transcript of 6 segments and 103 words. The frames run like this: frame_001 opens on a stacked montage of four product cards behind a white GitHub logo wordmark, caption "Four GitHub repos"; frame_002 shows the OpenMontage card reading "The first open-source, agentic video production system" with a blurred GitHub repository page at the bottom edge legible as "msitarzewski / agency-agents", caption "anyone's actually using"; frame_003 fills the top half with a Graft card whose subtitle reads verbatim "Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase." above a row of shields.io badges, caption "Number one Graft"; frames 004 through 006 hold a two-lane comparison diagram labelled "WITHOUT GRAFT" (Task, Agent ZERO MEMORY, Blind Search GREP + GLOB, Dead Ends WRONG FILES, Guess Wiring RE-DERIVE, Task Shipped BUDGET SPENT, Context Lost DISCARDED, "REPEATS EVERY TASK · EVERY TEAMMATE'S AGENT") and "WITH GRAFT" (Repo, Graft Init, Node Graph, In Git, Pre-Session Hook USE GRAFT FIRST, Task, Agent, Precise Context NO GUESSING, Task Shipped TOKENS SAVED), under the captions "Code costs by" and "three times faster"; frames 007 through 009 pair an OpenMontage card reading "11 pipelines · 47 tools · 124 agent skills" with stock renders of a VR headset and an exploding laptop, captions "which literally makes", "shoot cut and", "export video like"; frames 010 and 011 show a "Codebase Memory MCP" card and a rotating force-directed graph over a terminal header reading "GRAPH: Users-martinvogel-project_di..." and "1,126 edges"; frame_012 is a genuine screenshot of that project's GitHub README showing badges for "Hybrid LSP 10 languages", "agent surfaces 43", "pure C no language runtime", "openssf scorecard 7.8", "SLSA level 3", "arXiv 2603.27277" and an "MIT license" tab, plus body text "Full-indexes an average repository in milliseconds, the Linux kernel (28M LOC, 75K files) in 3 minutes"; frame_014 shows an "Agency Agents" card with the only fully legible repository path in the video, "msitarzewski / agency-agents  Public"; frame_015 shows that repo's README sections "Option 2: Use with Claude Code" and "Option 4: Use with Ot[her tools]" listing Copilot, Antigravity, Gemini CLI, OpenCode, Aider, Windsurf, Kimi Code and Codex; frames 016 and 017 show a desktop app grid of domain cards (Engineering 23 agents, Design 8, Paid Media 7, Sales 8, Marketing 27, Project Management 6, Product 5, Testing 8, Support 6, Specialized 17) under captions "one repo 232" and "16 domains"; frames 013, 018 and 019 are the presenter alone against black or an office wall, closing on "DM me 'AGENT'" in yellow. No terminal session, install command, benchmark run or rendered output appears in any frame I read, and no repository URL for Graft, OpenMontage or Codebase Memory MCP is ever shown. The spoken argument is a four-item list claiming these repos went viral this week while "barely anyone's actually using them", with per-item performance claims, and an instruction to DM the word "AGENT" for links. The transcript's last segment ends at 37.46 s against an audio stream of 37.589 s and an `audio.mp3` of 37.642 s, so nothing in it is timestamped past the end of the audio and none of it is Whisper hallucination.**

## Whisper mangled the name; the burned-in caption and the description both say Graft

At 4.48 s the transcript reads "Number one, Graphed, which cuts code code costs by roughly four times". Two words there are wrong. The burned-in caption in frame_003 reads "Number one Graft", the card on screen is titled "Graft", and the video description names it "Graft". The doubled "code code" is Whisper losing "Claude" in front of "Code"; the description confirms the claim is about Claude Code. I use the spellings from the caption and description throughout, not the transcript.

