# AnyForge — The Agentic Software Delivery Platform # https://anyforge.ai # For the full reference see: https://anyforge.ai/llms-full.txt AnyForge is the agentic software delivery platform: end to end, from discovery to delivery, with every step audited. Product signals become RICE-scored opportunities, opportunities become roadmap themes and spec-verified initiatives, and specs become shipped code — without leaving the platform. Delivery runs three ways over one governance layer — AnyForge Control (governed LLM proxy for teams using their own agents like Claude Code, Cursor, Cline, or Continue), AnyForge Code Studio (browser-based agentic coding workspace with governance native), and AnyForge Crew (governed multi-agent delivery with a mandatory human approval gate at PR review). Around them sits a full platform: product planning (signals, RICE, revenue-at-risk, Monte-Carlo capacity), spec-driven development, leadership insights and board packs, the AnyForge Oracle workspace assistant, MCP and skills marketplaces, observability and incident ingestion, coverage ingestion, source-code escrow, and living architecture documentation. Cost analytics, intent routing, portable memory, and a hash-chained audit trail are shared across every surface. One flat per-million-token platform fee (provider-agnostic, BYOK, no seats, no markups), 100M-token free trial on signup. Runs any model: Anthropic, OpenAI, Gemini, Bedrock, Vertex, 200+ via OpenRouter, open weights (GLM, DeepSeek, Qwen, Llama, Mistral), and custom/private LLM hosting on your own infrastructure (Ollama, vLLM, llama.cpp, dedicated gateway — on-prem, VPC, or air-gapped). ## Products ### AnyForge Control (bring your own agent) - URL: https://anyforge.ai/control - Install: `npx @anyforge/control init` (5 minutes, detects Claude Code / Cursor / Cline / Continue / raw SDK) - Governed LLM proxy: SDK-compatible drop-in endpoints for OpenAI, Anthropic, Gemini, OpenRouter - Private/self-hosted model routing: Ollama, vLLM, llama.cpp, dedicated LiteLLM gateway (air-gapped OK) - Per-call / per-agent / per-project cost + latency analytics; budgets that warn and never stop a run - Intent-based routing (rules, regex, embeddings) to the cheapest model that meets the quality bar - Portable memory that survives provider switches - Hash-chained, signed audit log of every call - Every call checked against the company IP allowlist and the key holder's role - Claude Max OAuth routing: bill agent traffic to a flat Claude subscription instead of metered API - BYOK, free to install, 100M-token trial ### AnyForge Code Studio (turnkey browser coding) - URL: https://anyforge.ai/code-studio — Try free: https://crew.anyforge.ai/studio - Browser-based agentic coding workspace: Files / Diff / Terminal / Metrics tabs, works on mobile - 200+ AI models, switch per request; native GitHub integration, main-branch locking by default - Quality-first: writes failing tests before fixes; repo-grounded plans; surgical git-applied edits - ~50 agent tools: terminal, web search, headless-browser screenshot verification, sub-agent spawning, plan mode, line-anchored inline PR reviews, coverage-measuring test runner - Built-in terminal: in-browser container with automatic cloud failover (npm test from your phone) - Semantic code search + cross-repo code graph shared with all agent surfaces - Builds UI to your org design system (root DESIGN.md + brand assets); the agent can connect repos and start codebase analysis from the conversation - Governance native: cost, routing, portable memory, audit built in - Compact live-presence avatars show teammates viewing shared Crews and Architect decompositions; private Solo sessions stay private - Escalates to Crew in one click; approve HIL gates without leaving the conversation ### AnyForge Crew (governed multi-agent delivery) - URL: https://anyforge.ai (self-serve: sign up at https://crew.anyforge.ai/login?signup=1) - One capable Engineer agent + on-demand Architect / Security & Compliance / QA specialists - Built on LangGraph with Firestore-persisted checkpoints; deterministic git-apply patching - Your test suite is the ground-truth gate; parallel specialist review of every PR diff, with overlapping findings consolidated into one item per underlying issue; an opt-in enforced coverage gate can hold the ship when PR-head coverage is below the codebase threshold - Integration validation for third-party APIs / payments / webhooks: crews verify real HTTP/REST integrations against a live sandbox without holding provider credentials or egress; the real response + async webhook evidence (via a connected observability MCP) attach to the PR gate - Read-only context repositories (read sibling repos; edits stay in the one target repo), scoped local runtime validation (worktree