Codeman/docs/local-models.md

A model that runs on this Mac

Codeman can set up a coding model that runs on the Mac the server runs on: no account and no API key, and the session's prompts and files go to a model on this Mac rather than a cloud provider (the coding agent itself may still check for updates). One button on the welcome screen opens a window that shows what fits this Mac, installs the choice and ends on a coding agent session that runs on it.

Experimental. The routes and their payloads may change in any release (versioning-policy.md, "Experimental local model routes"). Apple silicon macOS only: anywhere else the window says why nothing is offered.

This page is the reference for how it works. The HTTP contract is in api-reference.md ("Local model"). The rules that must not be broken are in architecture-invariants.md.

Which model fits which Mac

The window offers the best model of the catalogue (src/config/local-model-catalog.ts) that fits this Mac's memory at a context a coding agent can work with, and up to four more. What the catalogue answers today, before the runtime measured anything (the budget is then an estimate: 2/3 of the memory below 36 GB, 3/4 from 36 GB, less 1 GiB, and never more than the memory less 6, 8 or 12 GiB for macOS, Codeman, a browser and the agent):

Memory	Budget	What it gets
8 GB	2.1 GB	Nothing. The smallest model a coding agent can work with needs 10.2 GB at 65,536 tokens
16 GB	10.4 GB	Nothing. MiMo V2.6 9B at 65,536 tokens needs 10.2 GB, 98.7% of the budget, just past the 98% a tight fit may reach (it is 95% of the 10.7 GB a 16 GB Mac mini measured). On the edge either way: only a run on such a Mac settles it
18 GB	11.8 GB	A small model (limited): MiMo V2.6 9B at 131,072 tokens, 11.4 GB, tight. Its base, Qwen3.5 9B, is the one alternative
24 GB	16.1 GB	The same, with room. On a base M2 or M3 (100 GB/s) it computes to 9.3 tokens per second: slow
32 GB	21.8 GB	The same. North Mini Code 1.0 needs 23.2 GB at 65,536 tokens; once the runtime measures about 3/4 of the memory, as it did on a 16 and a 48 GB Mac, the Mac gets it at 131,072 tokens, or Qwen3.8 27B at 65,536 tokens on a Max chip
36 GB	27.9 GB	Qwen3.8 27B at 131,072 tokens on an M4 or M5 Max (27.0 GB, 97%); on a slower chip, where the dense 27B computes to under 10 tokens per second, Qwen3.6 35B-A3B as a GGUF build (27.1 GB, 97%). Both tight
48 GB	37.6 GB	Qwen3.6 35B-A3B (MLX) at 131,072 tokens: the pair we tested, "Recommended" on an Apple M5 Pro with 48 GB. On a Max chip Qwen3.8 27B at 262,144 tokens (32.1 GB, tight). The test was a short task: in a long session MLX may keep up to 8 GiB more, which brings this pair to the edge of the budget, and past it while a long prompt is processed
64 GB	50.5 GB	Qwen3.8 27B at 262,144 tokens on a Max chip; Qwen3.6 35B-A3B (MLX) on a Pro chip
96 GB	76.2 GB	Qwen3.8 27B at 262,144 tokens; on an M3 Max, Qwen3.6 35B-A3B (MLX)
128 GB	102 GB	Ornith 1.5 122B-A10B (MLX) at 131,072 tokens, 97.4 GB, tight
192, 256, 512 GB	154 GB+	Ornith 1.5 122B-A10B at 262,144 tokens

"Tight" means the need is between 85% and 98% of the budget; above 98% a model is not offered. A model that computes to under 10 tokens per second on the chip is "slow" and ranks after the fast ones, which is why the chip changes the pick at 36 to 96 GB. Every model is offered at the largest context it fits at, and never below 65,536 tokens: OpenCode's own prompt is 8,000 to 10,000 tokens, and at 32,768 it would compact after less than one default file read. test/local-model-recommend.test.ts ("every memory size") pins this table.

