MAI2API

Use Microsoft's MAI-Thinking-1 as a local API. MAI2API runs a small proxy on your PC that speaks the OpenAI (/v1/chat/completions), Anthropic (/v1/messages) and OpenAI Responses (/v1/responses) APIs and translates them to the Microsoft Playground web app, using your own signed-in session. Point Claude Code (or any OpenAI/Anthropic client) at it and the model answers — tools included, via emulation (see Limits).

Unofficial and unsupported. It drives a consumer web UI with your own account; if the Playground changes, this can break. Prefer the official API (Microsoft Foundry, model MAI-Thinking-1) if you need a contract.

How it works

Claude Code ──► mai_proxy.py (127.0.0.1:18787)
                  │  Bearer = your Playground session token (~65 min)
                  ▼
            playground.microsoft.ai  (tRPC + SSE)
                  ▲
   token supply   │  rpa_keeper.py: resident headless Edge on a
                  │  dedicated profile (CDP), renews the token every
                  │  few minutes, clicks through account-picker /
                  │  "stay signed in" screens when they appear
   supervision    │  guardian.py: restarts the proxy if it dies
   alerts         │  notify.py: Windows toast + optional ntfy push
                  │  when a human sign-in is needed again

Requirements

• Windows 10/11, Python 3.10+ (pip install -r requirements.txt)
• Microsoft Edge (preinstalled on Windows) for the dedicated profile
• A Microsoft account that can open https://playground.microsoft.ai
• Optional: Claude Code for the coding-agent setup

Setup (once)

git clone https://github.com/MasayukiTa/MAI2API.git MAI2API
cd MAI2API
powershell -ExecutionPolicy Bypass -File .\setup.ps1
python harvest_now.py --setup   # sign in once in the window that opens
python selftest.py              # expect: SELFTEST OK

setup.ps1 creates a dedicated virtual environment, registers logon auto-start for the guardian and the keeper, and starts both.

Runtime state lives outside this folder: the token, the dedicated browser profile, sessions, logs and pid files are kept in %LOCALAPPDATA%\MAI2API (override with the MAI2API_STATE_DIR environment variable), so the source tree never holds live credentials. The one-time sign-in creates the browser profile there; after that, everything is unattended.

Daily use

Nothing to start — the pieces auto-start at logon. For Claude Code:

mai-claude.cmd

(It sets ANTHROPIC_BASE_URL=http://127.0.0.1:18787, ANTHROPIC_API_KEY=dummy, ANTHROPIC_MODEL=sonnet.) Model names follow an explicit alias policy: any client-supplied name maps to the single upstream model, and responses always report the real upstream id (mai-thinking-1-latest). Set MAI_PROXY_MODEL_STRICT=1 to reject names outside the known alias families with a 400 instead. Any other client can use http://127.0.0.1:18787/v1/chat/completions directly.

Feeding a token from somewhere else also works: POST a raw Easy Auth id_token to http://127.0.0.1:18787/__harvest with the X-MAI-Harvest-Key header (the key is generated on first run in the state directory, harvest_key.txt), or write the token to token.txt in the state directory.

Optional: phone alerts

If the session ever truly dies (password change, policy re-evaluation), the keeper cannot fix it alone — it can at least tell you. Create notify_config.json:

{ "ntfy_topic": "your-random-topic-name" }

and subscribe to that topic in the ntfy app. The keeper alerts on sign-in loss, repeated failures, and recovery. A Windows toast is always attempted as well.

Files

file	role
mai_proxy.py	the API proxy (OpenAI + Anthropic surfaces)
rpa_keeper.py	token keeper: headless Edge + RPA clicks, 4-min passes
headless_harvest.py	one-shot harvest the proxy spawns on a stale token
guardian.py	crash watchdog for the proxy (PID-exact restarts)
harvest_now.py	one-time visible sign-in / manual harvest
notify.py	toast + ntfy notifications used by the keeper
selftest.py	health + one live completion
setup.ps1	one-time setup (auto-start, dependency, first launch)
mai-claude.cmd	Claude Code launcher

Stop switches: create guardian_stop.txt / rpa_stop.txt / harvest_stop.txt in the state directory to stand the respective piece down.

Endpoints: /v1/chat/completions, /v1/messages (plus count_tokens, an approximation), /v1/responses, /v1/models, and the operational trio /health (token validity), /health/live (process liveness), /health/ready (503 until the token is valid), plus /metrics (request/upstream/auth counters as JSON). All responses stream truly: upstream deltas are flushed as they arrive.

Limits (measured, honest)

• Throughput: the Playground enforces an account-level rate limit — bursts of ~20–25 streams are fine, sustained throughput is ≈10 completions/min. The proxy paces and backs off on 429s, so clients see waits, not errors.
• Tools are emulated: the Playground offers no native tool calling. The proxy teaches the model a text protocol and converts its replies into tool_calls/tool_use; execution happens in your client (e.g. Claude Code runs Read/Write/Bash/WebSearch itself). Calls are validated against the declared tools (name allowlist + argument schema): an invalid call is fed back for one repair attempt, then dropped — never executed blindly.
• Usage numbers are estimates: the Playground reports no token counts, so usage is estimated from character counts and marked (usage_estimated / "estimated": true). It is never silently 0.
• Not supported: image input, previous_response_id (rejected with a 400 — send full input each time), exact token counting. The legacy browser-extension harvester is superseded by the headless keeper and is not needed.
• System prompts: the Playground accepts only a short system field and largely ignores it, so the proxy folds system text and tool definitions into the conversation (cap via MAI_PROXY_STUFF_CAP, default 280k chars, oldest turns dropped).
• One machine, one account, localhost only. The token and the browser profile in the state directory are credentials — they are git-ignored and physically separated from the source; keep it that way. requests.log records request lines only (method, path, status) — never bodies or prompts.

Troubleshooting

• selftest.py says the token expired and nothing heals: the session needs a human again — run python harvest_now.py --setup.
• Proxy down: the guardian restarts it within about a minute (check proxy.log in the state directory). A process that stays alive but stops answering is terminated and restarted after ~5 minutes of failed liveness checks, capped at 3 per hour (guardian_state.json records every restart).
• Keeper quiet: check rpa_loop.log in the state directory; make sure no rpa_stop.txt exists there.

Development

python -m pytest tests -q   # 40+ unit tests, no network needed

Release checklist: pytest green · CI green · selftest.py OK on a real install · a streamed and a buffered completion through each surface (/v1/chat/completions, /v1/messages, /v1/responses) · /health/ready returns 503 with an expired token and 200 with a fresh one.