CLAUDE.md

@@ -143,6 +143,16 @@ store / metadata sidecar.
   four chips) with the CLI alias attached. A newly released model appears in the
   composer with no code change. Falls back to a small static list when the key is
   missing or the call fails.
+- `openrouter.py` — the **OpenRouter catalogue** (`GET /api/models/openrouter`):
+  every model a pi.dev session can run on, with its $/Mtok (input / output /
+  cache-read), context window, parameter size and capability tags. The data is
+  **APPE's** (`appe.dev.gabvdl.xyz/api/models/openrouter.json`, a daily
+  models.dev sync), cached in-process for 6h with a disk copy under `/data` so a
+  restart with no network still lists models. Deliberately separate from
+  `/api/models`: that endpoint is the composer's short chip list, this is the few
+  hundred rows the model browser searches. It also **prices** pi sessions —
+  `conversations.rates_for()` reads a model's real rates from here instead of the
+  flat qwen estimate, and only an id the catalogue doesn't know falls back.
 - `notify_audio.py` — the sound a notification made, replayable in the browser
   (`POST /api/notifications/audio` → a WAV, behind the ▶ button on an expanded
   notification card). There is **no synthesizer here**: it POSTs to the homelab
@@ -531,6 +541,23 @@ glyphs render synchronously all over the app, so they can't come from a lazy
 chunk, and pulling ~1500 icons into the main bundle to serve one settings page
 is a bad trade.
 
+### Picking a pi model: the OpenRouter browser
+
+The model select lists the curated entries per harness, but on **pi** the real
+catalogue is a modal: `Browse OpenRouter…` opens `OpenRouterBrowser` — a
+`FuzzyList` over every OpenRouter model, each row stating the three rates that
+bill an agent run (**in / out / cache-read**, $ per million tokens), the model's
+parameter size and its context window, with capability chips (tools, reasoning,
+vision, open weights, free) to narrow the list. Cheapest first, so an unfiltered
+open lands on what costs least to try.
+
+Picking one sends its full id (`openai/gpt-oss-20b`) as `--model`; the runner
+already routes **any** vendor-prefixed id to OpenRouter and bare ids to the local
+EVOX2 provider, so no allow-list has to be kept in step. The frontend reads the
+same shape: `isOpenRouterId()` (a slash — Claude ids never have one) is what
+makes an arbitrary pick resolve to the pi harness on resume, without the
+catalogue being loaded.
+
 ### Effort (claude) vs. thinking (pi)
 
 How hard a run works a turn is **one knob shown two ways**, per harness — never

backend/conversations.py

@@ -16,6 +16,8 @@ import json
 import pathlib
 import re
 
+import openrouter
+
 
 # Claude Code session ids (= main transcript file stems) are UUIDs.
 _UUID_RE = re.compile(r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}"
@@ -438,12 +440,17 @@ def _worktree_events(text) -> list[dict]:
 
 # Per-model price in USD / million tokens (input, output). Cache tiers are
 # multiples of the input price (read ≈ 0.1×, 5m write ≈ 1.25×, 1h write ≈ 2×).
-# `qwen` covers the pi-harness OpenRouter sessions; a *bare* qwen id (no vendor
-# prefix) is the local EVOX2 LM Studio provider, which is free. These are the
-# fallback estimates — a pi transcript record carrying the runner-mirrored
-# `costUSD` (pi's own per-message provider cost) overrides them (see `feed`).
+# `qwen` is the *last-resort* estimate for a pi-harness OpenRouter session whose
+# model isn't in the catalogue; a *bare* qwen id (no vendor prefix) is the local
+# EVOX2 LM Studio provider, which is free. These are the fallback estimates — a
+# pi transcript record carrying the runner-mirrored `costUSD` (pi's own
+# per-message provider cost) overrides them (see `feed`).
 PRICES = {"opus": (5.0, 25.0), "sonnet": (3.0, 15.0), "haiku": (1.0, 5.0),
           "fable": (10.0, 50.0), "qwen": (0.15, 1.0), "qwen-local": (0.0, 0.0)}
+# What a cache read costs, as a multiple of the input rate — Anthropic's flat
+# 10%. OpenRouter models carry their own absolute cache price instead (see
+# `rates_for`), because a cache read there is per-model policy, not a fixed cut.
+CACHE_READ_MULT = 0.1
 MAX_BLOCK = 8000  # cap a single text/result block so payloads stay bounded
 
 
@@ -456,19 +463,40 @@ def _real_model(model: str | None) -> bool:
     return bool(model) and not model.startswith("<")
 
 
-def price_for(model: str) -> tuple[float, float]:
+def rates_for(model: str) -> tuple[float, float, float]:
+    """(input, output, cache-read) price in USD / million tokens for a model.
+
+    **A pi-harness session is priced from the OpenRouter catalogue** (APPE's
+    daily models.dev sync — see `openrouter.py`), so a run on any of the several
+    hundred pickable models is costed at that model's real rates, cache read
+    included. Only an id the catalogue doesn't know falls through to the flat
+    qwen estimate below. Claude models keep the built-in table and the 10%
+    cache-read cut.
+    """
     m = (model or "").lower()
+    if "/" in m:  # vendor-prefixed ⇒ an OpenRouter model on the pi harness
+        hit = openrouter.rates(model)
+        if hit:
+            return hit
     if "sonnet" in m:
-        return PRICES["sonnet"]
-    if "haiku" in m:
-        return PRICES["haiku"]
-    if "fable" in m or "mythos" in m:
-        return PRICES["fable"]
-    if "qwen" in m:
-        # Vendor-prefixed ids (qwen/…) run on OpenRouter; bare ids are the
-        # local EVOX2 LM Studio provider — free.
-        return PRICES["qwen"] if "/" in m else PRICES["qwen-local"]
-    return PRICES["opus"]
+        pi, po = PRICES["sonnet"]
+    elif "haiku" in m:
+        pi, po = PRICES["haiku"]
+    elif "fable" in m or "mythos" in m:
+        pi, po = PRICES["fable"]
+    elif "qwen" in m or "/" in m:
+        # Vendor-prefixed ids run on OpenRouter (priced above when known); bare
+        # qwen ids are the local EVOX2 LM Studio provider — free.
+        pi, po = PRICES["qwen"] if "/" in m else PRICES["qwen-local"]
+    else:
+        pi, po = PRICES["opus"]
+    return pi, po, pi * CACHE_READ_MULT
+
+
+def price_for(model: str) -> tuple[float, float]:
+    """(input, output) $/Mtok — `rates_for` without the cache-read rate."""
+    pi, po, _ = rates_for(model)
+    return pi, po
 