How much memory a model needs

The need of a model is its peak: the weights, what stays held per token of context, what the runner holds beside both, and what it takes for a moment while a long prompt is processed. The runtime runs a model with one of two runners, which spend memory differently, so every row names its runner (runner) and its figures follow that runner's rule:

• llama.cpp (GGUF builds, on Metal) allocates the K/V cache for the whole context when the model loads and does not grow after that. The runtime is started with OLLAMA_KV_CACHE_TYPE=q8_0 (the cache in 8 bits, 17/32 of the default f16; the runtime's own documentation calls the loss very small), OLLAMA_FLASH_ATTENTION=1 (an 8-bit cache needs it), LLAMA_ARG_CACHE_RAM=1024 (llama-server keeps copies of earlier prompts, and this bounds them to 1 GiB) and LLAMA_ARG_CTX_CHECKPOINTS=8 (llama-server keeps copies of the state it cannot roll back, two or three per request, up to 32 by default: 5 GB for the 27B; 8 keeps the last few requests, which is what an agent's next request restores from). A row's figures are computed from its model's config.json: the K/V cache of its attention layers, its recurrent or sliding-window state, live and in each of the 8 checkpoints, 1.0 GB for the runner and the 1 GiB prompt cache. A build with a multi-token-prediction layer (the two -mtp tags and qwen3.8:27b-q8_0) costs more: the runtime turns prediction on, so llama-server keeps that layer's cache in f16 and one more copy of the recurrent state for rollback per drafted token (the tag's draft_num_predict: 2 for the 35B-A3B, 4 for the 27B). No peak: the buffers are allocated at load. The checkpoints and the prediction costs were read in the llama.cpp source the runtime bundles (b12111), not measured.
• MLX grows its cache with the tokens actually used, keeps snapshots of the recurrent state, and may keep up to 8 GiB of snapshots of earlier prompts, which no setting bounds. A Mac always takes the MLX build of a tag that has one. Below 48 GB that 8 GiB does not fit beside a useful model, so the catalogue lists GGUF builds there.

Where every figure comes from:

Model	Runner	Memory figures	Speed (tokens per second per 100 GB/s)
qwen3.5:9b-q4_k_m	llama.cpp	computed from the Qwen3.5-9B config	9.3: a published run on an M1 Max
mimo-v2.6:9b-q4_k_m	llama.cpp	the same (a fine-tune of Qwen3.5-9B)	9.3: the same architecture and size
north-mini-code-1.0:q4_k_m	llama.cpp	computed from its config	13.6: the 35B-A3B's, the same 3 billion active parameters
qwen3.6:35b-a3b-mtp-q4_k_m	llama.cpp	computed from the Qwen3.6-35B-A3B config, prediction layer included	13.6: a published run on an M1 Max
qwen3.6:35b-a3b	MLX	measured on an Apple M5 Pro with 48 GB	16.9: a published run; 67.7 measured on the tested Mac
qwen3.8:27b-q8_0	llama.cpp	computed from the Qwen3.8-27B config, prediction layer included	1.7: the row below scaled by file size
qwen3.8:27b-mtp-q4_k_m	llama.cpp	computed from the Qwen3.8-27B config, prediction layer included	2.95: a published run without prediction (a floor)
ornith-1.5:122b	MLX	computed from the Qwen3.5-122B-A10B config, scaled by the measured MLX row, + 8 GiB	5.0: the 35B-A3B's, x 3/10 active parameters

Sizes are what the registry serves a Mac for each tag (every layer plus the config), read on 2026-11-08. Only one row was ever run on a Mac, so every other answer in the table above is a calculation. For a llama.cpp build the check after the download catches a load that was judged too kindly, but not what grows after it in host memory (the checkpoints and the prompt cache, which are counted, not checked); for an MLX build it catches neither (see "The check" below). The measured row's figures come from single long prompts, not a session of many requests, in which MLX may keep up to 8 GiB more. The catalogue header lists the models left out on purpose (no evidence of working tool calls, a restrictive licence, a known bug) and why.

A new row needs its model's architecture (attention layers, K/V heads, head size, and its linear-attention or sliding-window layers, from its config.json) in the table of test/local-model-catalog.test.ts, which recomputes every computed figure from it and fails for a row it does not know. Its answer for each memory size is then pinned in test/local-model-recommend.test.ts.

