asher · asher · Aug 10, 2026 · Aug 9, 2026 · Aug 9, 2026 · Aug 9, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,25 +8,27 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Added
 
+- gmlx chat --server: chat against a running server as a plain client,
+  without the assistant's tools and memory (no background requests).
+  Engages automatically when the config's server is already up and
+  serves the requested model; --local pins the in-process load. An
+  explicit GGUF path always loads the file on disk.
 - decode_prefill_ratio accepts "auto" and it is the new default: live
   streams keep at least half their decode rate while deep prompts admit,
   and pacing stands down wherever it would not help (simultaneous
   bursts, cheap chunks, stuck queues). A numeric ratio pins the previous
   static behavior; GMLX_DECODE_PREFILL_AUTO=0 reverts on a live server.
-
 - GMLX_SERVE_MEMSTATS=path.jsonl writes a per-tick serve memory trace:
   MLX counters, free-headroom estimate, and per-owner cache byte
   attribution with allocation shapes marked on change, for diagnosing
   serve memory growth under load.
-
 - Serve admission is gated on projected memory headroom: a request whose
   measured KV and prefill-transient projection does not fit is kept
   queued and retried each tick instead of committing memory the box does
   not have. Requests are never failed by the gate, an idle server always
   admits, and a request deferred past GMLX_ADMIT_DEFER_MAX_S (default
   60s) is admitted anyway with a loud log. GMLX_ADMIT_HEADROOM=0
   disables.
-
 - /v1/metrics reports residency budget vs resident bytes, live
   active/cache/headroom memory, and admission deferral counters.
 
@@ -35,19 +37,15 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 - mlx-kquant floor raised to 0.3.11: MoE prefill gather runs 12-28%
   faster per call at chat-chunk widths, lifting serve prefill 20-43%
   shallow and 8-14% deep on many-expert models.
-
 - Bench chart value axes clamp to the data range when a zero anchor
   would waste the panel height on empty space; nearby engine lines
   now read as visually distinct.
-
 - benchmarks.md tracks builds and measured date per model, and merged
   results carry the newest contributing run date: one model rebenched
   on newer releases no longer implies the rest was remeasured.
-
 - DeepSeek-V4-Flash IQ2_XXS rebenched on gmlx 0.2.2 + mlx-kquant
   0.3.11 vs ds4-server b030961 (2026-08-05): prefill 1.11-1.86x and
   decode 1.05-1.59x across the full d512-500k ladder.
-
 - DeepSeek-V4 single-token decode runs its hyper-connection glue as four
   native mlx-kquant ops instead of about 176 python kernel launches per
   step. `GMLX_HC_M1_FUSED=0` and `GMLX_HC_KQ=0` restore the previous
@@ -65,6 +63,8 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Fixed
 
+- Chat's bottom toolbar no longer clips the live tok/s readout on
+  narrow terminals; sampling knobs are dropped first instead.
 - The serve free-headroom estimate went negative on models whose load
   materializes weights into MLX-tracked memory (the same bytes counted
   twice); the loader now registers only the truly untracked mmap

diff --git a/docs/cli.md b/docs/cli.md
@@ -275,6 +275,7 @@ decode tok/s, MTP acceptance when speculating, and context fill
 ```sh
 gmlx chat model.gguf --temp 0.7 --system-prompt "You are terse."
 gmlx chat --assistant                 # the tool-loop assistant on the managed server
+gmlx chat --server                    # plain server client, no assistant extras
 ```
 
 With `--assistant` the REPL loads nothing locally: turns run through the
@@ -285,6 +286,18 @@ positional becomes a served model id (or is omitted for the server default).
 unchanged; local-load flags do not apply. Full contract:
 [assistant.md](assistant.md#text-chat-gmlx-chat---assistant).
 
+`--server` is the same server-backed REPL minus the assistant extras: no
+tools, no memory store, no config `assistant:` block, just plain streamed
+turns against the served model. Use it when the terminal should be a thin
+client and every server request should come from the conversation itself.
+
+When the config's server is already running, a bare `gmlx chat` (or one
+naming a served model id) becomes a `--server` client automatically
+instead of loading a second copy in-process; `--local` forces the local
+load, and any local-load flag does the same. An explicit GGUF path always
+loads the file on disk - if the running server serves that same file, a
+note points at the served id. Chat never auto-starts a server this way.
+
 `/exit` (or Ctrl-D) quits, `/reset` restarts the conversation, `/help` lists
 every command. The terminal is upgraded on top:
 
@@ -399,10 +412,12 @@ every command. The terminal is upgraded on top:
 
 | Flag | Default | Meaning |
 |------|---------|---------|
-| `gguf` (positional) | - | Path to the GGUF (sharded ok) or a config model id; with `--assistant`, a served model id (optional: server default). |
+| `gguf` (positional) | - | Path to the GGUF (sharded ok) or a config model id; with `--assistant`/`--server`, a served model id (optional: server default). |
 | `--assistant` | - | Chat through the built-in tool-loop assistant on the managed server: MCP tools + long-term memory from the `assistant:` block ([assistant.md](assistant.md)). Local-load flags don't apply. |
-| `--base-url URL` / `--host` / `--port` / `--api-key` | managed server | Assistant mode: target server (as in [`talk`](#gmlx-talk)). |
-| `--no-start` / `--start-timeout S` | - / `180` | Assistant mode: never auto-start the server / auto-start wait. |
+| `--server` | - | Plain server client: the `--assistant` REPL minus tools, memory, and the config `assistant:` block. Automatic when the config's server is up and serves the request. |
+| `--local` | - | Load in-process even when the config's server is running (skips the automatic `--server` mode). |
+| `--base-url URL` / `--host` / `--port` / `--api-key` | managed server | Server modes: target server (as in [`talk`](#gmlx-talk)). |
+| `--no-start` / `--start-timeout S` | - / `180` | Server modes: never auto-start the server / auto-start wait. |
 | `--max-tokens N` | until EOS | Per-reply decode-token cap; default `0` = each reply runs until the model stops (diffusion models fall back to a bounded 2048-token canvas; in `--assistant` mode `0` defers to the server's own default). Pass N to cap (adjustable via `/max-tokens`; `0` removes the cap, and a note says when the cap ended a reply). |
 | `--temp` / `--top-p` / `--top-k` / `--min-p` | family default | Sampling; unset flags seed from the model's [family defaults](#family-defaults-intent-and---profile) (`0.0`/`0.95`/`0`/`0.05` under `--no-family-defaults`). All adjustable in-chat. |
 | `--xtc-probability` / `--xtc-threshold` | `0.0` | XTC sampling (text path, adjustable in-chat). |