 
 def norm_usage(u: dict) -> dict:
@@ -500,19 +528,20 @@ def usage_tokens(u: dict) -> int:
 
 
 def usage_cost(u: dict, model: str) -> float:
-    pi, po = price_for(model)
+    pi, po, pc = rates_for(model)
     pi /= 1e6
     po /= 1e6
+    pc /= 1e6
     return (u["input"] * pi + u["cacheWriteUnits"] * pi
-            + u["cacheRead"] * 0.1 * pi + u["output"] * po)
+            + u["cacheRead"] * pc + u["output"] * po)
 
 
 def usage_cache_cost(u: dict, model: str) -> float:
     """The share of a turn's cost that is replayed cached prompt (billed at 10%
-    of the input rate). The rest is fresh tokens the model actually had to read
-    or write this turn."""
-    pi, _ = price_for(model)
-    return u["cacheRead"] * 0.1 * (pi / 1e6)
+    of the input rate on Claude, at the model's own cache rate on OpenRouter).
+    The rest is fresh tokens the model actually had to read or write this turn."""
+    _, _, pc = rates_for(model)
+    return u["cacheRead"] * (pc / 1e6)
 
 
 def _block_text(content) -> str:

backend/main.py

@@ -45,6 +45,7 @@ import gitdiff
 import goal as goalmd
 import memories as memories_mod
 import models as models_mod
+import openrouter as openrouter_mod
 import cron as cron_mod
 import ctas as ctas_mod
 import forms as forms_mod
@@ -2053,6 +2054,22 @@ def list_models():
             "default": models_mod.DEFAULT_MODEL}
 
 
+@app.get("/api/models/openrouter",
+         responses=_r(schemas.OpenRouterModelsResponse))
+def list_openrouter_models():
+    """Every model a pi.dev session can run on OpenRouter, cheapest first.
+
+    Not part of `/api/models` on purpose: that endpoint is the composer's short
+    pickable list (one chip per Claude family + the curated pi entries), while
+    this is a few hundred models the model browser searches through. Prices
+    ($/Mtok in / out / cache), context window and parameter size all come from
+    APPE's catalogue — the homelab's model-data source of truth (openrouter.py).
+    """
+    return {"models": openrouter_mod.catalogue(),
+            "source": openrouter_mod.CATALOGUE_URL,
+            "generatedAt": openrouter_mod.generated_at()}
+
+
 class SpawnBody(BaseModel):
     prompt: str
     model: str | None = None
@@ -4027,6 +4044,10 @@ def _startup():
     threading.Thread(target=_watch_sidecar_runs, daemon=True).start()
     # Cron jobs: fire scheduled agent sessions (see cron.py).
     cron_scheduler.start()
+    # OpenRouter catalogue (prices + sizes from APPE): pulled off the request
+    # path so the first model-browser open and the first priced transcript
+    # already have it (openrouter.py).
+    openrouter_mod.warm()
     # Conversation CTAs: the buttons a finished conversation offers (ctas.py).
     # The seeded Complete prompt is written into the workspace on first run;
     # there is no scheduler — a CTA runs when it is pressed.