What an install does, in order

The person picks a model, presses "Install on this Mac" and confirms the sizes once. The server then runs one job of five steps. The window polls GET /api/local-model once a second; the job runs in the server, so the window can be closed.

Step	What it does
runtime	Downloads the pinned runtime, compares its hash, unpacks it, chooses a port once, starts it, reads the measured memory budget and judges the choice again
model	Downloads the model through the runtime
verify	Loads the model and checks it: one tiny chat, the memory and context the runtime reports, one tool call
agent	Installs the coding agent with its stock install command. Left out when the agent is installed
connect	Saves one model endpoint and switches the custom model endpoints setting on

Every step can be repeated: Resume and "Try again" enter the step the job stopped in.

The runtime

• One pinned build. Ollama 0.40.15 for macOS, a compiled constant (OLLAMA_RUNTIME_PIN in src/local-model/runtime.ts): the URL, the size (185,971,671 bytes) and the sha256. No file, setting, environment variable or request can change which binary is downloaded.
• Download. Streamed to *.partial while it is hashed. A cut stream is continued with a Range request, at most three times. The hash is compared before anything is unpacked, and a mismatch deletes the file. Unpacking is /usr/bin/tar into a temporary folder, then one rename.
• No administrator rights. Everything lives in the data directory.
• Supervised by Codeman. The runtime is a non-detached child of the server: started at boot once a model is installed and not stopped, restarted at most three times after a crash, sent SIGTERM at shutdown. It is not a login item: it stops with Codeman, and a leftover from a server that was killed is taken back at the next start (see below).
• Its environment is an allowlist. PATH, TMPDIR, LANG and the proxy variables are copied from the server. HOME is a private folder. Exactly these are set: OLLAMA_HOST=127.0.0.1:<port>, OLLAMA_MODELS, OLLAMA_CONTEXT_LENGTH, OLLAMA_NO_CLOUD=1, OLLAMA_KEEP_ALIVE=10m, OLLAMA_NUM_PARALLEL=1, OLLAMA_MAX_LOADED_MODELS=1, and four memory settings of the llama.cpp runner: OLLAMA_KV_CACHE_TYPE=q8_0, OLLAMA_FLASH_ATTENTION=1, LLAMA_ARG_CACHE_RAM=1024 and LLAMA_ARG_CTX_CHECKPOINTS=8 (see "How much memory a model needs" above). No CODEMAN_* variable and none of the user's own OLLAMA_* or LLAMA_ARG_* settings reach it.
• The port is chosen once, by binding, from 11435 across 20 ports, and stored. 11434 is left alone: it is the default of a runtime the user may run themselves, which is never touched.
• An orphan is recognized. The pid and the port are stored. At boot a listener on the port that answers /api/version and whose pid runs a binary under Codeman's runtime folder is Codeman's own leftover from a server that was killed: it is adopted or replaced. Any other listener is "Port 11435 is in use by another program", with "Use another port" as the way out.

The memory budget is measured

Before the runtime exists the budget is an estimate from the installed memory. Right after it starts, the runtime prints the device total in its log line "inference compute" (total="37.4 GiB" on a 48 GB MacBook). Codeman reads that line, stores the figure and judges the chosen model against it before a byte of the model is downloaded. A model that no longer fits ends the job there with "does not fit" and the next model to try. From then on the window shows the measured budget and the word "estimated" is gone.

The need of a model is its peak (see "How much memory a model needs" above).

The model download

POST /api/pull with stream: true. A 23 GB pull sends about 1.4 million progress lines, so the stream is parsed piece by piece and never kept. Cancel aborts the request, which stops the download. Resume pulls again and the runtime continues its partial files. An error line or a cut stream is retried up to five times. Once the pull reported its total, every size shown for the model is that total.

