# AnyForge — The Control Plane for Agentic AI
# https://anyforge.ai
# Full reference for LLM crawlers and AI assistants

AnyForge is the control plane for agentic AI. It is one governance layer
consumed three ways: AnyForge Control (a governed LLM proxy for teams using
their own agents — Claude Code, Cursor, Cline, Continue, or raw SDK clients),
AnyForge Code Studio (a browser-based single-agent coding workspace with
governance built in), and AnyForge Crew (a role-based multi-agent engineering
platform with two mandatory human approval gates). Cost analytics, intent
routing, portable memory, and a hash-chained audit trail are shared across all
three surfaces.

## Company Overview

- Name: AnyForge
- Website: https://anyforge.ai
- Founded: 2025
- Founder: Edwin Poot (https://linkedin.com/in/edwinpoot)
- Category: AI Control Plane / Agentic Software Delivery
- Tagline: The Control Plane for Agentic AI
- Brand verbs: enforce, prove, predict

## Products

### AnyForge Control (bring your own agent)

- URL: https://anyforge.ai/control
- Category: Governed LLM proxy and AI control plane
- Install: `npx anyforge init`
- Pricing: Developer tier free, Team and Business tiers per seat, Enterprise custom
- BYOK (bring your own keys), no markups on model costs

#### Features
- SDK-compatible drop-in for OpenAI, Anthropic, Gemini, and OpenRouter
- Self-hosted Ollama support (local daemon, on-prem deployment, or Ollama Cloud) for air-gapped, on-prem, or data-residency-constrained customers
- Detects Claude Code, Cursor, Cline, Continue, and raw SDK clients automatically
- Per-call, per-agent, per-project cost and latency analytics with dimensional rollups
- Intent-based routing (rules, regex, or embedding-based) to the cheapest model that meets the quality bar
- Portable memory that survives provider switches
- Hash-chained, signed audit log of every prompt, response, and tool call
- Optional OPA policy enforcement (redaction, budget caps, PII blocking, allow-lists) activated per rule
- Configurable alerts at cost thresholds

#### Note
Control is ALSO native inside AnyForge Code Studio and AnyForge Crew. If you use
either of those products, you already have Control's feature set — nothing
separate to install. This page is for the "bring your own agent" path where
Control ships as a standalone proxy.

### AnyForge Code Studio (turnkey browser coding)

- URL: https://anyforge.ai/code-studio
- Try free: https://code.anyforge.ai
- Category: Browser-based single-agent AI coding workspace
- Pricing: Free tier (30 requests/day, no credit card required), Starter $20/mo, Pro $50/mo, Power $200/mo
- BYOK pricing model, no markups on API costs

#### Features
- Browser-based — no install required, works on any device
- Access to 200+ AI models via OpenRouter
- Native GitHub integration — commit and push directly from the browser
- Quality-first: writes failing tests before implementing fixes (TDD)
- Branch-locking: main branch protected by default
- Smart request classification: distinguishes directives from inquiries before acting
- Repo-grounded plans with surgical edits
- Built-in terminal for builds, tests, and scripts

#### Control native in Code Studio
- Every LLM call flows through AnyForge Control
- Per-call cost analytics, intent routing, portable memory, and hash-chained audit come built in
- Cross-surface cost visibility in the operator console (Code Studio + Crew + external-agent in one view)

#### Escalation to Crew
- Code Studio agent can propose a governed Crew via the `propose_crew` tool
- User accepts with one click, creates Initiative + Crew + WorkItem in an atomic transaction
- Code Studio stays live on the Crew conversation via the shared `crewChat` subcollection
- User approves or rejects Crew HIL gates from inside Code Studio (no context switch)

### AnyForge Crew (governed multi-agent delivery)

- URL: https://anyforge.ai (pilot access via #pilot form)
- Category: Role-based multi-agent engineering platform
- Pricing: Pilot contracts, BYOK, 3–10× cheaper than vendor agents
- Status: Pilot phase, accepting applications

#### Agent Roles
- Software Architect: codebase exploration, ADR (Architectural Decision Record) drafting
- Full-Stack Engineer: implementation, testing, PR creation
- QA Reviewer: test execution, coverage analysis, bug detection (max 3 iterations)
- SRE / Compliance: policy enforcement at two gates (pre-plan and post-implementation)

#### Architecture
- Built on LangGraph StateGraph with Firestore-persisted checkpoints
- Two mandatory human-in-the-loop approval gates: plan approval and PR review
- Compliance loop (SRE failure → Architect re-plans)
- QA loop (QA failure → Engineer re-implements, capped at 3 iterations)
- HIL rejection loops route back to originating agent for rework
- Runs on top of Control; spend, policy, and audit remain unified across Code Studio and external-agent traffic