backend/openrouter.py

@@ -0,0 +1,194 @@
+"""
+The OpenRouter catalogue — every model a pi.dev session can be spawned on.
+
+The composer's model select used to offer two hand-written pi entries (one
+OpenRouter qwen, one local EVOX2 qwen). OpenRouter serves several hundred, and
+picking between them is a *price* decision, so the picker now lists the lot with
+what each one costs.
+
+**The prices are not ours.** They come from APPE
+(https://appe.dev.gabvdl.xyz), the homelab's model-data app, which syncs
+models.dev daily and publishes its catalogue as static JSON
+(`/api/models/openrouter.json`, one entry per model with $/Mtok in / out /
+cache-read, context window, mined parameter count, capability tags and a
+measured tokens/sec). Keeping a second price table here is exactly the bug this
+avoids: when OpenRouter re-prices a model, APPE's next daily sync carries it and
+this module picks it up within the TTL.
+
+Three layers so the picker is never empty:
+  1. in-process cache (``CACHE_TTL_S``);
+  2. a disk copy under ``/data`` — written on every successful fetch, read when
+     APPE is unreachable, so a restart with no network still lists models;
+  3. nothing — the picker falls back to the curated ``PI_MODELS_DEFAULT`` list
+     in ``models.py`` and the flat qwen price estimate in ``conversations.py``.
+
+The catalogue is also what **prices** a pi session: ``rates()`` hands
+``conversations.py`` the per-model $/Mtok, so an OpenRouter run is costed at its
+real rates instead of a one-size-fits-all qwen guess. (A pi transcript that
+carries pi's own ``costUSD`` still wins over any estimate — see
+``conversations.feed``.)
+"""
+
+import json
+import os
+import pathlib
+import threading
+import time
+import urllib.error
+import urllib.request
+
+# APPE's published slice for the openrouter provider. Point at a file:// URL or
+# another host with APPE_OPENROUTER_URL if APPE moves.
+CATALOGUE_URL = os.environ.get(
+    "APPE_OPENROUTER_URL",
+    "https://appe.dev.gabvdl.xyz/api/models/openrouter.json")
+CACHE_TTL_S = 6 * 3600
+# Survives a restart when APPE is unreachable (same writable volume as the
+# conversation metadata).
+CACHE_PATH = pathlib.Path(
+    os.environ.get("OPENROUTER_CACHE_PATH", "/data/openrouter-models.json"))
+FETCH_TIMEOUT_S = 15
+
+_lock = threading.Lock()
+_cache: tuple[float, list[dict]] | None = None
+_index: dict[str, dict] = {}   # model id → entry, for the pricing lookups
+_generated_at: str = ""
+
+
+def _shape(entry: dict) -> dict | None:
+    """One APPE catalogue row → one pickable model.
+
+    ``api_id`` is the id **without** the ``openrouter/`` prefix — exactly the
+    ``--model`` argument the pi runner passes to OpenRouter. Everything else is
+    renamed to the viewer's camelCase and left otherwise untouched: prices stay
+    USD per million tokens, ``modelSize`` is billions of parameters (null for
+    models that publish none), ``speedTps`` is median output tokens/sec.
+    """
+    mid = entry.get("api_id") or ""
+    if not mid:
+        return None
+    return {
+        "id": mid,
+        "name": entry.get("name") or mid,
+        "vendor": mid.split("/")[0],
+        "description": entry.get("description") or "",
+        "inputCost": float(entry.get("input_cost") or 0.0),
+        "outputCost": float(entry.get("output_cost") or 0.0),
+        # None (not 0) when the model has no prompt cache: a missing cache price
+        # and a free cache are different facts and the picker shows them so.
+        "cacheCost": (float(entry["cache_cost"])
+                      if entry.get("cache_cost") is not None else None),
+        "contextWindow": entry.get("max_token") or None,
+        "modelSize": entry.get("model_size") or None,
+        "tags": entry.get("tags") or [],
+        "tier": entry.get("tier") or "",
+        "license": entry.get("license") or "",
+        "speedTps": entry.get("speed_tps") or None,
+    }
+
+
+def _parse(payload: dict) -> tuple[list[dict], str]:
+    rows = payload.get("models") or []
+    models = [m for m in (_shape(r) for r in rows) if m]
+    # Cheapest first (blended 3:1 in:out, the usual agent mix), so the browser's
+    # unfiltered list opens on what a run costs least to try.
+    models.sort(key=lambda m: (m["inputCost"] * 0.75 + m["outputCost"] * 0.25,
+                               m["id"]))
+    return models, str(payload.get("generatedAt") or "")
+
+
+def _from_disk() -> tuple[list[dict], str]:
+    try:
+        return _parse(json.loads(CACHE_PATH.read_text(encoding="utf-8")))
+    except (OSError, ValueError):
+        return [], ""
+
+
+def _fetch() -> tuple[list[dict], str]:
+    req = urllib.request.Request(
+        CATALOGUE_URL, headers={"accept": "application/json",
+                                "user-agent": "ai-agent/openrouter-catalogue"})
+    try:
+        with urllib.request.urlopen(req, timeout=FETCH_TIMEOUT_S) as r:
+            raw = r.read().decode()
+        payload = json.loads(raw)
+        models, generated = _parse(payload)
+        if models:
+            try:
+                CACHE_PATH.parent.mkdir(parents=True, exist_ok=True)
+                CACHE_PATH.write_text(raw, encoding="utf-8")
+            except OSError:
+                pass  # read-only volume — the in-process cache still serves
+            return models, generated
+    except (urllib.error.URLError, OSError, ValueError):
+        pass
+    return _from_disk()
+
+
+def _store(models: list[dict], generated: str, stamp: float | None = None) -> None:
+    """Publish a catalogue. `stamp=0` marks it as not-yet-refreshed (the disk
+    copy), so the next `catalogue()` still goes and asks APPE."""
+    global _cache, _generated_at, _index
+    with _lock:
+        # A failed refresh keeps whatever we had rather than emptying the picker.
+        if models or not _cache:
+            _cache = (time.time() if stamp is None else stamp, models)
+            _generated_at = generated
+            _index = {m["id"]: m for m in models}
+
+
+def catalogue(force: bool = False) -> list[dict]:
+    """Every OpenRouter model, cheapest first. Empty only if APPE was never reached.
+
+    May block on the network (once per TTL). `warm()` calls it off the request
+    path at startup; the pricing lookups never do.
+    """
+    with _lock:
+        if not force and _cache and time.time() - _cache[0] < CACHE_TTL_S:
+            return _cache[1]
+    _store(*_fetch())  # outside the lock: a slow call blocks nobody
+    return (_cache or (0.0, []))[1]
+
+
+def warm() -> None:
+    """Refresh the catalogue in the background (called at startup)."""
+    threading.Thread(target=catalogue, kwargs={"force": True},
+                     name="openrouter-catalogue", daemon=True).start()
+
+
+def generated_at() -> str:
+    """When APPE generated the catalogue we are serving (ISO 8601, "" if unknown)."""
+    catalogue()
+    return _generated_at
+
+
+def find(model_id: str) -> dict | None:
+    """The catalogue entry for a `--model` argument, or None if it isn't OpenRouter's.
+
+    **Never fetches.** Transcript parsing prices every turn through here, and a
+    parse must not stall on an HTTP call; it reads the in-process index, falling
+    back to the disk copy once. A cold, network-less start therefore prices at
+    the flat estimate rather than hanging.
+    """
+    mid = (model_id or "").strip()
+    if not mid or "/" not in mid:
+        return None  # bare ids are the local EVOX2 provider, not OpenRouter
+    if _cache is None:
+        models, generated = _from_disk()
+        _store(models, generated, stamp=0.0)
+    return _index.get(mid)
+
+
+def rates(model_id: str) -> tuple[float, float, float] | None:
+    """(input, output, cache-read) $ per million tokens for an OpenRouter model.
+
+    None when the id isn't in the catalogue, which is the caller's cue to fall
+    back to its own estimate. A model with no prompt cache reports its cache-read
+    rate as the input rate — a cache read it can't serve is billed as a fresh
+    read, which is what the tokens would actually cost.
+    """
+    m = find(model_id)
+    if not m:
+        return None
+    cache = m["cacheCost"] if m["cacheCost"] is not None else m["inputCost"]
+    return m["inputCost"], m["outputCost"], cache

backend/schemas.py

@@ -1153,6 +1153,37 @@ class ModelsResponse(Schema):
     default: str                     # model a run uses when nothing is picked
 
 
+class OpenRouterModel(Schema):
+    """One OpenRouter model a pi.dev session can be spawned on.
+
+    The catalogue behind it is APPE's (models.dev, synced daily) — see
+    `backend/openrouter.py`. Prices are USD per **million** tokens.
+    """
+    id: str                          # the `--model` argument, e.g. "openai/gpt-oss-20b"
+    name: str                        # display name ("GPT OSS 20B")
+    vendor: str                      # the id's first segment ("openai")
+    description: str = ""
+    inputCost: float
+    outputCost: float
+    # Cache-read price. None = the model has no prompt cache (a cache read is
+    # then billed as a fresh read), which is a different fact from a free cache.
+    cacheCost: Optional[float] = None
+    contextWindow: Optional[int] = None
+    # Parameters in billions, mined from the id/name by APPE. None for models
+    # that publish no size (most closed ones).
+    modelSize: Optional[float] = None
+    tags: list[str] = []             # vision | reasoning | tools | opensource | …
+    tier: str = ""                   # small | medium | big (blended-price heuristic)
+    license: str = ""
+    speedTps: Optional[float] = None  # median output tokens/sec
+
+
+class OpenRouterModelsResponse(Schema):
+    models: list[OpenRouterModel]    # cheapest first (blended 3:1 in:out)
+    source: str                      # the APPE URL the catalogue came from
+    generatedAt: str = ""            # when APPE generated it (ISO 8601)
+
+
 class SpawnResult(Schema):
     sessionId: str
     pid: Optional[int] = None