The check

After the pull, in this order:

1. One tiny chat with reasoning_effort: "none". It loads the model.
2. GET /api/ps must show the whole model on the GPU (size_vram equal to size) at the chosen context length. For a llama.cpp build both figures are the buffers llama-server reports allocating (weights, K/V cache, compute), so this proves the whole load is on the GPU. For an MLX build it proves only the weights: the runtime books an MLX model at its weights (runtime issue 19417), so for those rows the catalogue's figures are all there is.
3. One tool probe on /v1/chat/completions (a tools array, reasoning_effort: "none", max_tokens 1024). It passes on a tool call whose arguments parse.

A failure is a state with a code and one next step, never a bare "failed": does-not-fit and no-tool-call offer the next model, probe-unfinished (the token limit ended the probe, so it proves nothing) and load-failed offer "Try again". The digest the runtime reports for the model is stored when the check passed.

The coding agent

When the chosen coding agent is not installed, the job installs it with the CLI registry's own stock install command, through the same runner as App Settings > CLIs (src/cli-install-runner.ts). The rules:

• Only a CLI that is a harness row of the local model catalogue.
• Only the compiled stock command. A ~/.codeman/clis.json override cannot change what runs here.
• Only after the person's consent click, on a screen that showed the literal command.
• Admin only in multi-user mode, and written to the admin audit log (cli_install, with via: "local-model").
• One installer per CLI at a time, shared with App Settings: while one runs, the other answers that it is busy.
• Afterwards the cached lookups of the CLI's binary are forgotten and the binary is looked for again. Still missing is the failure harness-failed, with the installer's last lines.

This step does not need the cliManagementEnabled setting. That setting gates the general CLI management surface; this path can only ever run one vetted command for one catalogue row.

The connection

Two writes outside the local model folder, both undone by Remove:

1. One record in custom-model-hosts.json:

   {
     "id": "local-ollama",
     "label": "This Mac",
     "baseUrl": "http://127.0.0.1:11435",
     "authStyle": "bearer",
     "models": ["qwen3.6:35b-a3b"],
     "defaultModelId": "qwen3.6:35b-a3b",
     "defaultContextLength": 131072
   }
   

   defaultContextLength is the chosen context: the runtime's model listing states none, so this field is what reaches the coding agent as its context limit.

2. customModelEndpointsEnabled: true in settings.json. The value it had before is kept in the state file.

The id local-ollama is reserved. A new endpoint cannot take it, and a record with that id that this feature did not write is never overwritten: the install says so before anything is downloaded. While the record is the managed one it cannot be edited or deleted in App Settings and the five-minute discovery sweep leaves it alone, so its model list stays the one model the install checked.

Switching the setting on also starts that five-minute sweep for every other saved endpoint.

Starting a session

"Start a session" starts the coding agent on the endpoint local-ollama through the existing custom model launch. Before a coding agent is STARTED on that endpoint the server checks, in this order:

• the runtime child Codeman started is alive (whatever else answers on the port is not ours);
• the digest the runtime reports for the installed model is the one stored at install time;
• the model asked for is the one that was installed and checked.

That check runs wherever a new agent process comes to be on the endpoint: a session created with the model, a model applied to a running session (which restarts its agent), the attach that starts an agent in a session that has none or whose agent exited, and the rebuild of a session after a reboot (that entry is skipped with "the local model is not ready" and stays on offer). An agent that is already running when Codeman restarts is not started by Codeman, so it is not asked: it keeps running on its port, see below.

A different digest refuses the launch: "The model on this Mac changed since it was installed. Remove it and install it again." The digest is also read again at boot and when the runtime is started.

From a Claude Code session: cloud plus local

The codeman skill (skills/codeman) lets an agent in a Codeman session hand work to this model, so Claude in the cloud plans and reviews while the model on this Mac does the bulk: summaries, classification, first drafts, and files that should not leave the machine. local_model reads GET /api/local-model and says what is installed or why nothing can run; spawn_worker <name> local starts the runtime when it was stopped (as Start does), then starts the installed coding agent on local-ollama through the same quick-start launch and the same checks as above; sendmark gives the worker a task and reads its answer back from a file it writes in its own folder. The skill never installs a model and never answers a confirmation, the runtime still answers one request at a time, and while a worker it opened is live, Stop and Remove are refused until it is closed.

