claude third party
A third-party Claude client doesn't have to run on someone else's server
Most guides about “Claude third party” stop at the definition, the app list, and the new billing rule. The part that decides whether you should trust one is the part none of them cover: where your account actually runs when the request goes out.
Direct answer (verified 2026-06-22)
A Claude third-party app is any tool that reaches Claude's models through your own Claude account or API key instead of Anthropic's first-party apps (claude.ai, the Claude desktop and mobile apps). If it asks you to sign in with Claude or to paste an Anthropic API key, it is a third party. Since early 2026, those requests draw from a separate prepaid balance Anthropic calls extra usage credits rather than your plan's normal allowance.
The detail that matters for trust is architectural. Some third-party apps take a token and call Claude from their server. A few, including Fazm, run Anthropic's own Claude Code agent locally on your Mac under the login already on your disk, so the request goes straight to Anthropic and nothing transits a third party's backend.
Source on third-party access and billing: Claude Help Center, “Use the Claude Agent SDK with your Claude plan”.
A few tools that reach Claude through your own account, not Anthropic's own apps:
The split nobody names: token-on-a-server vs agent-on-your-machine
“Third party” is a billing and authentication category, not an architecture. Two apps can both be third-party Claude clients and handle your account in completely different ways.
The first kind asks you to sign in with Claude, receives an OAuth token, and stores it. When you send a message, the request is made from their server using your token. That is convenient, and for plenty of tools it is fine. But it means a long-lived credential and the content of your prompts both live on infrastructure you do not run, and you revoke access by deleting the grant in your Claude settings.
The second kind never holds a token on a server because it does not have a server in the loop. It launches an actual Claude agent process on your computer, lets that process read the Claude login already sitting in your home directory, and the request leaves your machine directly. There is no grant to revoke because nothing was ever granted to a remote app; you just quit the program.
How Fazm, a local third-party Claude client, routes a request
Your Mac (Fazm)
Native app, no Fazm server in the request path
claude-agent-acp
Anthropic's Claude Code agent, spawned as a local child process
~/.claude login
Reuses the Claude Pro/Max login already on your disk
Anthropic API
Request goes direct; usage hits your own plan
An OAuth-style third-party app inserts its own server between the first and last box. A local agent client does not.
The part you can verify yourself
Claims about architecture are worth nothing if you cannot check them. Fazm is open source, so the “runs Anthropic's own agent locally” claim is not marketing, it is a line in a manifest and a process you can watch. In the repo, the ACP bridge declares the exact adapter and version it wraps:
Three things fall out of those lines. The agent is Anthropic's, not a reimplementation, so you get the same agent loop you already trust. It runs as a local process, so the request does not pass through a Fazm backend. And the session log is a file on your machine under ~/.claude/projects, which is exactly why Fazm can restore a chat after a Mac restart and fork it into a new window with the full prior context, without re-uploading anything.
Before you trust any third-party Claude client
- Can you read its source, or is the architecture a black box?
- Does the request run on your machine, or on the vendor's server?
- Does it reuse your own Claude login, or hold a token remotely?
- Does your chat history stay on your disk, or in their database?
- Can you point it at your own endpoint or proxy if you need to?
You do not need every answer to land the same way. A server-side app can be a perfectly reasonable choice for a quick integration. But you should know which side of each line a tool sits on before you sign it into your Claude account, not after.
The billing change that put “third party” in everyone's search bar
A lot of the recent interest in this topic traces to one shift: in early 2026 Anthropic separated third-party access from your plan's ordinary allowance. Requests made through third-party apps now draw from a prepaid pool called extra usage credits, distinct from both your subscription and per-token API billing. That is why a tool that worked yesterday can suddenly report that a request was rejected and ask you to add credit.
It does not change the architectural point. Whether your client runs locally or on a server, it still authenticates as you and still draws from your balance. What local execution changes is the blast radius if something goes wrong: a credential that never leaves your disk cannot leak from a vendor you have never audited. If you want the mechanics of the three billing pools, the Claude usage credits breakdown covers which pool a third-party request actually hits.