frontend/openapi.json

@@ -1122,6 +1122,25 @@
         }
       }
     },
+    "/api/models/openrouter": {
+      "get": {
+        "summary": "List Openrouter Models",
+        "description": "Every model a pi.dev session can run on OpenRouter, cheapest first.\n\nNot part of `/api/models` on purpose: that endpoint is the composer's short\npickable list (one chip per Claude family + the curated pi entries), while\nthis is a few hundred models the model browser searches through. Prices\n($/Mtok in / out / cache), context window and parameter size all come from\nAPPE's catalogue \u2014 the homelab's model-data source of truth (openrouter.py).",
+        "operationId": "list_openrouter_models_api_models_openrouter_get",
+        "responses": {
+          "200": {
+            "description": "Successful Response",
+            "content": {
+              "application/json": {
+                "schema": {
+                  "$ref": "#/components/schemas/OpenRouterModelsResponse"
+                }
+              }
+            }
+          }
+        }
+      }
+    },
     "/api/spawn": {
       "post": {
         "summary": "Spawn Conversation",
@@ -9461,6 +9480,133 @@
         ],
         "title": "OkResponse"
       },
+      "OpenRouterModel": {
+        "properties": {
+          "id": {
+            "type": "string",
+            "title": "Id"
+          },
+          "name": {
+            "type": "string",
+            "title": "Name"
+          },
+          "vendor": {
+            "type": "string",
+            "title": "Vendor"
+          },
+          "description": {
+            "type": "string",
+            "title": "Description",
+            "default": ""
+          },
+          "inputCost": {
+            "type": "number",
+            "title": "Inputcost"
+          },
+          "outputCost": {
+            "type": "number",
+            "title": "Outputcost"
+          },
+          "cacheCost": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Cachecost"
+          },
+          "contextWindow": {
+            "anyOf": [
+              {
+                "type": "integer"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Contextwindow"
+          },
+          "modelSize": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Modelsize"
+          },
+          "tags": {
+            "items": {
+              "type": "string"
+            },
+            "type": "array",
+            "title": "Tags",
+            "default": []
+          },
+          "tier": {
+            "type": "string",
+            "title": "Tier",
+            "default": ""
+          },
+          "license": {
+            "type": "string",
+            "title": "License",
+            "default": ""
+          },
+          "speedTps": {
+            "anyOf": [
+              {
+                "type": "number"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "title": "Speedtps"
+          }
+        },
+        "type": "object",
+        "required": [
+          "id",
+          "name",
+          "vendor",
+          "inputCost",
+          "outputCost"
+        ],
+        "title": "OpenRouterModel",
+        "description": "One OpenRouter model a pi.dev session can be spawned on.\n\nThe catalogue behind it is APPE's (models.dev, synced daily) \u2014 see\n`backend/openrouter.py`. Prices are USD per **million** tokens."
+      },
+      "OpenRouterModelsResponse": {
+        "properties": {
+          "models": {
+            "items": {
+              "$ref": "#/components/schemas/OpenRouterModel"
+            },
+            "type": "array",
+            "title": "Models"
+          },
+          "source": {
+            "type": "string",
+            "title": "Source"
+          },
+          "generatedAt": {
+            "type": "string",
+            "title": "Generatedat",
+            "default": ""
+          }
+        },
+        "type": "object",
+        "required": [
+          "models",
+          "source"
+        ],
+        "title": "OpenRouterModelsResponse"
+      },
       "ParentConvRef": {
         "properties": {
           "id": {

frontend/src/api.ts

@@ -21,6 +21,7 @@ import type {
   MemoryDetail,
   MemorySummary,
   ModelsResponse,
+  OpenRouterModelsResponse,
   NotifyLogEntry,
   NotifyResult,
   EnvSaveBody,
@@ -112,6 +113,16 @@ export async function fetchModels(): Promise<ModelsResponse> {
   return r.json();
 }
 
+/** Every OpenRouter model a pi.dev session can run on, with its prices — the
+ *  model browser's catalogue (sourced from APPE; see backend/openrouter.py). */
+export async function fetchOpenRouterModels(): Promise<OpenRouterModelsResponse> {
+  const r = await apiFetch("/api/models/openrouter", {
+    headers: { Accept: "application/json" },
+  });
+  if (!r.ok) throw new Error(`openrouter models: ${r.status}`);
+  return r.json();
+}
+
 export interface ConversationsPage {
   conversations: ConversationSummary[];
   /** Total conversations matching the filter (pre-pagination). */

