Recorded v1.0.0 engineering showcase · 3 AI coding CLIs · LiteLLM v1.105.1 · 2026

Stop runaway AI coding loops.

I built a team policy console, background spend watcher, and in-terminal request tool on top of open-source LiteLLM so every engineer gets a clear weekly AI allowance and forgotten night or weekend loops pause automatically.

When a small engineering team shares one raw API key across Claude Code, OpenCode, and Codex, a single stuck retry loop on Friday evening can quietly burn the month’s budget. Here, each engineer works on a virtual key with an 80% soft warning, a 5% grace buffer, and one-command top-up requests from the terminal.

TypeScript
Node.js
LiteLLM
MCP
PostgreSQL
Cost control

Explore walkthrough View on GitHub Hire Me

Static showcase of recorded v1.0.0 runs. No live model calls or cookies run in your browser.

Human approval boundary: Routine work inside a member’s budget runs automatically. Soft warnings at 80% notify without blocking, hard caps stop after a 5% grace buffer, and temporary budget increases or model exceptions require a manager’s click in the console.

Real example: hitting a cap

Same terminal session

1. Lena hits her weekly cap HTTP 422 · USD 21.00

Lena’s weekly allowance is USD 20.00 with a 5% grace buffer. At USD 21.00, the gateway pauses her key and returns a recovery command.
2. She asks for USD 5.00 from her CLI MCP tool or CLI

Instead of chasing a manager in chat, she requests a top-up directly from her coding session:

npx allowance request --amount 5 --reason "Finish release test suite"
3. Priya approves USD 3.00 until Friday +USD 3.00 approved

Manager Priya adjusts the grant to USD 3.00 in the console. Lena’s next prompt in the same terminal succeeds immediately, and the audit log records both numbers.

Manager console screenshot showing Priya approving USD 3.00 on Lena’s USD 5.00 request.

USD 144.97 → 0.55 Simulated weekend runaway loop cut from 5,000 calls to 40 calls when budget, off-hours cap, and 30-minute unattended pause combine.

3 CLIs · 18 / 18 Live verification passed across Claude Code (2.1.316), OpenCode (v2.1.3), and Codex (0.165.0) at USD 0.170827 total gateway spend.

69 audit events Recorded across 6 Crew accounts and 11 requests in a real 4-agent team test (USD 0.082512 metered across 369 spend rows).

317 / 369 rows Honest finding: streamed CLI responses logged USD 0 cost across 12,896,452 tokens, exposing a streaming cost-metering gap.

What I built

Open-source LiteLLM v1.105.1 enforces per-key hard budgets, but its built-in per-user management screens and approval workflows sit behind the enterprise license. I built the missing team layer using only open-source endpoints.

TypeScript Node.js 24 LiteLLM MCP PostgreSQL Playwright

Manager & member web console Server-rendered console on port 18080 with Overview, Member detail, Requests queue, Usage breakdown, Audit log, and CSV/JSON export.
Background spend & off-hours watcher Polls spend every 30 seconds, fires 80% soft warnings, blocks keys outside working hours once off-hours caps cross, and expires temporary grants.
In-terminal MCP & CLI request tools Engineers check remaining allowance (allowance_status), ask for more budget (allowance_request_more), or check approval status without leaving their editor.
OpenCode unattended loop guard Local plugin pauses a session after 30 minutes (40 calls at one call every 45 seconds) without a fresh human message.

Ten-step product walkthrough

Follow how engineers Lena, Omar, and Jonas interact with team manager Priya across ten recorded screens. Filter by scenario, swipe the screenshot carousel, or click any screen to zoom and pan.

1. Lena works inside budget

Actor: Lena (Member) · Weekly budget: USD 20.00 · Spend: USD 7.20 (36%)

Allowed · 200 OK

What happened

Lena runs her normal AI coding CLI against the team gateway on port 14000. Because her weekly spend (USD 7.20) is well below the 80% warning mark, the gateway forwards the request with zero extra prompts.

What to notice in the screenshot

In the manager overview, Lena shows a healthy green status bar at 36% of her USD 20.00 weekly allowance.

JavaScript is disabled. Direct links to all 11 walkthrough screenshots: 1. Lena works, 2. Soft warning, 3. Hard stop, 4. Request, 5. Partial approval, 6. Continues, 7. Model denied, 8a. Off-hours cap, 8b. Unblocked, 9. Audit log, 10. Usage.