Adjacent guides
Keep reading
Stop Claude Code from compacting away your decisions
Why long Claude Code sessions silently drop earlier decisions, and how a client that keeps full history in context avoids it.
Point a Mac GUI at your proxied Claude subscription
Run CLIProxyAPI locally, then consume that endpoint from a native macOS app with persistent sessions and one model-selection gotcha.
How Claude usage credits actually work
Plan allowance, extra usage credits, and API billing are three separate pools. Here is which one a third-party app draws from.
Want a third-party Claude client whose source you can read?
Walk through how Fazm runs Anthropic's own agent locally, on your own login, with sessions on your disk.
Claude third party: common questions
What is a Claude third-party app?
A Claude third-party app is any tool that reaches Claude's models through your own Claude account or API key rather than through Anthropic's own first-party apps (the claude.ai web app, the Claude desktop app, the Claude mobile app). If a tool asks you to sign in with Claude or to paste an Anthropic API key, it is a third-party app. Since early 2026, requests made through third-party apps draw from a separate prepaid balance Anthropic calls extra usage credits rather than your plan's normal allowance. Examples of third-party Claude clients include code editors and agents like Cursor, Cline, Zed, Aider, and Fazm.
Are third-party Claude apps safe to use?
It depends entirely on the architecture, not on the label. The question that matters is where your account actually runs. Some third-party apps take an OAuth token (or your API key) and make the request from their own server, which means your prompts, your code, and a long-lived credential pass through infrastructure you do not control. Others run a Claude agent locally on your machine using the login that already lives on your disk, so the request goes from your Mac straight to Anthropic and nothing transits a third party's backend. Before trusting any third-party Claude client, find out which of these two it is, and prefer ones whose source you can read.
How is Fazm different from other third-party Claude apps?
Fazm does not reimplement Claude or proxy it through a Fazm server. It bundles Anthropic's own Claude Code agent (the @agentclientprotocol/claude-agent-acp adapter, pinned at version 0.29.2 in the repo) and spawns it as a local child process on your Mac. That process authenticates with the same ~/.claude login Claude Code already uses, so your Claude Pro or Max plan is what pays, and the request goes directly to Anthropic. Your full chat history is written to ~/.claude/projects on your own disk. Fazm is open source, so every one of these claims is checkable in the source at github.com/mediar-ai/fazm.
Do third-party Claude apps use my subscription or cost extra?
Both can be true depending on how the app connects. If it signs in with your Claude account, requests draw from your subscription plus the separate extra usage credit balance Anthropic introduced for third-party access in early 2026. If it uses an Anthropic API key, it bills per token against that key. Fazm uses your own Claude Pro or Max login, so usage hits your existing plan, and it also supports a custom API endpoint if you would rather route through your own gateway or proxy.
Where does my chat history go when I use a third-party Claude client?
With a server-side OAuth app, your conversation typically lives in that company's database. With a local agent client like Fazm, history is written to ~/.claude/projects/<encoded-working-dir>/<sessionId>.jsonl on your own machine, the same place Claude Code stores its sessions. That on-disk file is what lets Fazm restore a chat after a Mac restart and fork a conversation into a new window with the full prior context intact, without re-uploading anything to a server.
Can I point a third-party Claude app at my own API endpoint?
Only if the app exposes the setting. Many do not. Fazm has a Custom API Endpoint field that sets ANTHROPIC_BASE_URL under the hood, so you can route Claude-model traffic through a corporate proxy, a self-hosted gateway like CLIProxyAPI, GitHub Copilot, or any Anthropic-compatible endpoint. One gotcha: the override only affects Claude-model requests, so if your selected model is not a Claude model the endpoint silently receives nothing.
How do I revoke a third-party app's access to Claude?
For OAuth-based third-party apps, go to claude.ai settings and remove the app under connected applications; that invalidates the token it was issued. For a local agent client that uses your machine's existing Claude login rather than a granted token, there is no separate grant to revoke: you simply quit or uninstall the app, and it stops being able to read the local credential.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.