Brainer

Your docs and your landing page are nodes in your code graph.

Change a function → know which paragraph of marketing copy just became a lie. Start a task → get the files that matter, without the search loop.

no API key · no network · no telemetry · 24MB · works offline

---

Two problems, one graph

1. Prose rots silently

Every code-intelligence tool can tell you what calls what. None can tell you that yoursite.com/product/ still advertises the feature you renamed last Tuesday.

$ brainer surfaces getSession

  getSession()  src/lib/auth/helpers.ts

  3 surfaces describe this:
    docs/architecture.md › Server vs client         98%  declared
    docs/auth.md › Authentication › File Layout     98%  declared
    docs/security.md › Cookie hardening             90%  semantic

2. Agents search blind, and you pay for it

An agent starting a task greps, globs, reads, discards, greps again. Every round costs tokens and latency, and it still misses things.

$ brainer context "add rate limiting to the login flow"

  src/lib/auth/server.ts        91%  semantic + imported by login
  docs/auth.md › Auth Flow      88%  prose match
  src/app/auth-form.tsx         84%  imports auth/server
  src/lib/auth/actions.ts       81%  semantic
  ...
  Nothing matched: "throttle" — search for this directly.

Measured against a keyword-grep baseline on four real tasks:

	grep	Brainer
recall of the files that mattered	48%	94%
tokens spent	21.1k	12.8k

Twice the recall, ~60% of the cost, one call instead of a loop.

---

Quick start

npm install -g @stack256org/brainer
brainer install       # detect your agent, wire everything up
brainer init          # build the index

Installed from GitHub, not npm — the name brainer on npm belongs to an unrelated package. The @stack256org scope is registered; @stack256org/brainer publishes there when it is ready.

After that it is on by default — nobody has to remember Brainer exists or ask for it each session.

brainer install --global   # wire up every session, not just this project
brainer serve              # the whole graph in a browser

init downloads a 24MB model once. After that Brainer never touches the network.

On by default, including cold sessions

Host	How it becomes default
Claude Code	SessionStart + UserPromptSubmit + PostToolUse hooks
Codex, Copilot CLI, Gemini CLI	skill in ~/.agents/skills/ + AGENTS.md
Cursor, Windsurf, Continue, Aider	AGENTS.md / .cursorrules + MCP instructions
Any other MCP client	MCP instructions sent on connect

A cloud session starts with no index — a fresh clone, an ephemeral container. SessionStart detects that, builds the index in the background, and says so. The query hooks stay quiet until it is ready, so a cold session degrades to no Brainer, never to Brainer but wrong.

Always-on costs 36 tokens per turn, measured across a realistic mix of task openers, follow-ups and off-topic prompts. One grep round-trip in an agent loop is 800–2000.

---

Works with your agent

Built in tiers, so it degrades instead of breaking.

Tier	Surface	Works with	Automatic
1	CLI	anything, including CI	no
2	MCP	every MCP client	on request
3	Skills	Claude Code, Codex, Copilot CLI, Gemini CLI	model decides
4	Hooks	Claude Code	every time

The CLI is the substrate — MCP, skills and hooks are thin wrappers over it, so a host Brainer has never heard of still works at tier 1.

MCP config is identical everywhere; only the key differs (mcpServers, servers for VS Code, context_servers for Zed). brainer install writes the right one. Full matrix, plus an AGENTS.md snippet for hosts without skills: integrations/README.md.

Three hooks, three jobs

Event	When	Job	Tuned for
SessionStart	session opens	build/refresh index, warm the model	availability
UserPromptSubmit	before work	seed context with relevant files	recall
PostToolUse	after an edit	report prose describing the change	precision

The recall/precision split is deliberate. A missed file makes an agent do incomplete work confidently — worse than searching slowly. A false surface cries wolf, and a noisy hook gets muted within a week. Same graph, opposite tuning.

Both fail silent: no index, tool missing, timeout or stale data prints nothing and leaves your session untouched.

---

How it works

     PROSE                                          CODE
  ┌──────────────────────┐                 ┌──────────────────────┐
  │ landing page copy    │───── describes ─│ runMigrations()      │
  │ docs/architecture.md │═════ declared ══│ getSession()         │
  │ README › Themes      │───── semantic ──│ ThemeProvider        │
  └──────────────────────┘                 └──────────────────────┘
        ═══ corroborated          ─── guess, awaiting confirmation

Prose is a node type. A heading and its body, a JSX text run, a doc comment — indexed, linked to the symbols it describes, traversed like any other edge.

Three properties keep it honest:

Provenance is never blended. An edge from a resolved import and an edge guessed from similarity are different kinds of thing. Structural and semantic results come back in separate buckets; a guess never renders as a fact.

