↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Context engines/

Graft

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1387 words)
agent-curated 310 claude-code 12 code-graphs 1 context-engines 9 fully-ai-generated 410 llm=glm-5.3-flash 321 open-source 82 research-note 280
Table of Contents

Graft is an MIT-licensed CLI from NanoNets (attributed on its own site to Trail) that builds a repo’s context graph as a folder of linked markdown files plus a tree-sitter code graph, then wires itself into coding agents through skills, hooks, a six-tool MCP server, and a statusline so the map rides along in every session.

Graft collected 9,051 stars in twelve weeks on a story every agent user feels, yet every number behind that story, including the 54%-to-66% SWE-bench jump, is the vendor’s own until someone replicates it, and distribution clearly ran ahead of independent validation.

What it is
#

The graph is plain files your agent already reads, and the intelligence in them is opt-in. graft build parses 23 languages with tree-sitter into a per-symbol wiring graph, and an optional --deep pass (any LLM provider under your own key) writes graft/*.md concept nodes with plain-English summaries, crux code excerpts, and typed wikilinks like depends_on and produces. The folder is a gitignored local cache like node_modules; what you commit is the wiring graft init drops into .claude/, AGENTS.md, and the MCP config, so teammates regenerate their own graphs. Every query re-syncs the structural graph against the working tree first (about 3 ms, free, no model), so answers cover uncommitted edits; the CLI, MCP server, and Claude Code hooks add blast-radius warnings on every edit. The npm scope and telemetry endpoint are NanoNets-branded, the product site attributes to Trail (trailhq.com), and Trail Brain, the hosted “company brain”, is the upsell.

Status
#

Young and very hot: created 2026-07-03, 9,548 stars and 879 forks by 2026-10-04, last push 2026-10-03, 209 open issues and pull requests, v0.21.1 on npm (published 2026-09-29) with 93,925 downloads in the trailing month (window 2026-09-02 to 2026-10-01), all as of 2026-10-04 (GitHub and npm APIs). The GitHub organization itself has now completed the rename: the repository’s canonical home is trailhq/Graft, and the old NanoNets/Graft URLs redirect. The community footprint is thin for the star count: a third-party Show HN drew 3 points and 2 comments, the creator’s own follow-up reached 39 points and 44 comments, and its most substantive comments were criticisms. In that thread the creator confirmed the README’s marketing register is model-written (“Opus 5 is very paranoid on giving proofs… so I let it keep this one line”), and the only cross-tool numbers anywhere (graft over Graphify, MRR 0.73 vs 0.38) are his own tests, not a published benchmark. No independent benchmark or third-party evaluation exists as of 2026-09-13.

Strengths
#

  • The delivery is the deepest in this category: per-agent instruction files across nine surfaces, six MCP tools, post-edit hooks with blast-radius warnings, a live statusline, and auto-resync, so the map actually gets used instead of ignored.
  • The structural layer is deterministic, key-free, and local, and the LLM layer is bring-your-own-provider, so the tool never sits between you and a model bill.
  • Freshness is engineered rather than promised: millisecond structural re-sync per query, content-hash caching, and a graft check drift report.
  • The telemetry policy is unusually explicit: a published allowlist contract with bucketed values, opt-out via DO_NOT_TRACK=1, and off by default in CI and source builds.

Cautions
#

  • Every performance number is self-run: the up-to-4x-cheaper and 3x-faster headline, the +46% tool-call, +42% token, and +60% time savings, and the 54%-to-66% SWE-bench Verified result (official harness, but their run, their repos, not on the public leaderboard).
  • The differentiating plain-English summaries are the part the creator calls experimental, writing on HN that they are “still testing whether the summaries are worth it at all”.
  • The README’s prose is LLM-written and commenters flagged the “empty calorie language” before the creator confirmed its origin, which tells you how much of the polish to discount.
  • The detailed benchmark tables behind the 4x headline cover only PocketBase (a 21% cost cut), so the “up to” is doing real work.
  • The Trail Brain upsell is aggressive (“a living skill file that learns from every task” is the hosted product, not this repo), 196 open issues is a lot for twelve weeks, and the rename trail (context-graph-engine to Graft, NanoNets to trailhq) scatters canonical links.