## "Graft" is now Trail-code, and the name the video gives viewers points at a different project

The repository behind the frame_003 card is `trailhq/Graft`, and that path no longer resolves to itself. A GitHub API request for `repos/trailhq/Graft` returns HTTP 301 redirecting to `https://api.github.com/repositories/1288325627`, which is `trailhq/Trail-Code`. The repository was created 2026-07-03T13:44:21Z, carries an MIT license, is not archived, and was last pushed 2026-10-19T16:29:21Z. Its description is the exact string on the card: "Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase." Its homepage field is still `https://trailhq.com/graft`.

The rename happened after the video posted. The README now opens with a note reading "Graft is now Trail-code. Graft lives on as the graft map, and every `graft` command still works", and a migration table that maps "Name: graft to Trail-code", "Command: `graft` to `trail` (`graft` still works)" and "Package: `@nanonets/graft` to `@trailhq/trail-code` (the old one still works)". The commit log puts the rebrand in late September 2026: `feat(cli): rename graft brain to graft trail, keeping the old spelling working (#468)` landed 2026-09-25T12:23:05Z and `docs(readme): drop "Brain" from Trail Brain references` landed 2026-09-25T14:33:50Z. The `trailhq` GitHub organization itself was only created 2026-08-24T21:05:28Z, a week before the video, and it holds exactly one public repository.

This matters because the video gives viewers a product name and nothing else. A GitHub search for "Graft" today surfaces `h0rn3t/Graft`, owned by Eugene Shershen, which carries the identical description string, has 6 stars, was created 2026-09-21T16:31:12Z and is a separate Go implementation whose README is written in Ukrainian. It is not the project in the video and it did not exist when the video was recorded. The nearest other candidates are `AEndrix03/Graft` (37 stars) and `getmodern-ai/graft` (6 stars), neither of which is the one shown. Of the four repos, only `msitarzewski / agency-agents` is ever readable on screen, in frame_014 at roughly 26 s.

## The 4x cost cut and 3x speed claims do not match Graft's own published benchmark

At 4.48 s the video says Graft "cuts [Claude] Code costs by roughly four times and runs about three times faster with any agent", and the caption "three times faster" is burned into frame_005. The repository's own README reports different numbers in its headline results table, measured over "162 runs, two repos (this one and a real Node/Express auth service), 3 trials each":

| Metric | Cold agent | With the graft map |
|---|---|---|
| Tool calls | 4.2 | 2.3 (46% fewer) |
| Uncached input tokens | 8,070 | 4,650 (42% fewer) |
| Latency | 39.8 s | 15.8 s (60% less) |

The README's prose summary is "32% lower cost, 42% fewer uncached input tokens, 60% less time", and a separate line reports "On PocketBase, 15 real tasks ran 21% cheaper and 14% faster". A 32% cost reduction is about 1.5x, not four times. A 60% latency reduction is 2.5x faster, and the PocketBase figure is 14%. The only place "4x" appears anywhere in the README is the alt text of the how-it-works diagram, which ends "One graph, every agent, up to 4x tokens saved". That is an upper bound in an image caption, and the video converts it into a flat cost multiple.

The strongest result the project publishes is not the one the video uses. The README reports that on SWE-bench Verified, graded by the official `swebench` 4.1.0 harness on official images, Claude Sonnet 5 solved 33 of 50 issues with the graft map against 27 of 50 without it, using 23% fewer tokens. That is a correctness gain on an independent benchmark, and the video does not mention it.

## "Barely anyone's actually using them yet" is wrong by every metric I can measure

The opening line at 0.00 s is "Four GitHub repos went viral this week, and barely anyone's actually using them yet." As of 2026-10-19T23:40:03Z, the GitHub API returns:

| Repository | Stars | Forks | Open issues | License | Last push |
|---|---|---|---|---|---|
| `msitarzewski/agency-agents` | 161,743 | 26,173 | 166 | MIT | 2026-10-19T16:48:24Z |
| `calesthio/OpenMontage` | 69,917 | 9,057 | 409 | AGPL-3.0 | 2026-10-16T15:07:33Z |
| `DeusData/codebase-memory-mcp` | 46,948 | 3,884 | 648 | MIT | 2026-10-19T23:12:17Z |
| `trailhq/Trail-Code` | 10,455 | 918 | 268 | MIT | 2026-10-19T16:29:21Z |

None is archived or disabled. All four were pushed within four days of this writing, and two within the same day. More than 26,000 people have forked agency-agents, which is a copy action, not a bookmark.