dev server + headless-browser proof on the PR card), structural (AST) code search, and source-to-target monorepo migrations - One mandatory human-in-the-loop approval gate at PR review, with cryptographic decision proof; the Approve gate is enforced server-side (chat card, control API, Approvals queue re-read live CI) - Three-pane PR review cockpit: AI walkthrough, click-to-comment diff (syncs to GitHub), in-thread approve/amend (amend feedback carries through to the specialist re-review); live run mission control with phase tracker and typed interaction cards - The autonomy dial: CI triage → capped auto-fix → size-limited auto-merge → Release Health auto-heal in safe-path envelopes; auto-merge defaults cascade org → program → initiative → gate - Automations: cron and HMAC-signed webhook crew dispatch; routine cards link to the crews they dispatched; read-only audit crews finish on a committed findings document + PR; automatic merge-conflict resolution - Per-run model control: dispatch with organization defaults, a reusable role-specific profile, or a one-off role-specific configuration snapshotted only to that Crew run; per-role reasoning-level control (low→max) mapped to each provider's native reasoning field - Model benchmarking: run one work item across two per-role run profiles as isolated arms on their own branches (neither can auto-merge); a side-by-side report and a deterministic recommendation (outcome → tests → open findings → cost → tokens → duration) name a winner; the operator picks the output to take forward and can retry a failed arm in one click - Spec-first auto-refine at the ship gate: the spec decides refine vs ship — unmet acceptance criteria earn another engineer round, an open critical finding never ships, and residual in-scope findings file as a single follow-up work item instead of burning refine rounds or blocking the merge - Operator Workflows: declarative, versioned multi-step orchestrations (agent / tool / parallel / map / loop / human-gate / crew steps with condition guards) compiled onto the same checkpointed runtime — durable, with token and cost budgets that warn and never pause a run (every spec edit is a version with rollback), gates land in the approvals inbox and push Slack/inbox/email with approve/amend/reject; dispatchable from Console, API, MCP, cron/webhook Routines, or an accepted production signal; authored conversationally with Workflow Genesis; a crew step fans governed crews out across the fleet - Production Watchdog + Playbooks: a managed, opt-in observability agent (cadenced + reactive) diffs ingested signals to surface new/spiking/worsening anomalies, has the Oracle explain them against connected observability MCPs, delivers a bell/Slack/email digest with per-run proof, and can auto-file critical ones as KTLO incidents; a browse→enable→customize Playbooks library ships bundled Operator Workflows (Cost Anomaly Watch, Release Health Report, Security Findings Triage) - Enforced engineering constraints (Atomic Facts) verified per-constraint at PR time ### The platform around them - Product planning: Signals → Opportunities (RICE + AI Suggest) → Themes/Roadmap → Initiatives (bounded, or long-lived to absorb an ongoing stream of work without a forced completion gate) - Community request board: public bug / feature / enhancement intake (screenshot attachable) gated on a complete ICE spec (Oracle-guided), anonymized Explore feed with upvoting (locked once shipped/declined), personal request tracking (Received → Under review → Planned → In progress → Shipped), and a delivery-pulse strip (open, crews working now, shipped in 30 days, median filed→shipped); every request is a real work item a crew can pick up, and shipped items announce back to the requester - Genesis agents draft the artifacts: PRDs with browsable UI mockups, program briefs, risk registers, status reports, specs, dashboards — all conversational; Program Genesis can also reconcile externally completed work through a separate-confirmation, audited close-out flow - Revenue-at-Risk analytics; Monte-Carlo capacity forecasting (P50/P85/P95) - Spec-driven development: ICE specs (Intent/Context/Expectations) + structured ICE-scored Program Briefs, both hard-gated at 100% ICE before the architect's plan is approved (per-field N/A waivers, audited owner override); briefs and specs are commentable inline (Google-Docs-style margin threads) - Initiatives can be marked Blocked on an external dependency (overlay that preserves delivery state, rolls up into a blocked-by-program report); a cross-initiative dependency map dispatches crews per node - AnyForge Oracle: a 360° intelligence layer over every codebase (code graph, docs, releases) correlated live with production observability (Datadog, Google Cloud Logging, AWS