Local model guide
A Mac or Linux machine with Ollama, put to work only when it is idle. About ten minutes, no bill per call, and a verified result at the end.
Where things stand. SpareTokens is a private pilot. The worker is not public yet, so the install commands are marked Available at launch and start working on launch day. Everything else on this page is how the worker behaves today. The waitlist is how you get an invite.
Before you start
This guide is for a computer you own: a Mac, a Linux box or a gaming PC with a local model through Ollama. LM Studio and MLX servers work the same way. Small local models can handle computation and solver checks. Tasks that need web search or a frontier model are matched to other sources, so your machine is never asked to do work it would fail.
Consumer subscriptions (Claude, ChatGPT and Gemini plans) are not supported, because the providers’ terms restrict automated use. The worker refuses a subscription login and points you here. Use an API key or a local model.
Step 1: Start Ollama and pull a model
About 3 minutes plus the download
Install Ollama from ollama.com if you do not have it, then pull any model your machine can run. For example:
ollama pull llama3.2
Ollama serves an OpenAI-compatible API on port 11434. The worker looks for it there, and for LM Studio on 1234 and MLX on 8080.
If something goes wrong
- The pull is slow: models are several gigabytes. You only do it once.
could not connect to ollama: start the Ollama app, or runollama servein another terminal.
Step 2: Install the worker Available at launch
About 2 minutes
On Linux, install bubblewrap first (sudo apt install bubblewrap), because computation tasks run in a sandbox. macOS has one built in.
pipx install sparetokens
sparetokens --version
You will see
sparetokens 0.1.0Other systems and installers are on the install page.
If something goes wrong
command not found: pipx: install it withbrew install pipxorsudo apt install pipx, then open a new terminal.
Step 3: Run the self-check
About 1 minute
sparetokens doctor
You will see
SpareTokens 0.1.0 self-check
os: Darwin 27.2.0 (arm64)
python: 3.12.8
sandbox: available (macOS sandbox-exec)
providers:
anthropic-api not ready: no Anthropic API key found at env:ANTHROPIC_API_KEY (set ANTHROPIC_API_KEY or point the source at a keychain item; it stays on this machine)
openai-api not ready: OPENAI_API_KEY is not set in the environment (it stays on this machine)
xai not ready: XAI_API_KEY is not set in the environment (it stays on this machine)
openrouter not ready: OPENROUTER_API_KEY is not set in the environment (it stays on this machine)
local ready
terms: subscription logins are not used (API keys and local models only); https://sparetokens.ai/docs/providers
config dir: ~/.config/sparetokens (not created yet); registered as: not registered
coordinator: http://127.0.0.1:18795 reachable (HTTP 200)
All good.Real output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~. Here a scripted test server stands in for Ollama, so the local line reads ready. The lines for Claude and ChatGPT subscription logins are left out, because those are never used.
The local provider should read ready and the sandbox line should say available.
If something goes wrong
local not ready: no local OpenAI-compatible server reachable: Ollama is not running. Start it and run the check again.no working OS sandboxon Linux: install bubblewrap. Inside Docker or on a hardened host the sandbox may be blocked, so run the worker on the host or in a VM. See sandbox problems.
Step 4: Run setup
About 4 minutes
Press Enter through every question. For a local model the defaults are idle time only, after ten idle minutes and while plugged in, with the best model you have. The wizard ends with one tiny test task so you see the whole loop.
sparetokens setup

Text version of this walkthrough
SpareTokens setup. Every question has a safe default; pressing Enter keeps it. Nothing here spends money unless you say so.
Step 1 of 9: sign in or stay anonymous
Signing in with GitHub is not available yet. You will get a random contributor ID stored on this machine; it can be linked to an account later.
Registered a new anonymous contributor (contributor-2349a7c4).
Step 2 of 9: how you appear publicly
How should you appear on the leaderboard?
1. Anonymous (Contributor #N)
2. My handle (a username I choose, no real name)
3. First name and last initial (Paul F.)
4. Full name
Choose [1] 2
Handle (3-24 letters, digits, _ or -) [contributor-2349a7c4] rowan_lab
Show a link (GitHub or website)? (y/n) [n]
Show a country? (y/n) [n]
Make models used public? (y/n) [n]
Make providers used public? (y/n) [n]
Make hardware type public? (y/n) [n]
Make tokens donated public? (y/n) [n]
You will appear as: rowan_lab
Step 3 of 9: connect what you have
1. local-ollama: ollama is running with 1 model(s) (models: fake-codefill)
API keys are off unless you pick them; only free local models are pre-selected. Consumer subscriptions are not offered (provider terms).
