Public beta

Route, govern and record every AI agent, on infrastructure you own.

Gateways route, security proxies enforce policy, cost tools cap spend. In line, PromptForce does all three on every request and cuts what it costs. Over time it compounds: the record of what worked moves tasks from rented models to models you own.

Watch the 1-minute film

See how it works ↓Compare with gateways →

Your machine or your cloud, never ours · paid from savings, never per seat

A lemon in a steel hand press, one drop falling, on a yellow background
01 / The squeezeLess waste, more good stuff.Get 18 tools in one local binary, combining 520k+ GitHub stars
Pay for outcomesSavings, not seats.

Free until PromptForce saves you $500. After that, you pay a share of what it saves. No per-seat fees.

Private firstNo SaaS. Ever.

Your laptop or your private VPC. Prompts and code never pass through us.

Rented to ownedWe grow with you.

We help you move work from rented frontier models to models you own, one task at a time, when your own record says it's ready.

How it works

One layer between every agent and every model.

Modular: swap any part for one you already run.

YOUR AGENTS Coding · pass-through option: a local sandbox per agent · no Docker Non-coding · brokered key per agent Claude Code Codex OpenCode Research Docs and ops Your own apps pass-through brokered tokenpool 20 PROVIDERS · ONE POOL · YOUR KEYS MODEL LABS Anthropic OpenAI Google Gemini xAI Mistral DeepSeek Meta Z.ai CLOUDS AWS BedrockBedrock Azure AI FoundryAzure AI Google VertexVertex GATEWAYS AND FAST INFERENCE Groq Cerebras SambaNova NVIDIA OpenRouter Vercel Cloudflare OpenCode O OFFLOAD LAYER · SELF-HOSTED Ollama vLLM SGLang llama.cpp your hardware · personal data stays here + any OpenAI- or Anthropic-compatible API PROMPTFORCE · AGENTOPS HUB modular · every component swappable PromptForce, in lineON THE WIRE · IN-LINESavings add up on every turn, beyond your provider's own cachingsaved+$+$+$ Multi-level routing Cache-awarecache hits Model familyright tier Effortthinking Free tokensby policy In-line coachingstatus line nudges System-1Jev · Laya Compressionrecoverable Cachestable prefixes FinOpsper agent Budget capsperson · team Policy + auditallow · ask · deny Agent profilesrules per agent recordcheaper next turn PLUGS INTO YOUR OBSERVABILITY OpenTelemetry Datadog Grafana Honeycomb Prometheus Langfuse traces, cost and audit events, exported OVER TIME · THE COMPOUNDING ZONE Recordtrainslocal models Sovereignrented toowned FluencyAI-nativefrom usage KV ZonesvLLMSGLang PLUGS INTO YOUR IDENTITY Any OIDC Okta Entra ID Auth0 Keycloak budgets, profiles and audit per person Option: runs in its own local sandbox Option: runs in its own local sandbox Option: runs in its own local sandbox Option: runs in its own local sandbox Option: runs in its own local sandbox Option: runs in its own local sandbox Evidence mapped for your assessor to review (preview) IMMUTABLE COMPLIANCE LEDGER signed · hash-chained · write-once · verify offline SOC 2 ISO 42001 EU AI Act NIST AI RMF NIST 800-171
Coding agents (Claude Code, Codex, OpenCode) pass through to their own provider after optimisation; non-coding agents get a brokered key each, with the provider swappable behind it. Both connect to the PromptForce AgentOps Hub, which runs on your machine or in your VPC with no SaaS. Agents can optionally run in their own local sandbox. PromptForce handles coaching, multi-level routing with System-1 models (Jev and open-weight Laya), recoverable compression, cache and token savings, FinOps and budget caps, and policy packs with audit, then calls your model providers with your keys. It plugs into your observability tools (OpenTelemetry, Datadog, Grafana, Honeycomb, Prometheus, Langfuse) and your identity provider (any OIDC: Okta, Entra ID, Auth0, Keycloak), writes a signed, hash-chained compliance ledger (immutable on write-once storage) with evidence mapped to SOC 2, ISO 42001, the EU AI Act, NIST AI RMF and NIST 800-171, and feeds its record into the compounding zone so each turn makes the next one cheaper.

Your agents

Coding agents · pass-through

  • Claude Code
  • Codex
  • OpenCode

Keep the tool and sign-in you use. Each request is optimised, then passed through to your provider.

Non-coding agents · brokered

  • Research
  • Docs and ops
  • Your own apps

Each agent gets one key with its own policy and budget cap. The providers behind it are abstracted away.

