alex 0.1.8

Alexandria is a local daemon that turns your AI subscriptions (Claude Max, ChatGPT/Codex, SuperGrok, Gemini) into one OpenAI/Anthropic-compatible endpoint on 127.0.0.1:4100 — with a credential vault, automatic token refresh, model routing, format translation, full trace capture, and usage/limit tracking.

Point any coding harness (Claude Code, Codex CLI, grok, opencode, …) at it and every request is authenticated with the right subscription, billed to the right bucket, and captured to SQLite for inspection.

Features

• Credential vault — imports OAuth tokens from Claude Code, Codex, grok, and gemini CLIs; stores them in ~/.alexandria/accounts/ (0600); refreshes them itself (Anthropic, OpenAI, xAI)
• One endpoint, every provider — /v1/messages, /v1/chat/completions, /v1/responses, /v1beta/models/{model}:generateContent with cross-format translation and model routing (claude-* → Anthropic, gpt-* → Codex, grok-* → xAI, gemini-* → Gemini)
• Login flows built in — alex auth login claude|codex|grok (PKCE paste, loopback, and xAI device-code flows); also exposed over HTTP so GUIs can drive re-auth
• Trace capture & sessions — every request/response stored with tokens, cost, latency; group runs with x-session-id, tag with x-alexandria-* headers, search body text, stitch transcripts
• Trace Browser & TUI — a two-pane live trace browser in the menu bar app, alex tui in the terminal, alex traces --json for scripts
• Limits & health — subscription plan windows (5h/7d) with utilization and reset times, per-provider heartbeats, alex ping, alex status
• Cost analytics — per-model requests/tokens/cost with subscription-vs-API billing buckets (/admin/analytics)
• Dario mode — optional generational supervisor for the @askalf/dario Anthropic upstream with health probes, npm auto-update, and rolling restarts
• macOS menu bar app — live gauges, re-auth windows, ping checks, window-reset alerts in macos/ (AlexandriaBar)
• Harness smoke tests — alex harness run executes frozen CLI harnesses (claude, codex, grok, …) in Docker against the proxy and verifies traces land
• Zero-downtime upgrades — ./install.sh --upgrade blue-greens the daemon on a shared port (SO_REUSEPORT)
• Cross-platform CLI — Linux, macOS, and Windows binaries on every release (cargo install alex)

Install

cargo install alex        # installs the `alex` and `alexandria` binaries
# or from a checkout:
./install.sh              # release build → /usr/local/bin/alex (+ alexandria symlink)
./install.sh --service    # also run at login (launchd/systemd)

Quick start

alex auth import          # pull credentials from your existing CLI logins
alex daemon --background  # start the daemon on 127.0.0.1:4100
alex status               # accounts, limits, health at a glance
eval "$(alex env)"        # point ANTHROPIC_/OPENAI_/XAI_ env at the proxy

Re-auth a subscription any time with alex auth login claude|codex|grok, watch live traffic with alex tui, and check window utilization with alex limits.

Format translation

Alexandria speaks four API dialects on the way in and routes to whichever provider owns the model you name — translating the request, the response, and the streaming events in between. Point any client's SDK at the proxy and use any model; the wire format is converted for you (all conversions pivot through the Anthropic Messages shape internally).

The upstream is chosen purely by the model name, so e.g. an OpenAI Chat request naming claude-opus-4-8 is translated to Anthropic and back; a Gemini request naming gpt-5.5 is translated to Codex and returned as Gemini candidates. Streaming works across the matrix — SSE events are re-synthesized in the client's dialect.

Serving the Gemini CLI with a different model

Point the Gemini CLI at the proxy and let it run on, say, gpt-5.5 — requests arrive as Gemini, get rewritten to OpenAI/Codex upstream, and responses convert back to Gemini shape:

eval "$(alex env)"   # sets GOOGLE_GEMINI_BASE_URL, GEMINI_API_KEY, bearer mechanism, GOOGLE_GENAI_USE_GCA=false
gemini --model gpt-5.5 --prompt 'Reply with only PONG'

One-time: the Gemini CLI must use API-key auth, not its Google login. Set it in ~/.gemini/settings.json:

{ "security": { "auth": { "selectedType": "gemini-api-key" } } }

(or pick "Use API key" once in the CLI's /auth menu). Then alex env points it at the proxy — GOOGLE_GEMINI_BASE_URL has no /v1, since the SDK appends /v1beta itself, and GEMINI_API_KEY_AUTH_MECHANISM=bearer sends the key as Authorization: Bearer. Temp tokens from POST /admin/run-keys carry the same env block in their exports field, so a tagged run can drive the Gemini CLI too.