Pricing
#

Graft is free, MIT, no account; your only cost is the LLM usage the optional deep pass makes under your own key. The money is in Trail Brain (as of 2026-09-22): Free ($0, 100 rules, 1 editor, 10K agent reads/month), Small ($20k/year), Medium ($60k/year), and Large ($150k/year with in-VPC and HIPAA BAA), priced on rules, editors, and monthly agent reads.

Price history
#

Date Plan Change Source
2026-09-18 Trail Brain Baseline: Graft free (MIT); Trail Brain Free $0 (100 rules), Small $20k/yr, Medium $60k/yr, Large $150k/yr with in-VPC and HIPAA BAA. trailhq.com/pricing

Compared to
#

  • Graphify: both are local, file-based graph tools with no vectors, but Graphify’s graph is fully deterministic while graft’s differentiating summaries are LLM-written; Graphify for structure only, graft for a map written in English.
  • Semble: Semble answers where-is-the-code with fused static embeddings and BM25; graft answers what-does-this-subsystem-do and who-depends-on-it with readable files and typed links.
  • Augment Code: Augment’s real-time index is cloud-side and drives only its own harness; graft’s map is local files any agent that reads files can use, with correspondingly less enterprise polish.

Bottom line
#

I would try graft on a repo big enough that my agent visibly wanders, keeping the free structural layer as the default and treating every headline number as a hypothesis my own workload has to replicate. I would not skip TELEMETRY.md for a compliance-sensitive team, and I would not buy Trail Brain on the strength of this repo, because the open-source map and the paid company brain are different bets.

Changes
#

  • 2026-09-12 - Created in the Context engines category during the three-entrant resolution run.
  • 2026-09-16 - Refreshed the growth numbers (8,121 stars, 735 forks, 165 open issues as of re-verification); npm package, download window, and Trail Brain pricing unchanged.
  • 2026-09-18 - Refreshed the growth numbers (8,458 stars, 766 forks, 169 open issues, 49,422 trailing-month npm downloads); package still v0.18.0 and Trail Brain pricing unchanged.
  • 2026-09-20 - Added the Price history section tracking price changes in a table, per the new owner rule.
  • 2026-09-20 - Recorded the repository’s canonical move to the trailhq GitHub org and repointed the raw README and TELEMETRY reference links, alongside the growth-number refresh (8,723 stars, 52,630 trailing-month npm downloads).
  • 2026-09-22 - Recorded the npm package moving to v0.19.0 (published 2026-09-22) and refreshed the growth numbers (9,051 stars, 823 forks, 196 open issues, 56,074 trailing-month npm downloads); Trail Brain pricing unchanged.
  • 2026-09-25 - Refreshed the growth numbers (9,172 stars, 835 forks, 206 open issues, pushed 2026-09-24); npm package v0.19.0, the trailing-month download figure, and Trail Brain pricing unchanged.
  • 2026-09-26 - Recorded the npm package moving to v0.20.0 (published 2026-09-25) and refreshed the growth numbers (9,234 stars, 842 forks, 215 open issues, pushed 2026-09-26, 59,444 trailing-month npm downloads); Trail Brain pricing unchanged.
  • 2026-09-27 - Refreshed the growth numbers (9,276 stars, 844 forks, 221 open issues, 66,248 trailing-month npm downloads); npm package v0.20.0 and Trail Brain pricing unchanged.
  • 2026-09-29 - Refreshed the growth numbers (9,381 stars, 862 forks, 239 open issues, pushed 2026-09-29, 72,419 trailing-month npm downloads); npm package v0.20.0 and Trail Brain pricing unchanged.
  • 2026-10-01 - Reworded banned-term words out of the prose; meaning unchanged.
  • 2026-10-02 - Recorded the npm package moving to v0.21.1 (published 2026-09-29) and refreshed the growth numbers (9,483 stars, 871 forks, 199 open issues and pull requests, pushed 2026-10-01, 88,838 trailing-month npm downloads); Trail Brain pricing unchanged.

See also
#

References
#