PromptForce AgentOps Hub · on your machine or VPC

  • Token optimisationstable cache, less repeat work
  • Multi-level routingrules, then System-1 models, then frontier
  • System-1 modelsJev and open-weight Laya
  • Compressionsmaller context, same answer
  • Budget capsper person, team and task
  • Policyallow, ask or deny, audited
  • Recordcost and outcome per turn

20 providers and any compatible API, one token pool

  • Anthropic
  • OpenAI
  • Google Gemini
  • xAI
  • Azure AI Foundry
  • AWS Bedrock
  • Google Vertex
  • Mistral
  • Meta
  • Vercel
  • OpenCode
  • DeepSeek
  • Groq
  • Cerebras
  • NVIDIA
  • OpenRouter
  • SambaNova
  • Cloudflare
  • Z.ai
  • Offload layer, self-hosted: Ollama, vLLM, SGLang, llama.cpp
  • + any OpenAI- or Anthropic-compatible API

Your keys, your accounts. Free and local models only if you turn them on.

Over time: the compounding zone

Every turn makes the next one cheaper.

  1. The record trains your own models. What worked, at what cost, on your work, becomes the evidence for moving a job to a local model.
  2. Less code to read. A short brief of your codebase replaces whole files, so each turn needs less context.
  3. KV Zones (preview). Cache-placement hints for your own vLLM, SGLang and llama.cpp servers, so private and neo-cloud infrastructure keeps the right context warm.

Launch film · 1:07

See it in a minute.

The app's setup, its dashboard and the codebase brief, as they run. Figures are 30 days on our own machine; repository and contributor names are stand-ins. Your savings depend on your workload.

Where it lives

Start with one team. Grow to your cloud. Never ours.

Built in Go. Four signed binaries with optional model sidecars. No container platform to run, no runtime to manage, nothing hosted by us.

PromptForce sits at the centre of rings you own: your machine, your team hub, your cloud or cluster, and an air-gapped wall. Model providers sit outside, reached only on your keys. PF machine team hub cloud · cluster air-gapped wall model providers · your keys

Why not a gateway

Not another cloud gateway. A layer you own.

QuestionCloud AI gatewaysPromptForce
Where it runsTheir serversYour machine, or a server in your private VPC. No SaaS.
Your prompts and codePass through a third partyGo straight to your AI provider
Your usage historyStored in their databaseStays on your disk; one anonymous daily summary, off with one switch
If the vendor goes offlineYour agents stopNothing changes: we are not in the path

Full comparison with LiteLLM, Portkey, Cloudflare, Kong, TrueFoundry and OpenRouter →

What's under the hood?For anyone tired of yet another AI-generated site: the measurements, the features we switched off, and the bugs that taught us.

Pricing

Free until it saves you $500.

$0

No trial clock, card or account. After that, a share of what it saves. Never per seat.

$0 to $400Savings add up while you work.$400A heads-up at 80%.$50015 more days to add a licence key.

Only savings beyond your provider's own prompt caching count. After that, PromptForce hands Claude Code and Codex back to their direct connection, so your agents keep working.

Research we're working on

Questions we measure before we claim.

01Does agent-written code last?

Following each change to see whether it is kept, rewritten or reverted weeks later.

02Remember only what was checked.

Agents keep a lesson only when the work it came from passed its checks.

03When can a local model take the job?

Your own checks, not ours, decide when routine work moves off frontier models.

04Only the context that gets used.

Measuring how much of what an agent reads actually changes the answer.

05Passing tests isn't proof.

Early tests suggest "tests passed" alone says little about whether a change is right.

We publish what we find, including what didn't work. Under the hood →

A few good questions

Before you take a sip.

Is the $500 a credit or a guaranteed saving?

Neither. It is the savings threshold for PromptForce's free use, not a cash credit and not a promise of how much you'll save. Savings depend on your workload. PromptForce counts only what it adds beyond your provider's own prompt caching, from the provider's billing fields.

Will my AI subscription actually last longer?

PromptForce cuts avoidable token use and can route suitable turns to other models. How far your plan stretches depends on your workload, provider limits and settings. The app shows plan usage and warnings so you can decide what to do next.

Do I have to switch AI providers?

No. Start with Claude Code or Codex and the provider you already use, and add other providers or local models when you want to. You bring your own subscriptions and API keys.

What does "vitamin C" mean here?

The C stands for context. PromptForce prepares the meaningfully viable context for each turn, using repeatable checks and fast small models, before the expensive model gets to work. Graph databases and context stuffing are the heavy dose; most turns do fine without them.

What leaves my Mac?

PromptForce runs locally. Your prompts and keys go only to the providers you configure. Once a day the licence check sends a salted machine hash and the savings meter reading, so a reinstall can't reset the offer; it never sends prompts, keys or your email. An anonymous daily usage summary (counts only) is on by default; turn it off in Settings. Details: what leaves your machine.

What if I'm on Linux?

The app ships for macOS and Windows. On Linux, two scripts install the parts the app runs:

curl -fsSL https://promptforce.ai/install.sh | bash && curl -fsSL https://promptforce.ai/rb-install.sh | bash

Each script is readable before you run it: install.sh and rb-install.sh.

Make every token worth the squeeze.

Less waste on every turn, better context before it.

Talk to us about your team
The squeeze
PromptForce.ai / Public beta · 01 / 05

0:00 / 0:20
Read the full pitch