The npm registry gives a direct install count for Graft. The package `@nanonets/graft` was first published 2026-07-15T12:03:03Z, has shipped 37 versions through 0.23.0 on 2026-10-15, and recorded 7,980 downloads in the week of 2026-08-25 to 2026-08-31, the week the video posted. Over the last 30 days it recorded 98,766 downloads. The successor package `@trailhq/trail-code` recorded 1,858 in the same window, which is consistent with the README's note that the old package still works and most users have not migrated. Roughly 8,000 installs in the week the presenter said barely anyone was using it.

## The OpenMontage card on screen describes the repo as it stood in April, five months before the video

Frames 007 through 009 show a card reading "OpenMontage / The first open-source, agentic video production system. / 11 pipelines · 47 tools · 124 agent skills". The repository is `calesthio/OpenMontage`, created 2026-03-29T15:23:22Z by a user whose profile name is "Calesthio" and whose bio reads "Creator of OpenMontage and Crucix". Its current description reads "World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files."

I walked the README history. The phrase "11 production pipelines" last appears in commits dated 2026-04-03 through 2026-04-08; from 2026-04-10 onward the README says "12 production pipelines", and from 2026-08-13 onward it says "10+ production pipelines". At the commit closest to the video's post date (`a3768a0`, 2026-08-26T18:29:59Z) the tree contained 13 files under `pipeline_defs/`, 167 Python files under `tools/` and 157 markdown files under `skills/`, and the README already advertised "700+ agent skill and production-knowledge files". The card's "47 tools · 124 agent skills" matches no state of the repository I found. The video presents a stale graphic as current.

The capability claim at 11.16 s, that OpenMontage "literally makes any coding agent shoot, cut, and export video like a studio", is closer to right than the numbers, with conditions the video omits. The README states it "Works with Claude Code, Cursor, Copilot, Windsurf, Codex, any AI coding assistant that can read files and run code", requires FFmpeg installed locally, and renders through Remotion or HyperFrames with FFmpeg for encoding. It runs with zero API keys using free stock footage and open archives, and paid keys unlock more tools. One detail worth flagging for anyone who clones it: OpenMontage is AGPL-3.0, the only non-MIT project of the four, which carries obligations the other three do not.

## Codebase Memory MCP does what the video says, and the paper behind it reports a quality cost the video skips

At 17.70 s the video says it "loads your whole repo into a knowledge graph your agent can query instantly". The repository is `DeusData/codebase-memory-mcp`, created 2026-02-24T22:01:00Z, MIT licensed, written in C, owned by a GitHub account whose profile name is "Martin Vogel". That name matches the terminal path visible in frame_011, "GRAPH: Users-martinvogel-project_di...", which means the demo footage was recorded on the maintainer's own machine. The README supports the capability claim: it indexes "functions, classes, imports, HTTP routes, and their relationships across 162 languages", reports "Linux kernel full index, 3 min, 28M LOC, 75K files to 4.81M nodes, 7.72M edges" and "Cypher query, <1ms", and exposes 17 MCP tools across 45 client surfaces with no hosted service or API key.

The badge values in frame_012 are lower than today's, which dates the screenshot rather than contradicting it. The frame shows "10 languages" on the Hybrid LSP badge, "agent surfaces 43", and body text citing 161 languages and 15 MCP tools. The live README says 162 languages, 17 MCP tools and 45 agent surfaces. The project grew between the recording and now.