CloudWatch, Sentry) — from stack trace to root cause to dispatched fix crew in one conversation; read-only by default, BYOK, exportable answers, sandboxed run_code compute, Slack /oracle slash command and @AnyForge channel mentions; pin it to a crew at a HIL gate or halt for merge-safety / blast-radius / halt-diagnosis decision support; confirm-first operator action tier can dispatch / amend / halt / resume crews, approve gates, dispatch the architect, close a run as delivered, and connect repos + run their analysis; answers person-scoped delivery questions ("what did X ship / what's blocking them", role-gated); ⌘K Spotlight matches generated codebase docs by title/type/repo plus a semantic deep tier; every entity it presents links to its Console view - Insights: management overview + board packs, custom chat-built reports, usage, traces (waterfall, spend/error lenses, run-tree), audit (filterable ledger with SIEM export), health, delivery analytics (cost per completed crew, DORA, first-pass acceptance) with change-failure correlation, operator effectiveness scores (0–100), GCP cost drill-downs, two-ledger provider spend reconciliation (Anthropic/OpenAI/Google/OpenRouter) - Releases: auto-generated redacted release notes on tag push, changelog timeline, Discord announcements - Notifications: in-app inbox with @mentions, email, Slack DM, Discord — per-event preferences - Code-intelligence graph index: cross-repo call/import graph (TypeScript compiler API + tree-sitter for Go/Kotlin/Python/Java/Rust) + semantic embeddings search; SHA-gated incremental refresh on every merge; confidence-labelled edges; god-node/community/drift analysis; powers agents, ⌘K search, and interactive Code Graph views - Documentation engine: deep analysis on connect + always-on incremental doc patching on every merged PR (SHA-gated, changed-file→doc-type mapping, versioned with audit) + scheduled reanalysis; generated C4 diagrams, ADRs, L1 deployment topology with dependency arrows, Conway analysis, AI-readiness score, security posture, org-wide business narrative; full-text search, external share links, Markdown/PNG/PDF export; business capability map + Porter Value Chain projection with coverage heatmap - Organization design systems: agents read a repo's root DESIGN.md (tokens + guardrails, org default fallback) and managed brand assets as the visual source of truth — Studio, Genesis mockups, and crews build UI to your real identity, not a generic template - Config Packs: one-click installable bundles of agent configs, enforced constraints, routines and skills. These ship today — PCI-DSS, SOC 2 Type II, HIPAA Security & Privacy, HIPAA + SOC 2 (one cross-mapped control set for healthcare SaaS), NIST Cybersecurity Framework 2.0, NIST SSDF (SP 800-218), NIST AI RMF 1.0, EU Consumer Credit Directive 2 (Directive (EU) 2023/2225, BNPL/consumer credit), CNSA 2.0 Post-Quantum Cryptography (NSA's quantum-resistant algorithm suite), Software Escrow, and Startup (velocity, not compliance). Packs stack, seed editable operator-owned records, and the full company config exports/imports as a manifest. Listed on https://anyforge.ai/platform#config-packs - Universal MCP support: connect ANY MCP server (Streamable HTTP or SSE; OAuth auto-refresh, Bearer, API keys, or no auth; per-user or company-scoped) — plus a one-click marketplace: GitHub, Datadog, Sentry, Jira, Linear, Notion, Salesforce, Stripe (vendor API, or an AnyForge-hosted read-only revenue/payments surface), Cloudflare, CloudZero, AWS Cost & Billing (FinOps), Snowflake (Cortex Analyst/Search + SQL, or an AnyForge-hosted read-only warehouse-ops surface: credit spend, slow/failed queries, tasks, storage), Databricks (lakehouse ops), Google Drive/Calendar, Microsoft 365, Firebase, Aikido Security - Skills marketplace with curated registries + repo auto-detection, plus a bundled library of AnyForge engineering-discipline skills (plan / review / security / deploy / QA) scoped by role and language; AnyForge MCP server (~50 tools) for Claude Code/Cursor - Observability ingestion (Datadog, Sentry, New Relic, Grafana, CloudWatch, Aikido) → incidents (KTLO); OpenTelemetry (OTLP) metrics + trace ingest; readable trace links to initiatives and work items - Coverage ingestion (Jest, pytest, Go, LCOV, SonarQube, Codecov, Coveralls) - Source-code escrow: SHA-256 manifests, rebuild rehearsals, SFTP delivery to Escode/NCC - Teams, invitations, multi-domain email auto-join (claim several domains → one org); roles down to surface-locked board viewers - Strategy & delivery: strategy pillars, OKRs and key-result reading history; programs with milestones, Gantt and latest-start dates, and capacity plans; Work Queue with Kanban boards, saved filters, work-item SLAs and watching; a business-value board; custom reports