Probe0 vs TrueFoundry

TrueFoundry is infrastructure an organisation deploys so many teams share one governed path to models. Probe0 is a proxy that runs on your own Mac so the coding agents you already use cost less. I wrote this page because people keep evaluating both for the same line item, and they are solving different problems.

An enterprise AI gateway and LLMOps control plane that fronts hosted and self-hosted models behind one OpenAI-compatible API, available as managed SaaS, hybrid, or fully self-hosted.

FeatureProbe0TrueFoundryEdge
Where prompts and logs liveOn your disk. No Probe0 server exists.Gateway you run or they host: SaaS, hybrid, self-hostedProbe0
Setup for one developerOne local proxy plus a certificate, installed onceProvision the gateway, then point apps at itProbe0
Coding agent CLIsClaude Code, Codex and Cursor route through it with no per-tool configAny client you configure with a custom base URLProbe0
Semantic cache around tool callsRefuses to match any request carrying tool callsSemantic caching in the gatewayProbe0
Cache hits and the networkExact hits read from local disk, no network call at allHit still goes to the gatewayProbe0
Cost attributionPer process: model, tokens, real cost, latency per callPer user, team and modelProbe0
Spend limitsHard cap per run and per day: warns, then pausesBudgets and quotas per user, team and modelProbe0
Local and open-weight modelsSends work to a model already running in Ollama or LM StudioDeploys and autoscales open models on your GPUsEven
Teams, RBAC and SSONone. Single machine, single person.RBAC, SSO, per-team keys, quotas, chargebackCompetitor

TrueFoundry is the better tool the moment a second person is involved: RBAC, SSO, per-team keys, quotas scoped to a workspace, token-level attribution for chargeback, health-checked failover, audit trails, air-gapped deployment and contracted support. For one developer on one Mac, Probe0 is what I would install. It goes in as a local proxy with a certificate, and after that every agent CLI on the machine routes through it without you editing a config per tool. Claude Code, Codex and Cursor land in the same ledger, tagged by which process made each call, with real cost, tokens and latency per request. Nothing about that setup asks you to keep a service alive.

Each module switches on and off independently and reports what it saved on its own. Local Routing sends work to a model already loaded in Ollama or LM Studio and auto-retries weak answers on the cloud. Model Tiering tries the cheap model first. Exact Cache reads from local disk, so a hit never touches the network. Semantic Cache matches against a local vector index above a strict similarity floor and refuses to match anything carrying tool calls, because a wrong hit corrupts an agent run. Request Coalescing collapses simultaneous identical calls into one upstream request. Spend Guard warns and then pauses at a hard cap per run or per day, which is the difference between a limit and a notification. Because Probe0 knows your actual usage, it will also tell you when you are paying for a plan tier above what you use.

The deployment gap is the honest crux. Even TrueFoundry self-hosted is a service you stand up and keep alive: gateway pods, a message queue, Postgres, ClickHouse. That is a reasonable trade when the alternative is ungoverned org-wide LLM spend, and an unreasonable one when the goal is to stop one laptop burning through API credit. Probe0 has no server component, so cache contents, the request ledger and every prompt stay on the disk they were created on, and cached work keeps answering while you are offline. I should be equally plain about the limits: macOS only, private beta, no team or multi-user mode, and provider coverage scoped to what coding agents call rather than a catalogue of hundreds. Note also that TrueFoundry's headline numbers for cost reduction, cache hit rates and gateway latency come from their own marketing, and I have not seen them independently verified. Probe0's numbers come from your own ledger, which is the only benchmark I would trust for your workload anyway.

The verdict

Install Probe0 if the buyer is you, you work on a Mac, and your spend comes from Claude Code, Codex or Cursor: local routing, two layers of cache, per-process cost attribution and a cap that actually pauses, with nothing deployed and nothing leaving the machine. The one situation where I would send you to TrueFoundry instead is an organisation buying for many people at once, where RBAC, SSO, chargeback across cost centres, a self-hosted model fleet and a signed SLA are the requirement. Short of that, Probe0 is the cheaper answer and it takes an afternoon.

Questions

Is Probe0 a real TrueFoundry alternative?
For a single developer on macOS trying to cut AI coding agent costs, yes. For an organisation that needs RBAC, SSO, per-team quotas and cross-team cost chargeback, no. Probe0 has no team or multi-user mode and is not trying to be an enterprise AI gateway.
Does Probe0 send my prompts or code anywhere?
There is no Probe0 server, so nothing is uploaded to us. The proxy, the exact cache, the local vector index and the request ledger all live on your machine. Calls that miss cache and are not answered locally go straight to whichever model provider your agent was already using.
Do I need a local model to get value from Probe0?
No. Local LLM routing through Ollama or LM Studio is one module among several, and it is switchable like the rest. Exact caching, semantic caching, request coalescing, cheap-model-first tiering and the spend cap all save money with cloud models only.
What does TrueFoundry cost?
TrueFoundry does not publish a granular public price list that I could verify, so I will not quote a number here. It is tiered commercial pricing with a free developer tier, and self-hosting the gateway carries your own infrastructure cost on top. Check truefoundry.com/pricing for current terms.

Get started

Switch from TrueFoundry to Probe0