Direct curl (non-streaming and streaming):

curl -H 'Authorization: Bearer <local_key>' -H 'Content-Type: application/json' \
  -X POST 'http://127.0.0.1:4100/v1beta/models/gpt-5.5:generateContent' \
  -d '{"contents":[{"role":"user","parts":[{"text":"Reply with only PONG"}]}]}'

curl -N -H 'Authorization: Bearer <local_key>' -H 'Content-Type: application/json' \
  -X POST 'http://127.0.0.1:4100/v1beta/models/gpt-5.5:streamGenerateContent?alt=sse' \
  -d '{"contents":[{"role":"user","parts":[{"text":"Reply with only PONG"}]}]}'

To use your Gemini subscription as an upstream instead (name a gemini-* model), alexandria authenticates via the gemini-cli OAuth token it imported and Google's Code Assist API. Individual accounts are onboarded automatically; Workspace/Enterprise accounts must supply a GCP project with the Code Assist API enabled — set gemini_project = "your-gcp-project" in ~/.alexandria/config.toml (or GOOGLE_CLOUD_PROJECT).

Sessions & trace tagging

Generate client credentials for the proxy, then tag requests so every trace from a run lands in one named session:

eval "$(alex env)"        # exports ANTHROPIC_/OPENAI_/XAI_ vars pointing at the proxy
alex credentials --json   # same thing as JSON (alias: creds); --host host.docker.internal for containers

SESSION="experiment-42"
curl -s "$OPENAI_BASE_URL/chat/completions" \
  -H "authorization: Bearer $OPENAI_API_KEY" \
  -H "x-session-id: $SESSION" \
  -H "x-alexandria-task: sparql-university" \
  -H "x-alexandria-job: cove-sparql-1" \
  -H "x-alexandria-trace-tag: attempt=1" \
  -d '{"model":"gpt-5.5","messages":[{"role":"user","content":"hi"}]}'

Every response carries an x-alexandria-trace-id. Typed headers (x-alexandria-harness|task|model|job) and free-form x-alexandria-trace-tag: key=value tags are captured with each trace. Collect a session back out:

alex traces --session "$SESSION" --json                 # CLI
curl -H "x-api-key: <local_key>" \
  "http://127.0.0.1:4100/admin/traces?session=$SESSION" # HTTP
curl -H "x-api-key: <local_key>" \
  "http://127.0.0.1:4100/traces/sessions/$SESSION/transcript"  # stitched transcript
curl -H "x-api-key: <local_key>" \
  "http://127.0.0.1:4100/traces/search?text=professor&session=$SESSION"  # body-text search

Trace Browser

The menu bar app's Trace Browser gives the same data a UI: two-pane sessions + live transcript, an omni bar combining free text with model:, harness:, task:, job:, tag:key=value, status:, run: and session: filters, live/pin modes, and per-turn token/cost breakdowns.

macOS menu bar app

macos/ contains AlexandriaBar, a Swift menu bar app that shows daemon health, subscription limit gauges, and account status — with in-app re-auth (device codes, paste flows), ping checks, and notifications when a subscription needs attention.

cd macos && ./Scripts/run.sh   # build + launch dist/AlexandriaBar.app

Re-auth helpers

When a subscription expires you get a notification and a one-click fix. Each provider gets the flow that suits it: Codex opens the browser and finishes automatically via the localhost callback (above), Grok shows an xAI device code you can enter from any device, and Claude takes the pasted code#state. The same flows are served by the daemon (POST /admin/auth/login/start, poll GET /admin/auth/login/<id>), so any UI can drive them — the terminal equivalent is alex auth login <provider>.

Crates

Development

./alex daemon             # dev shim (cargo run)
./test.sh                 # tiered tests: unit | wire | harness | dario
cd macos && swift test    # menu bar app tests

License

Dual-licensed under MIT or Apache-2.0, at your option.