frontend/src/business/composer/components/OpenRouterBrowser.tsx

@@ -0,0 +1,192 @@
+import { useMemo, useState } from "react";
+import { EmptyState, FuzzyList, Modal, TagFilter, type FuzzyRenderContext } from "@gabvdl/ui";
+import { Cpu, Pi } from "lucide-react";
+import { useOpenRouterModels } from "@/lib/models";
+import { cn } from "@/lib/utils";
+import type { OpenRouterModel } from "@/types";
+
+/**
+ * The OpenRouter model browser — the pi.dev harness's real model picker.
+ *
+ * The composer's select lists a handful of curated entries; OpenRouter serves
+ * several hundred, and choosing between them is a **price** decision more than a
+ * name decision. So this is a fuzzy list where every row states the three rates
+ * that actually bill an agent run (input / output / cache-read, $ per million
+ * tokens), the model's parameter count and its context window — the numbers you
+ * would otherwise go to OpenRouter's site to compare.
+ *
+ * The catalogue comes from APPE (`/api/models/openrouter` → `backend/openrouter.py`
+ * → appe.dev.gabvdl.xyz), which syncs models.dev daily. Cheapest first, so an
+ * unfiltered open lands on what is cheapest to try; typing re-ranks by relevance.
+ */
+
+/** `$/Mtok` as a compact figure: sub-cent rates keep their significant digits. */
+export function usd(n: number | null | undefined): string {
+  if (n == null) return "—";
+  if (n === 0) return "free";
+  if (n < 0.01) return `$${n.toFixed(4).replace(/0+$/, "")}`;
+  if (n < 1) return `$${n.toFixed(3).replace(/0+$/, "")}`;
+  return `$${n.toFixed(2)}`;
+}
+
+/** "35B" / "1.5B" — a model's parameter count, or null when it publishes none. */
+export function paramsLabel(size: number | null | undefined): string | null {
+  if (!size) return null;
+  return size >= 1 ? `${Number(size.toFixed(1))}B` : `${Math.round(size * 1000)}M`;
+}
+
+/** "262k" / "1M" — a context window in tokens. */
+export function ctxLabel(ctx: number | null | undefined): string | null {
+  if (!ctx) return null;
+  if (ctx >= 1_000_000) return `${Number((ctx / 1_000_000).toFixed(1))}M`;
+  return ctx >= 1000 ? `${Math.round(ctx / 1000)}k` : `${ctx}`;
+}
+
+/** The capability facets worth filtering on (a row can carry several). */
+const FACETS = [
+  { value: "tools", label: "Tools" },
+  { value: "reasoning", label: "Reasoning" },
+  { value: "vision", label: "Vision" },
+  { value: "opensource", label: "Open weights" },
+  { value: "free", label: "Free" },
+] as const;
+
+/** One price cell — label above, figure below, so three fit in a row's tail. */
+function Rate({ label, value, muted }: { label: string; value: string; muted?: boolean }) {
+  return (
+    <div className="flex min-w-[3.5rem] flex-col items-end leading-tight">
+      <span className="text-[9px] uppercase tracking-wide text-muted-foreground">{label}</span>
+      <span className={cn("font-mono text-[11px]", muted ? "text-muted-foreground" : "text-foreground")}>
+        {value}
+      </span>
+    </div>
+  );
+}
+
+function Chip({ children }: { children: React.ReactNode }) {
+  return (
+    <span className="rounded-full bg-muted px-1.5 py-0.5 text-[10px] text-muted-foreground">
+      {children}
+    </span>
+  );
+}
+
+export function OpenRouterBrowser({
+  open,
+  picked,
+  onPick,
+  onClose,
+}: {
+  open: boolean;
+  /** The `--model` argument currently in the composer, highlighted in the list. */
+  picked?: string | null;
+  onPick: (id: string) => void;
+  onClose: () => void;
+}) {
+  const { data, isLoading, error } = useOpenRouterModels(open);
+  const [facets, setFacets] = useState<string[]>([]);
+
+  const rows = useMemo(() => {
+    const all = data?.models ?? [];
+    if (!facets.length) return all;
+    return all.filter((m) =>
+      facets.every((f) =>
+        f === "free" ? m.inputCost === 0 && m.outputCost === 0 : (m.tags ?? []).includes(f),
+      ),
+    );
+  }, [data, facets]);
+
+  const renderRow = ({ item, active, highlight }: FuzzyRenderContext<OpenRouterModel>) => {
+    const params = paramsLabel(item.modelSize);
+    const ctx = ctxLabel(item.contextWindow);
+    const isPicked = picked === item.id;
+    return (
+      <div
+        className={cn(
+          "flex cursor-pointer items-center gap-3 rounded-lg border px-3 py-2 transition-colors",
+          active ? "border-primary/50 bg-primary/5" : "border-transparent hover:bg-muted/50",
+          isPicked && "border-teal-500/50 bg-teal-500/5",
+        )}
+      >
+        <div className="min-w-0 flex-1">
+          <div className="flex items-center gap-2">
+            <span className="truncate text-sm font-medium">{highlight("name")}</span>
+            {params && <Chip>{params}</Chip>}
+            {ctx && <Chip>{ctx} ctx</Chip>}
+          </div>
+          <div className="truncate font-mono text-[11px] text-muted-foreground">
+            {highlight("id")}
+          </div>
+        </div>
+        <div className="flex shrink-0 items-center gap-2">
+          <Rate label="in" value={usd(item.inputCost)} />
+          <Rate label="out" value={usd(item.outputCost)} />
+          {/* No cache price = no prompt cache: a "cache read" is billed as a
+              fresh read, which is why it shows the input rate, greyed. */}
+          <Rate
+            label="cache"
+            value={usd(item.cacheCost ?? item.inputCost)}
+            muted={item.cacheCost == null}
+          />
+        </div>
+      </div>
+    );
+  };
+
+  return (
+    <Modal
+      open={open}
+      onClose={onClose}
+      title="OpenRouter models"
+      description="Prices are USD per million tokens, from APPE's daily models.dev sync."
+      size="lg"
+      bodyClassName="flex min-h-0 flex-1 flex-col p-3"
+      className="h-[80dvh] max-h-[80dvh]"
+    >
+      <TagFilter
+        wrap
+        multiple
+        allLabel="All"
+        className="mb-3 shrink-0"
+        items={FACETS.map((f) => ({ id: f.value, label: f.label }))}
+        value={facets}
+        onChange={setFacets}
+      />
+      {isLoading ? (
+        <div className="p-4 text-sm text-muted-foreground">Loading the catalogue…</div>
+      ) : error || !data?.models.length ? (
+        <EmptyState
+          compact
+          icon={<Cpu className="h-5 w-5" />}
+          title="No catalogue"
+          description="APPE (appe.dev.gabvdl.xyz) couldn't be reached, so the model list is empty. The curated pi entries still work."
+        />
+      ) : (
+        <FuzzyList<OpenRouterModel>
+          items={rows}
+          keys={["name", "id", "vendor", "description"]}
+          getItemKey={(m) => m.id}
+          smooth
+          onSelect={(m) => {
+            onPick(m.id);
+            onClose();
+          }}
+          placeholder={'Search models…  (use "quotes" for exact)'}
+          estimateSize={62}
+          overscan={8}
+          showCount
+          emptyState={
+            <EmptyState
+              compact
+              icon={<Pi className="h-5 w-5" />}
+              title="No model matches"
+              description="Try another search, or clear the capability filters."
+            />
+          }
+          className="min-h-0 flex-1"
+          renderItem={renderRow}
+        />
+      )}
+    </Modal>
+  );
+}

frontend/src/generated/api.ts

@@ -82,6 +82,7 @@ import type {
   NotifyLogResponse,
   NotifyResult,
   OkResponse,
+  OpenRouterModelsResponse,
   PlanDetail,
   PlanDetailApiPlanGetParams,
   PlansResponse,
@@ -1733,6 +1734,56 @@ export const listModelsApiModelsGet = async ( options?: RequestInit): Promise<li
 
 
 