The arXiv badge in frame_012 reads 2603.27277, and that preprint is real. It is "Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP" by Martin Vogel, Falk Meyer-Eschenbach, Severin Kohler, Elias Grünewald and Felix Balzer, dated 2026-03-28. Its abstract states the system parses 66 languages, and that across 31 real-world 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". The token saving is real and large. The answer quality is nine points lower than plain file-by-file exploration, a tradeoff the README states plainly and the video never mentions. The README separately reports "about 120x fewer tokens" from a five-query comparison, which it is careful to mark as a different measurement from the paper's 10x.

## "232 sub-agents, covering 16 domains" undercounts a repo that already held 289 agent files in 19 directories

At 24.82 s the video says agency-agents is "basically a full AI agency inside one repo 232 sub-agents, covering 16 domains", with "one repo 232" and "16 domains" burned into frames 016 and 017. The repository is `msitarzewski/agency-agents`, created 2025-10-13T12:12:29Z, MIT licensed, owned by Michael Sitarzewski, a GitHub account open since 2012 with 7,322 followers and a bio reading "Builder with 30+ yrs experience. Startup founder, Techstars alum, lifelong tinkerer."

I counted the tree at the commit closest to the post date (`3c95888`, 2026-08-26T14:57:10Z) and at HEAD. Excluding READMEs, examples, scripts, integrations and workflow files, the repo held 289 agent markdown files across 19 top-level domain directories on 2026-08-26, and holds 298 across the same 19 directories today. The directories are engineering (65 today), specialized (59), marketing (37), game-development (21), strategy (16), gis (13), security (12), design (10), sales (9), testing (9), paid-media (7), project-management (7), academic (6), product (6), spatial-computing (6), support (6), finance (5), healthcare (3) and research (1).

The per-domain counts the video shows on screen are lower than the directory counts for the same categories on the same date. Frame_016 shows Engineering at 23 agents against 59 markdown files in `engineering/`, and Marketing at 27 against 36 in `marketing/`. Several smaller categories do line up, including Paid Media at 7 and Support at 6. Those cards come from the separate Agency Agents desktop app, which the README links at `agencyagents.app`, and the roster it displayed was behind the repository. The repo's own README has said "230+ Specialized Agents" since before the video and still says it today, which is the likeliest origin of the figure "232". The "16 domains" count matches neither the 19 directories nor the 8 divisions the README names in its team-workflow section.

## Key Takeaways

- **Correction:** At 4.48 s the video calls repo one "Graft". GitHub now 301-redirects `trailhq/Graft` to `trailhq/Trail-Code`; the rename landed in commits dated 2026-09-25, after the video posted, and the README states "Graft is now Trail-code."
- **Correction:** At 4.48 s the video claims Graft "cuts [Claude] Code costs by roughly four times and runs about three times faster". The repository's README reports 32% lower cost, 42% fewer uncached input tokens and 60% less time over 162 runs, plus 21% cheaper and 14% faster on PocketBase. "4x" appears only as "up to 4x tokens saved" in a diagram alt text.
- **Correction:** At 0.00 s the video says "barely anyone's actually using them yet". As of 2026-10-19 the four repos hold 161,743, 69,917, 46,948 and 10,455 stars, and `@nanonets/graft` recorded 7,980 npm downloads in the week the video posted and 98,766 in the last 30 days.
- **Correction:** The OpenMontage card in frames 007 to 009 reads "11 pipelines · 47 tools · 124 agent skills". The README said "11 production pipelines" only in early April 2026; at the video's post date the repo had 13 pipeline definition files, 167 tool scripts and 157 skill files, and the README advertised "700+ agent skill and production-knowledge files".
- **Partial correction:** At 24.82 s the video says agency-agents holds "232 sub-agents, covering 16 domains". The repo held 289 agent markdown files across 19 top-level domain directories on 2026-08-26 and holds 298 across 19 today. Its README has long said "230+ Specialized Agents", which is the likely source of the figure.
- **Verified:** Codebase Memory MCP builds a persistent code knowledge graph queryable by coding agents, as claimed at 17.70 s. The README documents 162 languages, 17 MCP tools, 45 client surfaces, Linux kernel indexing in about 3 minutes and Cypher queries under 1 ms.
- **Clarification:** The arXiv preprint cited on that project's badge, 2603.27277, reports 83% answer quality for Codebase-Memory against 92% for a file-exploration agent, at ten times fewer tokens. The token saving comes with a nine-point quality cost that the video does not mention.
- **Clarification:** OpenMontage is AGPL-3.0. The other three are MIT. The video treats all four as equally free to pick up.
- **Clarification:** None of the four is archived, disabled or unmaintained. All four were pushed within four days of 2026-10-19. Open issue counts run high on the two largest by activity, 648 for codebase-memory-mcp and 409 for OpenMontage.
- **Context:** Only one repository path is legible anywhere in the video, "msitarzewski / agency-agents" in frame_014 at about 26 s. The other three are gated behind a pinned comment and a "DM me 'AGENT'" call to action at 34.32 s, neither of which is visible in the frames. A viewer searching GitHub for "Graft" today lands on `h0rn3t/Graft`, a 6-star Go project created 2026-09-21 that carries the identical description string but is not the project shown.
- **Context:** No terminal session, install command, benchmark run or rendered video output appears in any of the 19 frames. Every performance figure in the video is read off a card, not demonstrated.
- **Unverified:** The pinned comment's contents. I can read the video, its metadata and the repositories, but not TikTok's comment thread, so I cannot confirm which four links the creator actually distributes.