Confidence is calibrated, not raw. Embedding similarity is compressed — a cosine of 0.50 means the link is right ~98% of the time. Reporting the raw number would say "coin flip" about a near-certainty. Brainer stores the score and reports the measured probability.

Guesses become facts. brainer confirm writes a marker into the file, so a link you approved once is never guessed again. Precision climbs with use instead of being frozen at whatever the model gives you.

---

Commands

brainer context "<task>"      # files a task touches — recall-tuned seed set
brainer surfaces <symbol>     # what prose describes this?
brainer impact <file>         # blast radius, bucketed by provenance
brainer trace <from> <to>     # how does A reach B?
brainer review                # walk unconfirmed links, y/n/skip
brainer doctor                # stale index, orphaned links, hubs
brainer serve                 # the graph, in a browser

Every command takes --json.

---

Measured, not asserted

Verified against a 103-file TypeScript repo and a separate marketing site, with a hand-labelled golden set.

Link recall@30	0.885
Link precision@3	0.677 (hand-adjudicated)
Agent context recall	94% shipped (100% in the spike) vs grep's 48%
Import resolution	100% (162/162, incl. tsconfig aliases)
impact() query	0.05ms worst case
Incremental reindex	15× faster than full
Always-on overhead	36 tokens/turn

Two probes failed before they passed, and a cross-encoder reranker was cut for making precision worse (0.391 → 0.198). Full methodology, including what we got wrong: docs/VERIFICATION.md.

---

Honest limitations

• Renames break confirmed links. doctor reports them; automatic rename detection is deferred.
• 59% of real links have no corroborating identifier and stay labelled guesses until reviewed. The review flow is load-bearing, not a nicety.
• Some prose has no code referent. ~1/6 of a landing page describes the project as a whole. Returning nothing is correct, and Brainer returns nothing rather than inventing a link.
• Cross-file call resolution is heuristic without a type checker, and labelled as such.
• The context seed set is not exhaustive. It is a starting point.

---

Language support

TypeScript, TSX, JavaScript, JSX. Markdown, MDX and HTML for prose. Grammars sit behind a registry, so more languages are a contribution, not a fork.

Brainer deliberately does not chase a 60-language count. Breadth is the competition's differentiator. Prose is ours.

---

Releasing

Publishing runs on npm trusted publishing (OIDC), so there is no NPM_TOKEN secret to leak or rotate. GitHub Actions requests a short-lived identity token, npm checks it against the trusted-publisher rule for this repository, and publishes if it matches.

npm version patch          # or minor / major
git push

That is the whole release. release.yml sees a version with no matching tag on main, runs the tests, tags it, cuts a GitHub release, and publishes.

Release and publish are one workflow, publish.yml, because npm trusted publishing validates the entry-point workflow filename. Splitting them cannot work: a release cut with GITHUB_TOKEN raises no event for another workflow to catch, and a reusable workflow presents the caller's filename in the OIDC claim. One file, named what npm trusts, is the only arrangement where a merge publishes by itself.

A merge whose version is unchanged releases nothing, on purpose — a typo fix should not burn a version number, and npm versions cannot meaningfully be unpublished.

First publish is manual. npm cannot create a new package over OIDC: a trusted-publisher rule can only attach to a package that already exists. So the first version goes up by hand, and everything after it is automatic.

With a passkey (WebAuthn) as your second factor there is no six-digit code to pass to --otp. Use a current npm, which falls back to a browser flow:

npx npm@latest publish --access public

It prints a npmjs.com/auth/cli/… URL, waits while you authenticate with the passkey, and then completes. This needs a real terminal — npm redacts that URL when stdout is not a TTY, so it cannot be driven from a script. npm 10 does not offer the fallback at all and only accepts --otp.

Then, on npmjs.com: the package → Settings → Trusted Publisher → repository stack256org/brainer, workflow publish.yml. After that, releases publish themselves and no token is ever stored.

workflow_dispatch runs the same job with --dry-run by default, so the whole path can be exercised without shipping anything.

Status

v0.1.0 — working. Indexes, links, queries, serves and installs. Verified against a real 127-file corpus: 281 symbols, 209 prose blocks, 2684 edges in 7.7s.

The confirmation ratchet is live: review, confirm and reject are built and tested, so a guess you approve is written into the file as a marker and read back as ground truth — precision now climbs with use rather than sitting wherever the model left it. trace and doctor are built too.

Still outstanding: true per-file incremental reindex. reindex rebuilds.

Usage · Design · Verification · Integrations

Part of Stack256 — 256 open-source applications you can host yourself.