Measured results & honest limits

Tested in three ways: a deterministic weekend runaway-loop simulation, a live 3-CLI verification against deepseek-v4.1-flash, and a 4-agent team test that exposed a real streaming cost-metering bug.

1. No limits (shared raw key)

5,000 steps · Ran until Mon 08:00 USD 144.97

Every retry succeeded all weekend until Monday morning arrival.

2. Weekly member budget only (USD 20.00 + 5% grace)

740 steps · Stopped Fri 23:02 USD 21.03

Hard budget stopped the loop once cumulative spend crossed the USD 21.00 grace limit (saving USD 123.94 vs. unguarded).

3. Off-hours cap only (USD 5.00 after 19:00)

289 steps · Stopped Fri 19:36 USD 8.03

USD 3.02 spent during work hours (17:30–19:00) plus USD 5.01 after 19:00 before the off-hours cap paused the key.

4. Budget + off-hours cap + 30-minute unattended pause

40 steps · Paused Fri 18:00 USD 0.55

The OpenCode plugin noticed 30 minutes without a human message and paused the loop after 40 calls.

Recorded weekend simulation outcomes (Friday 17:30 to Monday 08:00 Europe/Berlin, 45s interval)
Protection configuration	Stop trigger	Calls	Spend (USD)	Saved (USD)
No limits	None · ran until Monday 08:00	5,000	144.97	0.00
Member budget only	Weekly hard limit (USD 20.00 + 5% grace = USD 21.00)	740	21.03	123.94
Off-hours cap only	Off-hours cap (USD 5.00 after 19:00 reached at USD 5.01)	289	8.03	136.94
Budget + off-hours + unattended rule	Unattended pause after 30 minutes without user input	40	0.55	144.42

Live verification across 3 CLIs

18 / 18 steps passed

Each real coding CLI ran the same 6-step live check against deepseek-v4.1-flash with a test budget of USD 0.05: allowed call, blocked model (HTTP 403), budget/key stop (HTTP 422/401), MCP extra-budget request, manager approval in the web console, and a working retry.

Claude Code · 2.1.316 (Claude Code) USD 0.056733 6 / 6 steps passed · Allowed call, 403 model block, budget stop, MCP request, approval, retry

OpenCode · opencode v2.1.3 USD 0.058732 6 / 6 steps passed · Plus OpenCode plugin unattended rule check (USD 0.000025)

Codex · codex-cli 0.165.0 USD 0.055337 6 / 6 steps passed · Virtual key updated in place without restarting CLI

Total live-verification gateway spend across all 3 CLIs and the plugin check: USD 0.170827 (01 Nov 2026, using elevated test token pricing so a USD 0.05 budget is reached in two calls).

Real 4-agent team test

5 / 7 probes passed

Four AI coding workers and a team manager ran through the live gateway across two teams: Crew (6 accounts: 5 workers + 1 leader, 11 requests, 69 audit events) and Night (3 accounts: 2 workers + 1 leader, 0 requests). Total gateway spend was USD 0.082512 across 369 spend rows (including USD 0.022452 from scripted feature probes).

Mira (Crew) USD 0.056181

Hit her USD 0.05 test budget, requested USD 0.20, received a partial USD 0.10 approval, and continued working.

Nils (Crew) USD 0.000019

Blocked from an unapproved model (HTTP 403), submitted a model access request, and received a manager denial with a note.

Edda (Crew) USD 0.000000

Exercised MCP status checks and request cancellation; all 48 of her streamed CLI calls recorded USD 0 cost.

Bo (Night) USD 0.002574

Ran off-hours probes; non-streaming probe calls metered USD 0.002574 while streamed CLI calls recorded USD 0.

What went wrong: streaming USD 0 rows

The most valuable result of running a real multi-agent team through the proxy was catching a metering failure that unit tests missed. Out of 369 gateway spend rows in LiteLLM_SpendLogs, 317 rows recorded USD 0 cost even though they carried 12,896,452 tokens.

Why it happened: Interactive CLI sessions use server-sent event (SSE) streaming against deepseek-v4.1-flash via an OpenAI-compatible upstream. When the upstream stream omitted final usage-cost metadata matching the custom model price map, LiteLLM v1.105.1 wrote the token counts to Postgres with spend = 0. Non-streaming calls priced accurately, which is why Mira’s hard limit (USD 0.056181) and the 18 / 18 live CLI checks triggered as designed, whereas Bo’s Night off-hours cap did not trip during streamed turns (leaving 5 of 7 scripted probes passing).

