Deference MCP
Give your agentits own budget.
Connect your coding agent once. It checks credit, picks models and runs them within your limit.
- Claude Code
- Codex
- Cursor
- VS Code
- Claude
- Gemini CLI
- OpenCode
- Zed
- Cline
- Goose
One command. One sign-in.
Add the server to your agent and approve it in the browser. No key to paste, nothing to install.
Setup guideclaude mcp add --transport http --scope user deference https://deference.si/api/mcp
claude mcp login deferencecodex mcp add deference --url https://deference.si/api/mcp
codex mcp login deference{
"mcpServers": {
"deference": {
"url": "https://deference.si/api/mcp"
}
}
}{
"servers": {
"deference": {
"type": "http",
"url": "https://deference.si/api/mcp"
}
}
}https://deference.si/api/mcp{
"mcpServers": {
"deference": {
"httpUrl": "https://deference.si/api/mcp"
}
}
}{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"deference": {
"type": "remote",
"url": "https://deference.si/api/mcp"
}
}
}{
"context_servers": {
"deference": {
"url": "https://deference.si/api/mcp"
}
}
}{
"mcpServers": {
"deference": {
"type": "streamableHttp",
"url": "https://deference.si/api/mcp",
"disabled": false,
"autoApprove": []
}
}
}extensions:
deference:
name: Deference
type: streamable_http
uri: https://deference.si/api/mcp
enabled: true
timeout: 300Sign in once, in the browser
After that the agent refreshes its own access, headless runs on the same machine included.
It checks before it spends.
Your agent reads credit, prices and spend from your account, not from memory, then runs the model it picks.
What it can do> tidy the build script, then split the test step
● Read package.json, scripts/build.mjs
● Edited scripts/build.mjs (+12 −4)
● Bash npm run build
└ built in 38s, 0 warnings
Build tidied. Lint and tests now run side by side.
> How much can you spend?
deferencedeference_get_balance
└ credit $42.10 · budget $8.40 of $10
$42.10 of credit on the account. This connection has $8.40 left of its $10 budget.
>
> tidy the build script, then split the test step
● Read package.json, scripts/build.mjs
● Edited scripts/build.mjs (+12 −4)
● Bash npm run build
└ built in 38s, 0 warnings
Build tidied. Lint and tests now run side by side.
> How much can you spend?
deferencedeference_get_balance
└ credit $42.10 · budget $8.40 of $10
$42.10 of credit on the account. This connection has $8.40 left of its $10 budget.
> Find a model for a 100k-token review under $1.
deferencedeference_list_models
└ 3 models fit · sorted by price
Three fit. The lowest price is $0.31 for this review; the strongest is $0.88.
>
> tidy the build script, then split the test step
● Read package.json, scripts/build.mjs
● Edited scripts/build.mjs (+12 −4)
● Bash npm run build
└ built in 38s, 0 warnings
Build tidied. Lint and tests now run side by side.
> How much can you spend?
deferencedeference_get_balance
└ credit $42.10 · budget $8.40 of $10
$42.10 of credit on the account. This connection has $8.40 left of its $10 budget.
> Find a model for a 100k-token review under $1.
deferencedeference_list_models
└ 3 models fit · sorted by price
Three fit. The lowest price is $0.31 for this review; the strongest is $0.88.
> Get a second opinion on this diff.
deferencedeference_run_model
└ anthropic/claude-sonnet-5.5 · $0.0042
Done. 3 issues found. Cost $0.0042, $8.39 left.
>
> tidy the build script, then split the test step
● Read package.json, scripts/build.mjs
● Edited scripts/build.mjs (+12 −4)
● Bash npm run build
└ built in 38s, 0 warnings
Build tidied. Lint and tests now run side by side.
> How much can you spend?
deferencedeference_get_balance
└ credit $42.10 · budget $8.40 of $10
$42.10 of credit on the account. This connection has $8.40 left of its $10 budget.
> Find a model for a 100k-token review under $1.
deferencedeference_list_models
└ 3 models fit · sorted by price
Three fit. The lowest price is $0.31 for this review; the strongest is $0.88.
> Get a second opinion on this diff.
deferencedeference_run_model
└ anthropic/claude-sonnet-5.5 · $0.0042
Done. 3 issues found. Cost $0.0042, $8.39 left.
> What did you spend this week?
deferencedeference_get_usage
└ 247 requests · $4.11
$4.11 across 247 requests. Most went to one model.
>
Compares models before it spends
It lists models by price and context, and estimates a call's cost from token counts first.
Knows what it has left
Each connection has its own budget, starting at $10. The agent reads it before a paid call.
Every paid call has a ceiling.
Each paid call carries a cost cap and an idempotency key. A retry never charges twice, and every answer states what it cost.
Spend budgetSpend you can read back by app
Usage splits what each connection spent by app, model and key, over a day or up to 90 days.
A cap on every call
The cap defaults to $0.50 and goes up to $5. A call that could cost more is refused before any charge.
One charge per idempotency key
A completed retry with the same key and inputs gets the first receipt again, not a second charge.
Keys it makes, limits you set.
Your agent creates, rotates and revokes its own keys for code that calls the API. Each starts at $10 and never goes above $100.
Limits on agent keysMarked with the app that made it
Keys an agent creates carry an Agent tag and their own limit. Raise a limit or remove a key here.
You decide what it touches.
Reading is always on. Spending and key management take one switch at sign-in. Pause or disconnect an app in one step.
AgentsEvery app in one list
What each connection may do, what it has spent, and when it last ran. Open one to change it.
One switch for spending
Read is always on. Write lets the app run models and manage its own keys. Turn it off and it reads.
What stays with you
The agent never adds credit. Raising a budget or a limit, deleting keys and disconnecting stay in the dashboard.
Every call, on record.
Each tool call and model request shows up in Agents and Activity with the app that made it, the model, tokens and cost.
Activity