ANTON — operator-led AI harness for M&A and investment professionals

See it in action

The single-operator cockpit over a Markdown vault — the same deal flows through every lane, with the sensitivity guard and audit trail wired through for institutional compliance.^*

^* Shown with synthetic demo content

The cockpit

ANTON agent cockpit with the workflow drawer closed — chat-first — **Chat-first — workflow drawer closed**The default is a clean composer over the sessions for the project: live markets, your week, the project rail, and real-time token spend. Interrogate deal docs or search Vault content in plain language. Quick start new project or general sessions.

ANTON agent cockpit with the workflow drawer open — the full skill palette — **…one keystroke opens the workflow drawer**The full dealmaker's palette — target screening, valuation, management meetings, and transaction materials (teasers, CIMs, IC memos) — each a governed skill you fire with a click or a hotkey.

Per-deal project rail and vault-scoped chat — **A workspace per deal**Each deal gets a dedicated project overview — stage, target, sponsor, key dates — and a chat that runs recall strictly over that specific transaction's vault folder.

Valuation, three ways

Governed in-bridge LBO intake agent — **Governed intake, not a black box**The LBO agent reads the CIM and transcribes only sourced assumptions — anything unsourced becomes a question for the operator to sign off on.

Paste-ready orchestration prompt for an attended CLI run — **Or drive it from the CLI**A paste-ready, fully-specified prompt for an attended Claude Code run — same governed intake, same rigorous sourced-assumptions discipline.

Deterministic LBO engine intake form — **The engine does the maths**Operator approves inputs and assumptions; a deterministic Python engine drives your proprietary Excel model cell-by-cell. No LLM ever computes a number.

Confidentiality-aware routing & budgets

Per-skill and per-crew model routing, cloud-fallback ladder, MNPI attestations — **The guard, made legible**Per-skill and per-crew model assignment, the cloud-fallback ladder, and MNPI cloud attestations — confidentiality-aware routing you can actually see.

Per-model token budget caps and monthly usage — **Token budgets and caps**Set monthly caps globally, per project, or per model; usage tracks against the cap and warns before it bites — the hard cost block is separate.

Per-project token budget — a cap scoped to a single deal — **…or scoped to a single deal**Cap one project and track every model against it — Claude, ChatGPT, MiniMax, local Ollama — so a single engagement can't blow the month.

Governance, taxonomy & operator

Skill and routine sensitivity taxonomy — **Every skill carries a sensitivity**The taxonomy maps each skill and routine to a tier and an allowed lane — the rules the guard enforces, gathered in one place.

Tuned to one operatorExpertise sectors, watchlists, news coverage, and a profile that shapes what ANTON drafts — all editable without opening Obsidian.

Executive Differentiation Matrix

Issue	Generic AI Platforms
Valuation Engine	Black-Box Computations: Relies on non-deterministic LLMs to calculate figures, leading to frequent mathematical hallucinations and broken dynamic formulas.	Deterministic Execution: Zero numbers are computed by the LLM. Financial modeling tasks shell out to a native Python engine that drives versioned Excel templates (LBO, DCF, Comps) cell-by-cell.
Data Confidentiality	Opaque Cloud Risk: Relies on generic cloud privacy policies; high operational risk of leaking sensitive Virtual Data Room (VDR) contents or deal parameters.	Structurally Enforced Local Isolation: Governed by a strict sensitivity guard mapping files into four distinct tiers. `Confidential` and `MNPI` data tiers default to a local-only GPU, ensuring data never leaves the machine.
Execution Governance	Unchecked Background Swarms: Opaque agentic loops execute hundreds of hidden background calls, generating highly unpredictable, non-repeatable outputs.	Four Explicit Processing Lanes: The operator selects the exact lane (Chat, Single Skill, Composite, or Autonomous Crew) via specific verbs, dictating hard token ceilings and explicit determinism boundaries.
Budget Control	Unmonitored Runaway Costs: Opaque architectures lack granular tracking, risking surprise API credit card overruns during intensive deal sprints.	Granular Token Budgeting: Features strict monthly budget caps that can be configured globally, per individual model, or mapped directly to a specific deal project with proactive friction warnings.
Memory Governance	Silent Database Overwrites: Automatically appends unverified data or vector embeddings to a shared database, creating unvouched facts and corrupting deal history.	Principal-Gated Knowledge Base: Waterfall knowledge structure on a per project > per sector > expert essence, feeding from all available sources. System conclusions are delivered purely as isolated proposals. Fact updates to your permanent Markdown vault require explicit operator sign-off—never overwritten, never silent.
Institutional Audit Trail	Opaque Operations: Opaque systems offer zero underlying insight into why a model arrived at a particular conclusion, making validation impossible.	Comprehensive Telemetry: Every routine run and model call generates a structured, queryable trail through a sanitize-then-redact pipeline, logging provider, model, token cost, sensitivity tier, and execution lane.
Continuous Optimization	Static Prompt Architectures: Requires manual, complex prompt engineering or expensive code-level adjustments to align with changing workflow preferences.	Heuristic Feedback Loops: Actively monitors recurring follow-up query patterns to cluster behaviors and propose concrete, version-controlled modifications to its own underlying templates and operating rules.