+export type listOpenrouterModelsApiModelsOpenrouterGetResponse200 = {
+  data: OpenRouterModelsResponse
+  status: 200
+}
+
+export type listOpenrouterModelsApiModelsOpenrouterGetResponseSuccess = (listOpenrouterModelsApiModelsOpenrouterGetResponse200) & {
+  headers: Headers;
+};
+;
+
+export type listOpenrouterModelsApiModelsOpenrouterGetResponse = (listOpenrouterModelsApiModelsOpenrouterGetResponseSuccess)
+
+export const getListOpenrouterModelsApiModelsOpenrouterGetUrl = () => {
+
+
+
+
+  return `/api/models/openrouter`
+}
+
+/**
+ * Every model a pi.dev session can run on OpenRouter, cheapest first.
+ *
+ * Not part of `/api/models` on purpose: that endpoint is the composer's short
+ * pickable list (one chip per Claude family + the curated pi entries), while
+ * this is a few hundred models the model browser searches through. Prices
+ * ($/Mtok in / out / cache), context window and parameter size all come from
+ * APPE's catalogue — the homelab's model-data source of truth (openrouter.py).
+ * @summary List Openrouter Models
+ */
+export const listOpenrouterModelsApiModelsOpenrouterGet = async ( options?: RequestInit): Promise<listOpenrouterModelsApiModelsOpenrouterGetResponse> => {
+
+  const res = await fetch(getListOpenrouterModelsApiModelsOpenrouterGetUrl(),
+  {
+    ...options,
+    method: 'GET'
+
+
+  }
+)
+
+
+  const body = [204, 205, 304].includes(res.status) ? null : await res.text();
+
+  const data: listOpenrouterModelsApiModelsOpenrouterGetResponse['data'] = body ? JSON.parse(body) : {}
+  return { data, status: res.status, headers: res.headers } as listOpenrouterModelsApiModelsOpenrouterGetResponse
+}
+
+
+
 export type spawnConversationApiSpawnPostResponse200 = {
   data: SpawnResult
   status: 200

frontend/src/generated/model/index.ts

@@ -154,6 +154,8 @@ export * from './notifyResult.ts';
 export * from './ogImage.ts';
 export * from './ogImageKind.ts';
 export * from './okResponse.ts';
+export * from './openRouterModel.ts';
+export * from './openRouterModelsResponse.ts';
 export * from './parentConvRef.ts';
 export * from './planActual.ts';
 export * from './planConvRef.ts';

frontend/src/generated/model/openRouterModel.ts

@@ -0,0 +1,28 @@
+/**
+ * Generated by orval v8.20.0 🍺
+ * Do not edit manually.
+ * ai-agent
+ * OpenAPI spec version: 0.1.0
+ */
+
+/**
+ * One OpenRouter model a pi.dev session can be spawned on.
+ *
+ * The catalogue behind it is APPE's (models.dev, synced daily) — see
+ * `backend/openrouter.py`. Prices are USD per **million** tokens.
+ */
+export interface OpenRouterModel {
+  id: string;
+  name: string;
+  vendor: string;
+  description?: string;
+  inputCost: number;
+  outputCost: number;
+  cacheCost?: number | null;
+  contextWindow?: number | null;
+  modelSize?: number | null;
+  tags?: string[];
+  tier?: string;
+  license?: string;
+  speedTps?: number | null;
+}

frontend/src/generated/model/openRouterModelsResponse.ts

@@ -0,0 +1,13 @@
+/**
+ * Generated by orval v8.20.0 🍺
+ * Do not edit manually.
+ * ai-agent
+ * OpenAPI spec version: 0.1.0
+ */
+import type { OpenRouterModel } from './openRouterModel.ts';
+
+export interface OpenRouterModelsResponse {
+  models: OpenRouterModel[];
+  source: string;
+  generatedAt?: string;
+}

frontend/src/lib/costAnalysis.ts

@@ -34,8 +34,10 @@ const DAY_MS = 86_400_000;
 const MONTH_MS = 30.4375 * DAY_MS;
 
 // ── pricing ───────────────────────────────────────────────────────────────────
-// Mirrors the backend's `conversations.PRICES` (USD per million tokens, input /
-// output). Cache reads bill at 0.1× input; cache writes are priced off their
+// Mirrors the backend's `conversations.PRICES` **fallback** table (USD per
+// million tokens, input / output) — the backend additionally prices OpenRouter
+// models from APPE's catalogue, per model. Cache reads bill at 0.1× input (an
+// OpenRouter model's own cache rate can differ); cache writes are priced off their
 // weighted `cacheWriteUnits` (5m ≈ 1.25×, 1h ≈ 2×) at the input rate. Kept
 // client-side so the dashboard can split a conversation's single `cost` figure
 // back into per-token-type shares.
@@ -53,8 +55,15 @@ function priceFor(model: string | null | undefined): [number, number] {
   if (m.includes("sonnet")) return PRICES.sonnet;
   if (m.includes("haiku")) return PRICES.haiku;
   if (m.includes("fable") || m.includes("mythos")) return PRICES.fable;
-  if (m.includes("qwen"))
-    return m.includes("/") ? PRICES.qwen : PRICES["qwen-local"];
+  if (m.includes("/")) {
+    // A vendor-prefixed id is an OpenRouter model on the pi harness. Its *real*
+    // rates come from the catalogue backend-side (`backend/openrouter.py` prices
+    // each turn, and that is what a conversation's `cost` already reflects);
+    // this table only splits that figure across token types, so a single open-
+    // model estimate is enough — and beats charging it at Opus rates.
+    return PRICES.qwen;
+  }
+  if (m.includes("qwen")) return PRICES["qwen-local"];
   return PRICES.opus;
 }
 

frontend/src/lib/models.tsx

@@ -1,3 +1,4 @@
+import { useState } from "react";
 import { useQuery } from "@tanstack/react-query";
 import {
   Bot,
@@ -8,11 +9,13 @@ import {
   Gauge,
   Pi,
   Rabbit,
+  Search,
   Sparkles,
   SquareTerminal,
   Zap,
 } from "lucide-react";
-import { fetchModels } from "@/api";
+import { fetchModels, fetchOpenRouterModels } from "@/api";
+import { OpenRouterBrowser } from "@/business/composer/components/OpenRouterBrowser";
 import { SelectPopover, type SelectOption } from "@/technical/SelectPopover";
 import { cn } from "@/lib/utils";
 import type { ModelInfo } from "@/types";
@@ -41,6 +44,10 @@ import type { ModelInfo } from "@/types";
 /** The agent CLI a session runs on. */
 export type Harness = "claude" | "pi";
 