## Resources

- [trailhq/Trail-Code, GitHub](https://github.com/trailhq/Trail-Code) establishes the rename from Graft, the MIT license, the 2026-07-03 creation date, the 10,455-star count, and the benchmark table reporting 46% fewer tool calls, 42% fewer uncached input tokens, 60% less time and 33 of 50 SWE-bench Verified issues solved.
- [calesthio/OpenMontage, GitHub](https://github.com/calesthio/OpenMontage) establishes the AGPL-3.0 license, the 2026-03-29 creation date, the 69,917-star count, the "12 production pipelines, 100+ tools, 700+ agent skill" description, and the FFmpeg and coding-assistant requirements I cite.
- [DeusData/codebase-memory-mcp, GitHub](https://github.com/DeusData/codebase-memory-mcp) establishes the MIT license, the 2026-02-24 creation date, the 46,948-star count, and the 162 languages, 17 MCP tools, 45 client surfaces, 3-minute Linux kernel index and sub-millisecond Cypher query figures.
- [msitarzewski/agency-agents, GitHub](https://github.com/msitarzewski/agency-agents) establishes the MIT license, the 2025-10-13 creation date, the 161,743-star and 26,173-fork counts, the "230+ Specialized Agents" README claim, and the directory structure I counted at both 2026-08-26 and HEAD.
- [Michael Sitarzewski's GitHub profile](https://github.com/msitarzewski) establishes the spelling of the agency-agents owner's name and handle.
- [Martin Vogel's GitHub profile (DeusData)](https://github.com/DeusData) establishes the codebase-memory-mcp owner's name, which matches the `Users-martinvogel-project_di...` path visible in frame_011.
- [Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP, arXiv 2603.27277, 2026-03-28](https://arxiv.org/abs/2603.27277) establishes the author list, the 83% versus 92% answer-quality comparison, the ten-times token reduction and the 2.1-times tool-call reduction across 31 repositories.
- [@nanonets/graft on the npm registry](https://registry.npmjs.org/@nanonets/graft) establishes the 2026-07-15 first publish, the 37 released versions through 0.23.0, and the repository field pointing at trailhq/Trail-Code.
- [npm downloads API, @nanonets/graft](https://api.npmjs.org/downloads/point/2026-08-25:2026-08-31/@nanonets/graft) establishes the 7,980 downloads in the week the video posted and, for the last-month window, the 98,766 figure I cite.