317 of 369 rows

Recorded USD 0 cost while carrying 12,896,452 streamed tokens during the 4-agent team test. Disclosed in full so anyone adopting open-source LiteLLM with custom OpenAI-compatible streaming endpoints knows to verify stream usage accounting (stream_options: {"include_usage": true}).

How the architecture works

Budget enforcement happens on the team gateway rather than trusting the developer’s laptop, while local CLI helpers and MCP tools keep requests fast inside the terminal.

1. On the member machine

Claude Code, OpenCode, or Codex points at the gateway on 127.0.0.1:14000 using a member virtual key stored with 0600 permissions. The local MCP server exposes allowance_status and allowance_request_more, and the OpenCode plugin pauses unattended loops after 30 minutes.

2. On the team server

Open-source LiteLLM v1.105.1 and PostgreSQL (127.0.0.1:15432) enforce hard budgets and model allowlists on every call. The Team AI Allowance service (127.0.0.1:18080) runs a 30-second watcher for 80% soft alerts, off-hours caps, and grant expiry.

3. For the manager

Managers review requests in the web console, approve full or partial top-ups with an expiration time, or deny with a written reason. Alerts deliver to local SMTP (127.0.0.1:18025) and webhook (127.0.0.1:18090) endpoints.

Why local edits cannot bypass a cap

Even if an engineer disables the local OpenCode plugin or edits local config files, the virtual key’s hard budget and model allowlist live on the LiteLLM gateway and Postgres database.

30-second watcher polling

Hard budget caps block inline on the next gateway call, while off-hours caps and 80% soft warnings evaluate on the 30-second watcher tick.

Single-host Docker stack

Designed for a small engineering team running on a single Linux/macOS host with loopback-bound ports rather than multi-region federation.

Open-source LiteLLM boundary

Uses only open-source LiteLLM proxy key routes and Postgres spend tables, with zero dependency on enterprise-licensed modules.

Local setup & verification

The stack runs on Node.js 24 and Docker Compose, binding all five services to 127.0.0.1 by default: console on 18080, LiteLLM gateway on 14000, PostgreSQL on 15432, mail catcher on 18025, and webhook catcher on 18090.

Start the stack and seed demo data

git clone https://github.com/sabers13/team-ai-allowance.git
cd team-ai-allowance
make setup
make demo
# Open http://127.0.0.1:18080

Member CLI & verification targets

node src/cli/index.ts status
node src/cli/index.ts request more-budget --amount 5 --reason "Finish tests"
make release-check
make e2e
make sim

LiteLLM boundary & 9 limits

Team AI Allowance uses only open-source LiteLLM proxy routes (/key/generate, /key/update, /key/info, /key/block, /key/unblock) and its Postgres spend tables. Nine operational boundaries are documented:

Nine documented operational boundaries in v1.0.0
Boundary	Behavior in v1.0.0	Operational impact
1. Streaming USD 0 cost rows	317 of 369 team-test rows logged USD 0 across 12,896,452 streamed tokens	Upstream OpenAI-compatible streams must emit usage chunks or use token-count fallback pricing
2. Watcher poll interval	Off-hours caps and 80% warnings evaluate every 30 seconds	A burst of calls inside one 30-second window can briefly overshoot an off-hours threshold
3. Unattended plugin scope	Client-side inactivity pause ships as an OpenCode plugin	Claude Code and Codex rely on gateway hard budgets and off-hours caps unless wrapped
4. Client-side plugin trust	The unattended pause runs on the member workstation	Treated as a convenience brake; hard financial enforcement stays on the server
5. Shared team timezone	Working hours are configured per team (e.g. Europe/Berlin)	Distributed teams across continents should group members into regional teams
6. Passphrase session auth	Console uses role passphrases and HTTP-only session cookies	Built for internal VPN/SSH access rather than public internet exposure without SSO
7. In-flight call completion	Budget checks happen before a model call starts	A large single call that starts at 99% budget completes into the 5% grace buffer
8. Single-node Postgres	State lives in one local PostgreSQL container	Requires standard volume backups for production durability
9. Custom model price table	Non-standard model IDs need explicit input/output token rates in LiteLLM config	Unmapped models default to zero cost unless added to the gateway model list

Project code and documentation are released under the MIT License (Copyright 2026 Saber Sojudi). Self-hosted typography and icons are used under their respective open licences: IBM Plex Sans (SIL OFL 1.1), Inter (SIL OFL 1.1), and Feather Icons (MIT).