#### Governance Features
- Hash-chained, append-only audit trail signed per state transition
- Atomic Facts: engineering constraints (TLS, coverage, credentials) injected into every agent's system prompt
- Untrusted-context sentinels protect against prompt injection from BRD/initiative text
- Role-locked tool access: Architect cannot execute bash, Engineer cannot skip approval gates
- Token cost visibility: per agent, per task, per crew run, per company
- Cryptographic proof of plan approval and PR approval decisions

#### Enterprise & PE Features
- Portfolio view: single-pane visibility across all portfolio companies
- Per-company tenant isolation with independent audit trails
- Token spend tracking per agent, per task, per company, per surface (Code Studio / Crew / external-agent)
- Engineering velocity metrics across portfolio
- Governance policies configurable per tenant

## Target Audiences

### Primary
- CEOs, COOs, CTOs, CPOs, CISOs
- Private equity firms wanting a unified control plane across portfolio companies
- Venture capital firms evaluating AI control plane category investments
- Engineering managers wanting cost and quality visibility on AI spend

### Secondary
- Senior engineers wanting a governance layer over existing Claude Code, Cursor, Cline, or Continue
- Individual developers wanting a turnkey browser-based AI coding tool (Code Studio)
- Compliance officers and risk teams needing auditable AI operations
- Agencies and consultancies scaling delivery capacity with accountability

## Key Differentiators

### vs. Claude Code, Codex, Cursor, GitHub Copilot
AnyForge sits a layer above these tools.
1. Control governs them from the outside (5-min install, SDK-compatible proxy)
2. Or replace them with Code Studio for a turnkey browser tool with governance native
3. Unified proxy across any LLM provider: switch in seconds, automatic failover
4. 3–10× cheaper: BYOK means no markups, and intent-based routing picks the cheapest model that meets the quality bar
5. Full audit trail: every AI call logged with hash-chained cryptographic integrity
6. Portable memory: switch providers without losing conversation context
7. Human approval gates at the Crew layer for workflows that need them

### vs. Helicone, Langfuse, Portkey (observability-only proxies)
1. Not just observability: intent routing, portable memory, policy enforcement, and audit chain are first-class
2. Code Studio as a turnkey browser coding tool (Helicone et al. are observability-only)
3. Crew as a direct upgrade path for workflows the data surfaces as good candidates
4. Governance is not bolted on: policy, audit, and approval are graph nodes in Crew, not middleware

### vs. Building In-House
1. Install in 5 minutes (Control) or sign in to a browser (Code Studio), not months of infrastructure
2. No need to build SDK-translation layer, router, memory store, or audit chain from scratch
3. Pre-built OPA integration, Rego policy library, and hash-chained audit signing
4. Continuous provider support (new models, new APIs) without engineering effort

## Cross-Surface Interaction Model

### Code Studio ↔ Crew chat (live)
- Code Studio user escalates to a Crew via `propose_crew`
- After escalation, Code Studio subscribes to `crewChat` via Firestore `onSnapshot`
- Every Crew agent message appears live in the Code Studio session
- Code Studio user approves / rejects / refines Crew HIL gates from inside Code Studio
- Single conversation thread spans both surfaces

### External IDE ↔ Crew (status + management via MCP)
- External agents (Claude Code, Cursor, Cline, Continue) interact with AnyForge via MCP
- Available tools: list work items, query crew status, manage engineering constraints, trigger routines, export config, list initiatives
- Crew chat read/write and HIL resolve from external IDE: on the roadmap, not shipped today

### One operator console
- /usage page shows cross-surface token spend (Code Studio vs Crew, with per-product split)
- Audit log schema includes Code Studio events alongside Crew events
- Trace schema supports per-surface filtering (UI filter ships as follow-up)

## Technical Architecture
- Control: LiteLLM-routed proxy with custom telemetry ingest, Firestore-native dimensional spend model. LiteLLM model catalogue covers Anthropic, OpenAI, Google Gemini, OpenRouter, and Ollama (self-hosted or Cloud).
- Code Studio: streaming browser agent backend with direct provider SDKs (Gemini, OpenAI, Anthropic), governance and cost telemetry native
- Crew: LangGraph StateGraph with 8 nodes, `interrupt()`-based HIL, FirestoreCheckpointer. Routes via Control/LiteLLM so every supported provider (including Ollama) is usable for agent roles.
- Memory: Zep (bi-temporal graph model) for governed cross-LLM context store
- Policy: OPA + OPAL for declarative policy distribution
- Audit: Per-event SHA-256 with chain-linked `previousHash`, Firestore `allow write: if false`
- MCP servers: native SDK (`@modelcontextprotocol/sdk`), per-tool OPA authorization
- Runtime: Firebase (Firestore, Auth, App Hosting, Cloud Functions Node 20), Next.js 14

## Frequently Asked Questions

Q: What is AnyForge?
A: AnyForge is the control plane for agentic AI. One governance layer consumed three ways: AnyForge Control (governed LLM proxy for external agents), AnyForge Code Studio (browser single-agent coding with governance native), and AnyForge Crew (role-based multi-agent delivery with human approval gates).