Five layers that meet only across a boundary

Layers communicate over the filesystem, HTTP, or a subprocess — never by importing each other. That discipline keeps the math auditable, data contained, and the heavy engines swappable.

vaultmarkdown memory

Plain-Markdown canonical store — every note, decision, and source is a .md file with mandatory frontmatter. Git-managed; the file is the record.

routinesthe bridge

The FastAPI core — the only layer that touches everything. ~111 loopback endpoints, the scheduler, encrypted credentials, telemetry, and the governance core.

Governance core— sensitivity guard · cost gate · audit. Every lane passes through here.

enginevaluation

A Python wrapper that drives Excel templates through a cell-map registry. It computes and self-checks; the model only picks inputs and narrates.

dashboardcockpit

A React + TypeScript cockpit — chat shell, workflow tiles, an inbox of pending proposals, burn-rate and routing panels.

sidecarscomposite · crew

Two heavier orchestration runtimes kept out-of-process — reached over HTTP or a subprocess, never vendored in.

The guard is one before_llm_call hook registered at startup. A single skill, a step inside a composite, and every role inside a crew all flow through it. No lane bypasses it.

One request, one lane — and you choose it

In the current LLM environment there is a tradeoff between intelligence (quality), privacy (LLM providers using data to train their models) and cost (API/subscription for cloud; very high costs for local hardware). Anton gives the operator the flexibility to use any provider or mix providers including local LLM to maximise budget.

Chat

type in the composer

Determinism
Latency: 2–10s
LLM calls: 1

Single skill

/recall · /comps · /lbo

Determinism
Latency: 3–30s
LLM calls: 0–1

Composite

/pitch · /teaser · /ic-memo

Determinism
Latency: minutes
LLM calls: declared

Autonomous crew

/triage · /explore · /debate

Determinism
Latency: minutes
LLM calls: 10s–100s

Determinism = will the same request give the same answer? More bars = more repeatable (a fixed engine call); fewer = emergent (agents reason freely, so output varies). The left lanes are cheap and exact; the right lanes are powerful but variable.

Confidentiality level decides where a model may run

Every note carries a tier. The guard maps tier to an allowed lane before dispatch, and defaults to the more restrictive lane when uncertain.

public

Listed-company financials, press releases, sector statistics.

any cloud lane

internal

Your internal thesis on public material; no confidential party names.

cloud frontier model

confidential

Project codenames, target & sponsor names, NDA contents, VDR docs.

local only

MNPI

Pre-announcement results, embargoed news, inside information.

never leaves

Single-operator, local-first, genuinely in use

A work in progress, shared for review. Honest about what's live and what isn't.

Live today

Three of the four lanes — chat, single skills, and 4 autonomous crews running.
~111 endpoints across 49 router files; 3,871 tests across 249 files.
Overnight automations — briefings, trackers, sector reads.
LBO end-to-end through the engine, captured back to the vault as a gated proposal.
Valuation comparables — Anton will understand target and identify comparable transactions and listed companies & provide strategic rationale for operator's approval.
Per-task routing, operator override windows, budget gating, per-provider ceilings.

In Progress

Buyer tracking & NDA — pending skill to track buyer engagement from email and batch personalised draft NDAs (Python engine, not LLM).
Composite lane — composite actions (pitch, teaser, IC paper).
DCF — blocked on per-template engine authoring.
Outlook & calendar ingestion — pending Microsoft Graph app registration.
Cross-machine deploy — local-first by design; no managed service.
HoT & SPA review — implement skill to draft Heads of Terms (based on learnings & expertise) and review SPA.

"He grimly does his work and sits motionless until it's time to work again, we can all take a page from his book" - Gilfoyle Bertram