Preston
getting started

Preston understands your codebase —
and remembers why.

Most AI tools write code. Preston is the layer underneath: a grounded, contestable understanding of what your code is, what worries it, and what you've decided — so your agents and your team work from the same memory a staff engineer carries in their head.

preston probe <your-repo>open the console →
the brain runs in our cloud · your code is never stored · no GitHub app
The shift it makes
Generation→Understanding

Cursor and Claude Code write the code. Preston is what they — and you — should be grounded in while doing it.

Stateless→Compounding

Review bots re-read each diff cold. Preston climbs your repo once and keeps an understanding that deepens — every decision you record sharpens the next answer.

Generic→Grounded

Not “best practices.” Your conventions, your prior decisions, your risks — every belief walkable to the code it rests on.

What you do with Preston

Five verbs

consult

“I'm about to change X.” Get a grounded posture — what's been decided, what it'd watch for, how confident it is — in one call. The thing your agent should ask before it acts.

review

It reviews PRs against its standing understanding and your conventions — and stays quiet when there's nothing for you to act on. Taste, not noise.

why

Every belief is walkable to its evidence. No black box; surprise is fine when the chain is walkable.

challenge · affirm

Disagree with it. It's a participant, not an oracle — your correction becomes part of the graph.

remember

Decisions persist with their justification. A decision without a justification is one the next person reverses.

For the people building agents

Point your agents at Preston.

Preston speaks MCP. Wire it into your coding agent and, mid-task, it asks one question and gets your codebase's real posture back — what's been decided, what conflicts, what to watch, how confident — every item walkable. No RAG to tune, no context to stuff. The grounding your agent was missing.

# wire Preston into your agent (Claude Code)
claude mcp add --transport http preston https://api.preston.bot/mcp \
  --header "Authorization: Bearer <your-key>"

# then just talk — your agent makes the call for you
consult({ intent: "extract the webhook handler", repo: "billing" })
→ { prior_decisions, conflicts_with, open_risks, confidence } // each walkable
Is Preston for you?
  • ✓You want your AI agents grounded in your codebase's real conventions and history, not generic advice.
  • ✓You want review that knows your code, calibrates its confidence, and doesn't spam.
  • ✓You want the “why” and the tribal knowledge to stop living only in senior heads.
  • ✕You want something to write the code — that's Cursor / Claude Code's job. Point Preston at them instead, as their grounding.
Get started

Five minutes, for real.

Beta access is by key. Paste the one you were given and every command below fills in with it.

  1. 1
    Install it
    curl -fsSL https://api.preston.bot/install.sh | PRESTON_KEY=<your-key> sh

    Installs the preston CLI (from us, not public npm) and saves your key.

  2. 2
    Point it at a repo
    preston connect ~/code/your-repo

    Read-only by default. No GitHub app, no setup.

  3. 3
    Let it climb
    preston probe your-repo

    Your source is read server-side (~15–20 min), then deleted. Keep it current later with preston catch-up.

  4. 4
    Wire it into your agent
    claude mcp add --transport http preston https://api.preston.bot/mcp --header "Authorization: Bearer <your-key>"

    So your agent uses everything Preston learned — grounded reviews, the risk map, the why. Pick your editor:

  5. 5
    Talk to it
    preston

    “What should I watch for if I touch the payment flow?” — answered from what it perceived, with how confident it is.

How it thinks
perceive→believe→consult · review→contest→remember↻

Understanding that compounds — every belief grounded in cause, calibrated by confidence, retractable by you. Your work is yours; the inferences Preston draws from it stay yours, and stay retractable.

What Preston is not

Not a code generator. Not a linter. Not a chatbot stapled to your repo. Not an oracle — it tells you when it's wary, and it's wrong sometimes; that's why every belief is walkable and contestable. And not a dashboard: it's a substrate — derivations, not summaries.