+/** Sentinel option value: "Browse OpenRouter…" opens the catalogue modal
+ *  instead of picking a model. Can't collide with a real id (no slash). */
+const BROWSE = "__browse_openrouter__";
+
 /** Per-harness label + icon for the harness select. */
 export const HARNESS_STYLE: Record<Harness, { label: string; hint: string; icon: typeof Bot }> = {
   claude: {
@@ -65,14 +72,27 @@ const FAMILY_STYLE: Record<string, { cls: string; icon: typeof Bot }> = {
   qwen: { cls: "bg-sky-500/10 text-sky-500", icon: Cpu },
 };
 const UNKNOWN = { cls: "bg-muted text-muted-foreground", icon: Bot };
+// Any vendor-prefixed id the families above don't claim — i.e. one of the few
+// hundred OpenRouter models the browser can pick (see OpenRouterBrowser).
+const OPENROUTER_STYLE = { cls: "bg-sky-500/10 text-sky-500", icon: Cpu };
 
 /** The family a model id belongs to (`claude-opus-4-8` → `opus`). */
 export function modelFamily(id: string): string {
   return Object.keys(FAMILY_STYLE).find((f) => id.includes(f)) ?? "";
 }
 
+/** Whether a `--model` argument is an OpenRouter model (`vendor/model`).
+ *  Claude ids never carry a slash, so the shape alone is the test — which is
+ *  what lets a conversation resumed on an arbitrary OpenRouter model still
+ *  resolve to the pi harness without the catalogue being loaded. */
+export function isOpenRouterId(id: string | null | undefined): boolean {
+  return !!id && id.includes("/");
+}
+
 export function modelStyle(id: string) {
-  return FAMILY_STYLE[modelFamily(id)] ?? UNKNOWN;
+  const family = modelFamily(id);
+  if (family) return FAMILY_STYLE[family];
+  return isOpenRouterId(id) ? OPENROUTER_STYLE : UNKNOWN;
 }
 
 /**
@@ -88,7 +108,9 @@ export function modelLabel(id: string): string {
     const ver = tail.match(/qwen(\d+(?:\.\d+)?)/)?.[1];
     return ver ? `qwen ${ver}` : tail;
   }
-  if (!family) return id.replace("claude-", "");
+  // An OpenRouter pick: the vendor prefix is the least informative half of
+  // `openai/gpt-oss-20b`, and the chip is narrow — show the model, not the shop.
+  if (!family) return isOpenRouterId(id) ? (id.split("/").pop() as string) : id.replace("claude-", "");
   const version = id
     .split(`${family}-`)[1]
     ?.split("-")
@@ -126,6 +148,25 @@ export function useModels() {
   });
 }
 
+/**
+ * The OpenRouter catalogue — every model the **pi** harness can run, with its
+ * $/Mtok (input / output / cache-read), parameter size and context window.
+ *
+ * Deliberately a second query rather than more entries in `useModels()`: that
+ * one is the composer's short pickable list (a chip each), this is a few hundred
+ * rows only the browser modal renders, so it is fetched when the modal first
+ * opens and then cached for the app session. The data itself is APPE's — see
+ * `backend/openrouter.py`.
+ */
+export function useOpenRouterModels(enabled = true) {
+  return useQuery({
+    queryKey: ["models", "openrouter"],
+    queryFn: fetchOpenRouterModels,
+    staleTime: 60 * 60 * 1000,
+    enabled,
+  });
+}
+
 /**
  * The `--model` argument for a model entry: a family's newest model goes out as
  * its CLI alias (`opus`), older ones as their full id.
@@ -156,7 +197,9 @@ export function harnessForModel(
   if (!arg) return "claude";
   const hit = findModel(models, arg);
   if (hit?.harness) return hit.harness === "pi" ? "pi" : "claude";
-  return modelFamily(arg) === "qwen" ? "pi" : "claude";
+  // Not in the list: an older transcript's model, or one of the several hundred
+  // OpenRouter models the browser can pick (which the short list never carries).
+  return isOpenRouterId(arg) || modelFamily(arg) === "qwen" ? "pi" : "claude";
 }
 
 /** The models a harness can run, in backend order (newest of each family first). */
@@ -242,6 +285,7 @@ export function ModelSelect({
   onChange: (arg: string) => void;
   disabled?: boolean;
 }) {
+  const [browsing, setBrowsing] = useState(false);
   const list = modelsForHarness(models, harness);
   const options: SelectOption[] = list.map((m) => {
     const { icon: Icon, cls } = modelStyle(m.id);
@@ -256,32 +300,65 @@ export function ModelSelect({
   // What the chip shows: the pick, else the harness's first model (what the run
   // will actually use) — never a blank chip.
   const shown = model || (list[0] ? modelArg(list[0]) : "");
-  if (!shown) {
+  // A model picked from the OpenRouter browser isn't in the short list, so it
+  // gets its own option — otherwise the popover would show no pick at all.
+  if (shown && isOpenRouterId(shown) && !options.some((o) => o.value === shown)) {
+    const { icon: Icon, cls } = modelStyle(shown);
+    options.unshift({
+      value: shown,
+      label: modelLabel(shown),
+      hint: shown,
+      icon: <Icon className={cn("h-3.5 w-3.5", cls.split(" ")[1])} />,
+      meta: "OpenRouter",
+    });
+  }
+  // The pi harness's real catalogue is the browser, not this list: several
+  // hundred models, picked by price. The select keeps the curated entries and
+  // hands off to the modal for everything else.
+  if (harness === "pi") {
+    options.push({
+      value: BROWSE,
+      label: "Browse OpenRouter…",
+      hint: "every model, with its $/Mtok and size",
+      icon: <Search className="h-3.5 w-3.5 text-muted-foreground" />,
+    });
+  }
+  if (!shown && harness !== "pi") {
     return <span className="text-[10px] text-muted-foreground">default model</span>;
   }
   const { icon: Icon, cls } = modelStyle(shown);
   const entry = findModel(models, shown);
   return (
-    <SelectPopover
-      title="Model"
-      width={250}
-      options={options}
-      value={shown}
-      disabled={disabled}
-      onChange={onChange}
-      trigger={
-        <span
-          title={`Runs on ${entry?.id ?? shown} — click to pick another model`}
-          className={cn(
-            "inline-flex items-center gap-1 rounded-full px-1.5 py-0.5 text-[10px] font-medium",
-            cls,
-          )}
-        >
-          <Icon className="h-2.5 w-2.5" />
-          {entry?.label || modelLabel(shown)}
-        </span>
-      }
-    />
+    <>
+      <SelectPopover
+        title="Model"
+        width={250}
+        options={options}
+        value={shown}
+        disabled={disabled}
+        onChange={(v) => (v === BROWSE ? setBrowsing(true) : onChange(v))}
+        trigger={
+          <span
+            title={`Runs on ${entry?.id ?? shown} — click to pick another model`}
+            className={cn(
+              "inline-flex items-center gap-1 rounded-full px-1.5 py-0.5 text-[10px] font-medium",
+              cls,
+            )}
+          >
+            <Icon className="h-2.5 w-2.5" />
+            {entry?.label || modelLabel(shown) || "pick a model"}
+          </span>
+        }
+      />
+      {browsing && (
+        <OpenRouterBrowser
+          open
+          picked={shown}
+          onPick={onChange}
+          onClose={() => setBrowsing(false)}
+        />
+      )}
+    </>
   );
 }
 