Stop, Start, Remove

• Stop stops the runtime and keeps it from starting at the next boot. Start undoes that.
• Remove stops the runtime, removes the endpoint record, switches the setting off again only when this feature switched it on and no other endpoint is saved, and deletes the runtime and its private home. The downloaded model is deleted only when asked. The coding agent stays. While a live session uses the model, Remove answers with the session names instead.

Where everything is

All under the data directory (~/.codeman, or the instance's own):

local-model.json                           the state file (mode 0600)
local-model/
  runtime/ollama-0.40.15.tgz.partial       the runtime archive while it downloads
  runtime/ollama-0.40.15/                  the unpacked runtime (about 579 MB)
  models/                                  the model store
  home/                                    the runtime's private HOME (its key pair is here)

Nothing is written or deleted outside local-model/ except the state file beside it and the two writes of the connection. The state file holds no path, so it can never point a delete somewhere else. The user's own ~/.ollama is never created, read or removed.

local-model.json is not part of a config backup. custom-model-hosts.json and settings.json are, so restoring an older backup can take the endpoint record away while the model stays installed. The record is put back the next time Codeman starts (the setting is left as the backup had it). A restore can also bring a record back after a Remove: it then points at a port nothing listens on, is an ordinary endpoint again, and can be deleted in App Settings.

The job's states

State	A server restart	Cancel	What can be done
idle	stays idle	refused	install
installing	becomes paused at the same step	becomes paused	cancel
paused	stays paused	refused	resume, install, remove
failed	stays failed, same error	refused	resume, install, remove
installed	stays installed; the runtime is started unless it was stopped	refused	start a session, stop or start, remove

One job at a time. A job a server restart interrupted comes back paused and never continues without a click. Cancel interrupts the step in flight (the download, the pull, the request, the installer's process group); in the last step it is honoured before the first write, and once the two writes began they finish.

On a Mac without developer tools

Coding agents call git and python3. On a Mac without Xcode or the Command Line Tools those are stubs that open macOS's "install the command line developer tools" window. So the window says one sentence wherever the person commits or finishes, and only when /usr/bin/xcode-select -p fails:

Coding agents use git and Python. This Mac has no developer tools yet, so macOS will offer to install them the first time an agent needs one.

Codeman itself only ever runs /usr/bin/xcode-select -p for this, bounded, and never git or python3.

What to know before relying on it

• The runtime's port has no login. Any program on this Mac, and any web page served from localhost, can use the model server and change its models. That is why the endpoint record lists one model, why the sweep leaves it alone, and why the digest is compared before a session starts.
• A Codeman restart stops the runtime. A session that is mid-answer on the local model loses that answer. The runtime is started again with the server, and the first request loads the model again (seconds).
• An open session keeps calling the runtime's port. The coding agent was given the address when it started. While the runtime is stopped, or for the seconds of a Codeman restart, its requests go to a port nothing of ours holds, and any program on this Mac could bind that port and answer them. That is why Stop and "Use another port" are refused while a session uses the model: close those sessions first.
• The runtime does not enforce the context length. A prompt longer than the context was accepted in testing. Do not rely on the runtime to stop one.
• Most memory figures are calculated, not measured. One model was run on one Mac; the others are computed from their architecture, and the runtime settings they assume (an 8-bit K/V cache, flash attention, a 1 GiB prompt cache) were read in its source, not yet observed on a Mac. "Recommended" appears only for the pair we ran, on the Mac we ran it on.
• The model stays in memory for ten minutes after the last request, then the runtime unloads it.

Keeping the catalogue current

The catalogue (src/config/local-model-catalog.ts) and the runtime pin are compiled into Codeman on purpose: no file or download decides what is recommended or run. What the registry serves can still move away from them: a tag is pushed again or removed, the build a Mac downloads changes size, Ollama ships a new version. The workflow .github/workflows/local-model-catalog.yml looks for that every Monday at 06:00 UTC. It runs scripts/check-local-model-catalog.mjs, which also runs locally: npm run check:local-model-catalog prints the report and exits 0 (clean), 1 (drift) or 2 (the check itself failed).