Your provider account and API key are yours to protect: use your own account, keep keys on your machine, follow your provider's terms, and set a spend limit at the provider too.
Numbers to use, comma separated, or 'none' [1]
Step 4 of 9: how much to give from local-ollama
Pick a rule:
1. idle only: when idle 10 minutes and plugged in (default)
2. manual only
Choose [1]
Step 5 of 9: which models for local-ollama
Pick:
1. Best available (default)
2. Cheapest that qualifies
Choose [1]
That picks: fake-codefill (strongest listed model)
unlocks: computation checks, small coding tasks, search heuristics, solver runs, re-checking earlier results; not larger searches (needs a stronger model); not literature checks (needs a stronger model); not Lean proofs (needs a stronger model); not refereeing (needs a stronger model)
Step 6 of 9: what to work on
Default: wherever it is needed most (all task types).
Opt out of any task type? (y/n) [n]
Step 7 of 9: when it may run
Never run on battery? (y/n) [y]
Quiet hours as HH:MM-HH:MM, or none [none]
Your own Claude Code and Codex sessions always take priority. These times are recorded now and applied by the background service.
Step 8 of 9: test run
Run one tiny test task now with local-ollama? It costs about a few seconds of compute. (y/n) [y]
Claimed, worked, submitted, verified in 0s. Cost $0.0000, 1000 tokens.
Step 9 of 9: turn it on
Run in the background now (a launchd or systemd user service)? (y/n) [n]
The background service is off. Turn it on any time with `sparetokens service install`, or donate by hand with `sparetokens run --source NAME`.
Your setup, in one paragraph:
You appear as rowan_lab. local-ollama: only when the machine has been idle 10 minutes and is plugged in. Working wherever needed most. Never on battery.
Every source also has a per-task and a per-day cap, three failures in a row pause it, and your own sessions always come first. Look at the rules with `sparetokens config explain`, see spend with `sparetokens budget`, and pause with `sparetokens pause 2h`. To remove everything: `sparetokens uninstall`.You will see, in full
SpareTokens setup. Every question has a safe default; pressing Enter keeps it. Nothing here spends money unless you say so.
Step 1 of 9: sign in or stay anonymous
Signing in with GitHub is not available yet. You will get a random contributor ID stored on this machine; it can be linked to an account later.
Registered a new anonymous contributor (contributor-2349a7c4).
Step 2 of 9: how you appear publicly
How should you appear on the leaderboard?
1. Anonymous (Contributor #N)
2. My handle (a username I choose, no real name)
3. First name and last initial (Paul F.)
4. Full name
Choose [1] 2
Handle (3-24 letters, digits, _ or -) [contributor-2349a7c4] rowan_lab
Show a link (GitHub or website)? (y/n) [n]
Show a country? (y/n) [n]
Make models used public? (y/n) [n]
Make providers used public? (y/n) [n]
Make hardware type public? (y/n) [n]
Make tokens donated public? (y/n) [n]
You will appear as: rowan_lab
Step 3 of 9: connect what you have
1. local-ollama: ollama is running with 1 model(s) (models: fake-codefill)
API keys are off unless you pick them; only free local models are pre-selected. Consumer subscriptions are not offered (provider terms).
Your provider account and API key are yours to protect: use your own account, keep keys on your machine, follow your provider's terms, and set a spend limit at the provider too.
Numbers to use, comma separated, or 'none' [1]
Step 4 of 9: how much to give from local-ollama
Pick a rule:
1. idle only: when idle 10 minutes and plugged in (default)
2. manual only
Choose [1]
Step 5 of 9: which models for local-ollama
Pick:
1. Best available (default)
2. Cheapest that qualifies
Choose [1]
That picks: fake-codefill (strongest listed model)
unlocks: computation checks, small coding tasks, search heuristics, solver runs, re-checking earlier results; not larger searches (needs a stronger model); not literature checks (needs a stronger model); not Lean proofs (needs a stronger model); not refereeing (needs a stronger model)
Step 6 of 9: what to work on
Default: wherever it is needed most (all task types).
Opt out of any task type? (y/n) [n]
Step 7 of 9: when it may run
Never run on battery? (y/n) [y]
Quiet hours as HH:MM-HH:MM, or none [none]
Your own Claude Code and Codex sessions always take priority. These times are recorded now and applied by the background service.
Step 8 of 9: test run
Run one tiny test task now with local-ollama? It costs about a few seconds of compute. (y/n) [y]
Claimed, worked, submitted, verified in 0s. Cost $0.0000, 1000 tokens.
Step 9 of 9: turn it on
Run in the background now (a launchd or systemd user service)? (y/n) [n]
The background service is off. Turn it on any time with `sparetokens service install`, or donate by hand with `sparetokens run --source NAME`.