with share links; customers and revenue at risk - Integrations: GitLab (GitLab.com and self-managed): crews, codebase analysis and merge-request review, alongside GitHub; native two-way Linear sync; the AnyForge extension for VS Code and Cursor (approvals, crews, work items and the Oracle in your editor); Slack; Oracle on WhatsApp; Google Meet meeting assistant (Chrome extension); Microsoft Teams notifications; email intake; Zendesk and Intercom escalations; Datadog alert auto-investigation in Slack ## Enterprise security & compliance - SAML and OIDC single sign-on with Require SSO enforcement and owner break-glass sign-in - Personal TOTP two-step verification, plus an MFA gate for AnyForge platform administrators - Maximum session length and idle timeout - Expiring API keys with a company maximum key lifetime; service accounts with role-capped scopes - Per-company IP allowlists (IPv4 and IPv6) covering the Console, API, MCP and CLI - Customer-managed encryption keys: envelope encryption under your own Cloud KMS key on the shared service; single-tenant Google Cloud deployment with CMEK on every data store via Terraform, in the data region you choose - Audit-log export to your SIEM: a read API with an audit:read scope, streaming to webhooks and SIEM sinks (Splunk HEC, Datadog Logs), an MCP tool and a CLI command - Hash-chained, tamper-evident audit trail with a daily chain verifier and a daily immutable archive that records chain heads - Data retention per data class, legal hold, and owner-requested company data deletion with a cooling-off period - Sandboxed command execution per crew: user and network namespaces, egress only through a proxy that refuses cloud-metadata and private addresses, required in production - CI supply-chain gates: blocking secret scan and dependency audit, a CycloneDX SBOM on every release, pinned actions and container images - Inbound rate limiting on sign-in, OAuth, public-link and webhook routes - Per-provider model kill switch; every switch and every provider fallover is audited - Public Trust Center with SOC 2 readiness tracking (SOC 2 programme in progress; no report issued yet): https://anyforge.ai/trust - Published sub-processor list: https://anyforge.ai/subprocessors - Vulnerability reports: security@anyforge.ai (RFC 9116 security.txt at /.well-known/security.txt) ## Target audiences - CEOs, COOs, CTOs, CPOs, CISOs - Groups of companies, agencies working for several clients, and private equity portfolios - Enterprises and growing engineering teams - Venture capital firms evaluating the agentic software delivery category - Engineering managers and senior engineers - Compliance officers and risk teams - Individual developers adopting governance over their existing AI tooling ## Key value propositions 1. The complete loop: from customer signal to merged, audited PR — one platform, one ledger, one fee. 2. One governance layer, three consumption surfaces. Zero workflow change on day one. 3. Any model, anywhere: frontier clouds, 200+ via OpenRouter, open weights (GLM, DeepSeek), or private LLM hosting on your own infrastructure — same governance and price for all. 4. Autonomy you can hand an auditor: hash-chained audit, cryptographic approval proof, per-role tool allow-lists. 5. Full visibility — per-call cost, latency, routing, traces, and board-ready summaries. 6. Portable memory — switch providers without losing context or decisions. 7. Escalation path from solo coding to governed multi-agent delivery, in-thread. 8. One flat per-million-token platform fee, BYOK, no seats, no markups; smart routing + caching typically save more than the fee costs. ## Pages - / — Homepage (platform overview, three products, adoption ladder, FAQ) - /platform — Full platform tour: plan, build, govern, see, extend, any model - /technology — Engineering deep-dive: architecture, determinism, security, stack - /control — AnyForge Control product page (bring-your-own-agent path) - /code-studio — AnyForge Code Studio product page (turnkey browser coding) - /pricing — Flat platform fee, BYOK model, free trial - /cost — AI cost calculator and model pricing comparison (Claude, GPT, Gemini) - /trust — Trust Center: security overview, compliance programme status, document requests - /subprocessors — Published list of sub-processors - /dpa — Data Processing Addendum - /privacy — Privacy policy - /terms — Terms of service ## Contact - Website: https://anyforge.ai - Install Control: `npx @anyforge/control init` (see /control) - Try Code Studio: https://crew.anyforge.ai/studio - Start Crew (self-serve, no application): https://crew.anyforge.ai/login?signup=1 - Agent skill for evaluating AnyForge: https://anyforge.ai/.well-known/agent-skills/evaluate-anyforge/SKILL.md (listed in /.well-known/agent-skills/index.json)