For every model row it checks:

• The tag still exists in the registry.
• The size of what a Mac downloads is within 0.1% of the row's bytes. A tag can be a manifest list with one build per runner, and Ollama on Apple silicon takes the mlx build when there is one, else llamacpp, else ggml. So the check asks for the tag the way Ollama asks, picks the build a Mac picks and sizes it the way a pull counts it (every layer plus the config). The plain manifest other tools get for the same tag is the llamacpp build, which is not what a Mac downloads.
• The runner of that build, once a row names its runner: a Mac still pulls an mlx build for an mlx row and a GGUF build (llamacpp, or a manifest that names no runner) for a llamacpp row. A row's memory figures follow its runner, so a tag that gains or loses an mlx build is drift even at about the same size.
• The build a Mac pulls runs on the pinned runtime: its config declares no newer Ollama (requires) than OLLAMA_RUNTIME_PIN.
• The pinned digest, once a row carries one: a Mac would still pull that manifest. The field is digest and its value is what /api/tags reports after the pull. The catalogue schema does not have it yet: LocalModelEntrySchema must gain an optional digest under exactly that name, or the check never compares it. Until then the report lists today's digests, rows with measured figures first, because without a digest a rebuild of about the same size goes unnoticed (a switch between the mlx and the gguf build does not: their sizes differ by 2% or more). Pinning digests is a decision, not a quick fix: today the catalogue holds no digest on purpose, and test/local-model-catalog.test.ts ("holds no download data") refuses a sha256 anywhere in it, so that rule changes with it.

For the runtime pin it checks whether a newer Ollama release exists, and whether the pinned ollama-darwin.tgz is still on GitHub with the pinned size and sha256.

It reads manifests and configs only, never a model layer. It does not check the memory figures, the speeds or whether a coding agent works with a model: no manifest says anything about those, and they change only when somebody measures on a Mac.

The issue. Drift opens one issue, "Local model catalogue: weekly check", or, while that issue is open, replaces its description with the new report and adds a comment. A clean week comments on the open issue and does nothing when there is none. A check that failed (the network, a rate limit, an answer that did not parse) fails the run and leaves the issue alone; what it found is in the run's summary. The schedule runs the copy of the workflow on master, and so does gh workflow run local-model-catalog.yml.

What a maintainer does with it, in a pull request each time:

• Size drift or a changed digest: correct the row. When the build itself changed (another runner or quantization), its memory figures and speed no longer describe it: mark the row estimated until somebody measures it again.
• A tag that is gone, or a build that needs a newer runtime: replace or drop the row, or move the runtime pin first.
• A newer Ollama: read its release notes, then move OLLAMA_RUNTIME_PIN (version, URL, size and sha256, checked against the release's sha256sum.txt) and run an install on a Mac before merging.
• A pinned archive that changed: a security question first. Every install refuses it at the hash compare, so nobody gets it, but find out why before moving the pin.

Close the issue once the fix is merged. A catalogue or pin change reaches users only in a release: an installed Codeman keeps the catalogue it was built with until it is updated.

Source

What	Where
Hardware probe, catalogue, recommendation	src/local-model/hardware.ts, src/config/local-model-catalog.ts, src/local-model/recommend.ts
Runtime pin, download, environment, supervisor	src/local-model/runtime.ts
Runtime HTTP client, pull parser, checks	src/local-model/ollama-client.ts
Developer tools check	src/local-model/developer-tools.ts
Job (state table, reducer, view, service)	src/local-model/install-job.ts
State file	src/local-model/state-store.ts
Connection and its undoing	src/local-model/connect.ts
Paths	src/local-model/paths.ts
Routes	src/web/routes/local-model-routes.ts
Shared CLI install runner	src/cli-install-runner.ts
Serialized writer of the endpoint store	src/custom-model-hosts-writer.ts
Weekly catalogue check	scripts/check-local-model-catalog.mjs, .github/workflows/local-model-catalog.yml