Your setup, in one paragraph:
You appear as rowan_lab. local-ollama: only when the machine has been idle 10 minutes and is plugged in. Working wherever needed most. Never on battery.
Every source also has a per-task and a per-day cap, three failures in a row pause it, and your own sessions always come first. Look at the rules with `sparetokens config explain`, see spend with `sparetokens budget`, and pause with `sparetokens pause 2h`. To remove everything: `sparetokens uninstall`.Real output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~.
If something goes wrong
- The test did not pass: a small model may not follow the test task. That proves nothing is wrong with your setup. Run
sparetokens setupagain, or go on to step 6. Nothing usable was found: no server answered. Start Ollama and run setup again.
Step 5: Choose when it gives, and preview it
About 2 minutes
Idle time is the default for local models. To run only overnight, pick the hours and preview the result:
sparetokens give mac --idle --nights
You will see
Saved local-ollama: idle time (after 10 idle minutes), only 22:00-07:00. Rules are in ~/.config/sparetokens/rules.toml. Check what it will do with `sparetokens status`, or preview it with `sparetokens preview local-ollama`.Real output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~.
sparetokens preview mac
You will see
Preview for local-ollama (idle): the last 7 days as an example week.
Example week, not your data: your own use is not recorded, so this shows what the setting would do on a typical week.
given 61.8 h
expired 29.2 h (unused by anyone)
kept 77.0 h (stays with you)
hours of the example day times 7; about 3.7 kWh at 60 WReal output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~. The preview uses an example week and says so, and the kilowatt-hour figure assumes 60 watts.
The machine stops and gives the task back the moment you return. Add --any-power to allow battery. More styles are on giving styles.
If something goes wrong
idle, power and temperature are not known yet, so it does not run: the worker will not guess. It runs when it can read that the machine is idle, plugged in and cool. Check withsparetokens status.- A run that names a source with an idle rule waits for idle too. For a first hand-started run, do not pass
--source.
Step 6: Run one task
About 2 minutes
Start one task by hand. Naming no source means only the flags limit this run, so it starts now:
sparetokens run --provider local --model llama3.2 --allow-bash --no-web --max-tasks 1
You will see
SpareTokens worker: provider local (local), tier 1, capabilities ['bash', 'solver', 'lean'], budget n/a, max tasks 1
No giving rule for local (it is not a configured source), so only these limits apply: --max-tasks 1.
Task 1 [code_fill] Write a small function: abundant count (instance 0) (cap n/a)
submitted as result 2 (cost $0.0000, tokens 2000, spend flag: none)
verification: verified: all 8 answers recomputed exactly
Done. Tasks submitted: 1. Dollars: $0.0000, tokens: 2000, wall: 0s.Real output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~. The scripted test model used here stands in for yours.
Local models are tier 1, which unlocks computation, code and solver tasks. sparetokens config models NAME shows what each model unlocks and why it does not unlock the rest.
If something goes wrong
failed locally: local request failedwith a 404: the model name does not exist on the server. Checkollama listand pass that exact name to--model.- A task was rejected: small models sometimes get the answer wrong, and the verifier says so. It costs you only time. Three failures in a row pause the source for 24 hours, and
sparetokens config reset-breaker NAMEclears it after you check why. - The machine got hot or busy:
sparetokens pause 2hpauses everything.
Your first verified result
Look at your history. The line ending in verified is your first verified result:
sparetokens history --limit 5
You will see
2026-11-21 06:34 local task 1 code_fill verified $0.0000 2000 tokens fake-codefill
2026-11-21 06:34 local-ollama task 13 canary verified $0.0000 1000 tokens fake-codefillReal output from the real worker, captured against a throwaway coordinator with throwaway signing keys. Paths are shortened to ~. The first line is the setup test, which carries no leaderboard weight. The model name reads fake-codefill because a scripted test model stood in for Ollama.
To let it run by itself, install the background service. It only runs when your idle rule says so:
sparetokens service install
Pause with sparetokens pause 2h, resume with sparetokens resume, and remove everything with sparetokens uninstall.
What “verified” means. A machine check written and owned by SpareTokens passed. It does not by itself mean a result is new or important, and no discovery has been claimed. See what has been verified so far and how each check works.
Schedules, pausing and uninstalling
Quiet hours, battery rules and how to stop cleanly.
More ways to give
Fixed amount, a share, everything spare or a one-off boost.
Add an API key too
Use expiring credit as well, and let the worker spend whatever expires first.
Troubleshooting
Every message the worker prints, and what to do about it.
Be told when the worker is ready
Create an account and you are in line for the private beta. We email you when it is your turn.