Quotaflow
Search docs... ⌘ K
llms.txtOpenAPIDashboard
Serverless & Agents

Quotaflow Local Connector

Use the Local Connector when a developer belongs to multiple workspaces and needs the agent to know who is paying for a task.

The connector is local-first:

Ask an AI agent to install it

The Quotaflow dashboard separates setup into two paths:

Prefer letting the local AI agent do the installation and verification:

Set up this coding agent for Quotaflow on this machine.
Read https://docs.quotaflow.ai/codex/local-connector/ and https://docs.quotaflow.ai/llms-full.txt.
Install or run the Quotaflow qf CLI, then run qf doctor.
Prefer qf run -- <agent>. For OpenAI-compatible manual config, use base URL https://api.quotaflow.ai/openai/v1 and a Quotaflow key that starts with qf_.
Do not ask me to paste personal provider refresh tokens, auth files, browser cookies, ~/.codex/auth.json, or raw secrets into chat.
If this repo is unknown, keep it Personal until I confirm the company workspace/project. Do not enable team fallback without an explicit qf link confirmation.

Commands

qf doctor
qf link --workspace acme --workspace-type company --project backend
qf run -- codex
qf run -- claude
qf run -- cursor

qf run -- <command> is the universal launcher. It injects Quotaflow context into the child process without replacing the user's personal provider login.

From inside a repository:

qf link --workspace acme --workspace-type company --project backend

This stores a local binding:

repo_fingerprint -> workspace/project

After linking, qf doctor should show Confidence: confirmed and team fallback can be considered by policy. If a repo is not linked, qf reports Personal/default context.

Launch any agent

qf run -- codex
qf run -- claude
qf run -- opencode
qf run -- your-agent --with-args

The launcher sets:

The generated instruction file tells the agent which workspace/project is active, what payer is expected, and what not to upload.

Local private state

Codex, Claude Code, Cursor, and other agents continue to own their own login/session files. Quotaflow does not copy those files.

qf keeps its own local metadata under the Quotaflow home directory:

~/.quotaflow/config.json          # local hints such as qf use personal
~/.quotaflow/bindings.json        # repo_fingerprint -> workspace/project cache
~/.quotaflow/instructions/*.md    # generated agent instructions
~/.quotaflow/receipts/            # optional local receipt queue

Files written by qf use private file permissions when the platform supports them. Treat these files as metadata/cache only. Do not put personal provider refresh tokens, browser cookies, or ~/.codex/auth.json content into Quotaflow JSON files.

Company billing and fallback are still validated by Quotaflow cloud policy. A local file can speed up context resolution, but it cannot by itself authorize team wallet usage.

Agents can also read local gateway guidance:

curl http://127.0.0.1:8787/v1/instructions
curl http://127.0.0.1:8787/v1/link-required

/v1/link-required returns a structured qf_link_required response when a repo must be confirmed before team fallback.

Experimental process-scoped proxy

The connector has an opt-in process-scoped proxy flag for local experiments:

qf run --proxy -- claude

The proxy is process-scoped. It does not change the system proxy. The current qfd provider proxy is an MVP skeleton and returns 501 for provider request paths until your dashboard policy and connector version explicitly enable runtime proxy fallback. Use qf run -- <agent> for normal context injection.

Do not enable a global OS proxy for Quotaflow unless your administrator explicitly requires it.

What goes to Quotaflow cloud

Allowed:

Not allowed:

AI agents: start at /llms.txt, fetch /llms-full.txt, and parse /openapi.yaml for exact endpoint schemas.