AnyForge Control

Govern every AI call.
Change nothing else.

AnyForge Control is the governance layer underneath the AnyForge platform. For teams using their own agents (Claude Code, Cursor, Cline, Continue, or a raw SDK client), Control ships as a governed LLM proxy that installs in five minutes. Zero workflow change. Immediate visibility.

Using AnyForge Code Studio or AnyForge Crew? Control is already native there — nothing separate to install. This page is for the "bring your own agent" path.

4-step install · 5 minutes

  1. Sign up at anyforge.ai — free, no card. You get a 100M-token free trial on signup.
  2. Add your provider API keys at /settings/secrets — Anthropic, OpenAI, Google, or OpenRouter. Keys live in Google Secret Manager, never in our database.
  3. Run the CLI in your repo:
    $ npx anyforge init
    Detects Claude Code, Cursor, Cline, Continue, or a raw SDK client. Mints a scoped AnyForge API key, writes the proxy config.
  4. Open your dashboard at crew.anyforge.ai — cost analytics, routing decisions, audit chain. Your next LLM call shows up within seconds.

Free to start with a 100M-token trial allowance. After trial: $1.00 per 1M tokens, provider-agnostic — same rate whether you use cloud LLMs or self-host. No subscription, no seats, no daily caps. Routing + caching typically save customers more on wasted model spend than the platform fee costs — the net effect is savings, not cost.

Capabilities

Visibility, routing, and governance. Opt-in per rule.

Unified proxy

One endpoint in front of OpenAI, Anthropic, Gemini, and OpenRouter. SDK-compatible drop-in, no prompt rewrites.

Self-hosted Ollama

Run models on your own Ollama daemon (local, on-prem, or Ollama Cloud). Same governance layer applied to every call — for air-gapped, data-residency, or regulated deployments.

Cost analytics

Per-call, per-agent, per-project spend. Dimensional rollups. Alerts at configurable thresholds.

Intent-based routing

Rules, regex, or embedding-based routing. Send each call to the cheapest model that meets the quality bar.

Portable memory

Shared context that survives provider switches. Same conversation history, different model, no rebuild.

Hash-chained audit

Signed, append-only log of every prompt, response, and tool call. Regulator-ready.

Optional OPA policy

Redaction, budget caps, PII blocking, allow-lists. Turn each rule on when you need it, not before.

Claude Max routing (OAuth)

Already on Claude Max? Paste an `sk-ant-oat-…` token and route agent traffic through your subscription instead of metered API rates — the flat Max quota absorbs the same workload at no marginal cost. Fallback-key slot keeps you running if Max quota saturates.

Already on Claude Max?

Route through your subscription, not a metered API account.

AnyForge accepts two flavours of Anthropic credential on the Settings → API Providers → Anthropic field: sk-ant-api-… (standard API key, per-token at Anthropic's API rates) and sk-ant-oat-… (OAuth token minted by claude setup-token, billed against your Claude Max / Pro subscription).

Why it matters: Per-token API rates add up fast on agent workloads. The flat Max quota absorbs the same traffic at no marginal cost (within Anthropic's rate limits). Optional fallback-key slot keeps the harness running if Max saturates — falls through to a metered key automatically, with a notice on /usage.

See docs › Claude OAuth tokens (beta) for setup.

What you learn on Control

Three questions the data answers in the first week.

  • What did this feature cost in tokens?

    Per-call spend, grouped by agent, project, or user. The answer nobody on your team can produce today.

  • Which model pays for itself on this kind of task?

    Model-by-model latency, success rate, and cost-per-outcome. Pick the cheapest model that meets the quality bar with evidence.

  • Which workflows are ready for multi-agent delivery?

    Repeated tasks with real stages, dependencies, and a human approver show up as Crew candidates on the dashboard.

When you are ready

Graduate to AnyForge Crew. Keep Control running underneath.

Crew is governed multi-agent delivery: a single capable Engineer drives the build with Architect, Security, and QA specialists available as on-demand advisors, plus a parallel specialist review of the PR diff and one mandatory human approval gate at PR review (with one opt-in exception: Release Health auto-merges small safe-path hotfix PRs without operator review when you enable it). Cryptographic audit trail on every call. You turn it on per project. Control keeps running underneath, so spend and audit stay unified across both planes.

You do not have to move to Crew. Many teams run Control indefinitely. It is a valid end state.

Apply for Crew pilot →Back to overview