frontend/src/mock/db.ts

@@ -16,6 +16,7 @@ import type {
   MemoryDetail,
   MemorySummary,
   ModelsResponse,
+  OpenRouterModelsResponse,
   NotifyLogEntry,
   PlanDetail,
   PlanSummary,
@@ -42,6 +43,8 @@ export interface MockDb {
   workspace: string;
   pricing: Pricing;
   models: ModelsResponse;
+  /** The OpenRouter catalogue the pi-harness model browser searches. */
+  openRouterModels: OpenRouterModelsResponse;
   files: FileEntry[];
   conversations: ConversationDetail[];
   /** Per-conversation commit rollup + net diff, keyed by conversation id. */

frontend/src/mock/handlers.ts

@@ -302,6 +302,7 @@ export async function handle(url: URL, init: RequestInit | undefined, request?:
   if (p === "/api/deploy-status") return json(db.deployStatus);
   if (p === "/api/bundle") return json({ workspace: db.workspace, totals: totals(), files: db.files, pricing: db.pricing });
   if (p === "/api/models") return json(db.models);
+  if (p === "/api/models/openrouter") return json(db.openRouterModels);
   if (p === "/api/dashboard") {
     const t = totals();
     return json({

frontend/src/mock/seed.ts

@@ -1425,6 +1425,20 @@ export function seedDb(): MockDb {
         { id: "qwen3-vl-8b", displayName: "Qwen3 VL 8B", family: "qwen", version: "3", alias: null, latest: false, harness: "pi", provider: "lmstudio", label: "Qwen VL", contextWindow: 32_000 },
       ],
     },
+    // A handful of the OpenRouter catalogue (the real one is ~340 models from
+    // APPE) — enough for the model browser to search, filter and price in mock
+    // mode. Cheapest first, like the API returns it.
+    openRouterModels: {
+      source: "https://appe.dev.gabvdl.xyz/api/models/openrouter.json",
+      generatedAt: "2026-08-01T04:30:00.000Z",
+      models: [
+        { id: "google/gemma-4-31b-it:free", name: "Gemma 4 31B (free)", vendor: "google", description: "Open Gemma instruction model for self-hosted chat and reasoning", inputCost: 0, outputCost: 0, cacheCost: null, contextWindow: 262_144, modelSize: 31, tags: ["vision", "reasoning", "tools", "opensource"], tier: "small", license: "opensource", speedTps: 120 },
+        { id: "mistralai/mistral-nemo", name: "Mistral Nemo", vendor: "mistralai", description: "Small multilingual instruct model", inputCost: 0.019, outputCost: 0.03, cacheCost: null, contextWindow: 131_072, modelSize: null, tags: ["tools", "opensource"], tier: "small", license: "opensource", speedTps: 190 },
+        { id: "openai/gpt-oss-20b", name: "GPT OSS 20B", vendor: "openai", description: "Open-weight GPT model for self-hosted reasoning", inputCost: 0.03, outputCost: 0.13, cacheCost: 0.03, contextWindow: 131_072, modelSize: 20, tags: ["reasoning", "tools", "opensource"], tier: "small", license: "opensource", speedTps: 258 },
+        { id: "qwen/qwen3-235b-a22b", name: "Qwen3 235B A22B", vendor: "qwen", description: "Large MoE reasoning model", inputCost: 0.13, outputCost: 0.6, cacheCost: 0.02, contextWindow: 262_144, modelSize: 235, tags: ["reasoning", "tools", "opensource"], tier: "small", license: "opensource", speedTps: 88 },
+        { id: "anthropic/claude-sonnet-5", name: "Claude Sonnet 5", vendor: "anthropic", description: "Balanced frontier model", inputCost: 3, outputCost: 15, cacheCost: 0.3, contextWindow: 1_000_000, modelSize: null, tags: ["vision", "reasoning", "tools"], tier: "medium", license: "commercial", speedTps: 71 },
+      ],
+    },
     files,
     conversations,
     commits,

runner/README.md

@@ -48,11 +48,16 @@ composer chip → POST /api/spawn {model, harness}
 
 | model id | provider |
 |---|---|
-| `qwen/qwen3.6-35b-a3b` (vendor prefix) | OpenRouter (`OPENROUTER_API_KEY`, auto-loaded from `~/homelab/.env.claude`) |
+| `openai/gpt-oss-20b`, `qwen/qwen3.6-35b-a3b`, … (vendor prefix) | OpenRouter (`OPENROUTER_API_KEY`, auto-loaded from `~/homelab/.env.claude`) |
 | `qwen3.6-35b-a3b` (bare) | EVOX2 LM Studio (`~/.pi/agent/models.json` `evox2` provider; box is WoL-woken automatically) |
 
-The pickable list lives in `backend/models.py` (`PI_MODELS_DEFAULT`, override
-with `PI_MODELS_JSON`).
+The routing is the **id shape**, not a list — so any of the several hundred
+OpenRouter models runs here with no runner change. Two pickers feed it: the
+curated chips in `backend/models.py` (`PI_MODELS_DEFAULT`, override with
+`PI_MODELS_JSON`) and the composer's **OpenRouter browser**, which lists the
+whole catalogue with per-model prices from `backend/openrouter.py` (sourced from
+APPE). Those same prices cost the session's transcript, so a run's `$` figure is
+the model's real rate even before pi reports its own `costUSD`.
 
 ## Permissions