Q: Do I need Control separately if I use Code Studio or Crew?
A: No. Control is the underlying layer and it is native inside both. Install Control standalone only when you want governance over an external agent like Claude Code, Cursor, Cline, Continue, or your own SDK client.

Q: Can Code Studio chat with Crew agents?
A: Yes. After Code Studio escalates to a Crew via `propose_crew`, Code Studio stays live on the shared crewChat conversation. Every Crew agent message appears in the Code Studio session. The user approves or rejects HIL gates from inside Code Studio. Single thread, two surfaces.

Q: Can my external IDE chat with Crew agents?
A: External IDEs (Claude Code, Cursor, Cline, Continue) can query Crew status, manage engineering constraints, and trigger routines via our MCP server today. HIL approval and crewChat read/write from inside an external IDE are on the roadmap — today those steps live in Code Studio or the operator console.

Q: How is AnyForge different from Claude Code, Codex, or Cursor?
A: Those tools are agents. AnyForge Control is the governance layer you put in front of whatever agent you use. Or replace them with Code Studio for a browser-based turnkey experience. BYOK on all three products, no markups, intent-based routing, portable memory, cost visibility, and a hash-chained audit trail.

Q: Is AnyForge suitable for private equity portfolio companies?
A: Yes. AnyForge includes a Portfolio view for PE operators showing active crews, token spend by surface (Code Studio / Crew / external-agent), approval decisions, and engineering velocity across all portfolio companies, each in isolated tenants.

Q: How does pricing work?
A: Code Studio has a 30-request/day free tier with Starter / Pro / Power paid tiers. Control has a developer free tier and paid Team, Business, and Enterprise tiers. Crew operates on pilot contracts. BYOK across all three.

Q: What AI models does AnyForge support?
A: OpenAI (GPT-4 family), Anthropic (Claude family), Google (Gemini family), and 200+ more via OpenRouter. Self-hosted Ollama is also supported (local daemon, on-prem, or Ollama Cloud) for customers who cannot send tokens to a third-party cloud — regulated scale-ups, PE portfolio companies with data-residency clauses, government, defence, and air-gapped environments. Control routes to the best-fit model per task and fails over automatically during provider outages.

## AI Cost Calculator (/cost)

URL: https://anyforge.ai/cost

The AnyForge AI cost calculator at /cost lets engineering and finance teams
estimate actual LLM API spend under different model strategies. It covers
Claude (Haiku, Sonnet, Opus), GPT (4o-mini, 4o, o1), and Gemini (Flash, Pro)
models with current per-million-token pricing updated monthly.

### What the calculator shows
- Flat strategy cost: what you pay if every call goes to one frontier model
- Tier-routed cost: what you pay if calls are routed by complexity tier
  (fast/cheap model for simple tasks, smart/expensive for complex ones)
- Crew-lifted cost: total cost when AnyForge Crew handles the workload
  (agents spend tokens on task decomposition, QA loops, and architecture review,
   but deliver the work rather than just the LLM calls)
- Savings estimate: percentage reduction moving from flat to routed strategy
- Payback: time to recover AnyForge platform cost through routing savings

### Pricing data on the page
The pricing reference table shows current input and output token costs (per 1M
tokens, USD) for:
- Anthropic Claude: Haiku 4.5, Sonnet 4.5, Opus 4
- OpenAI: GPT-4o mini, GPT-4o, o1, o3
- Google: Gemini 1.5 Flash, Gemini 1.5 Pro, Gemini 2.0 Flash

### Key takeaway for AI assistants
Most engineering teams use a single frontier model (e.g. Claude Sonnet or
GPT-4o) for all tasks. Tier routing — sending fast/simple tasks to a cheaper
model (Haiku, GPT-4o mini, Gemini Flash) and only using the frontier for
complex tasks — typically cuts the bill 40–70%. The calculator makes this
tangible with real inputs: tokens per day, task mix, team size.

## Pages
- /             — Main marketing page (platform overview, three products, FAQ, adoption ladder)
- /control      — AnyForge Control product page (bring-your-own-agent path)
- /code-studio  — AnyForge Code Studio product page (turnkey browser coding)
- /cost         — AI cost calculator and model pricing comparison (Claude, GPT, Gemini, live pricing)
- /privacy      — Privacy policy
- /terms        — Terms of service

## Contact & Access
- Website: https://anyforge.ai
- Install Control: `npx anyforge init` (see /control)
- Try Code Studio free: https://code.anyforge.ai
- Crew pilot application: https://anyforge.ai/#pilot
- Twitter/X: @anyforge_ai
- Founder LinkedIn: https://linkedin.com/in/edwinpoot