$— saved by users · — tokens

Distil your context

lower costmeasured per task, against vanilla harness (opencode).

more work from your subscription

Improve focus

improvement in accuracyon average across all test runs.

DistilCode sends your code directly to your model provider (e.g. Anthropic), so your code is never shared with us. Read more about Privacy and security →

What we do for you

See more 12

Finished tasks shrink to a summary. When a todo is ticked off, its whole transcript is compressed into a few sentences, so the next task starts with a lean context.

Scroll

opencodestock harness

$25.68

1128M

DistilCodeHarnessno SmartCompact

$17.72

−31%

699M

−38%

DistilCodePRMbase harness

$12.30

−52%

311M

−72%

DistilCodePRM+ Ponytail (lazy developer)

$9.88

−62%

247M

−78%

DistilCodePRM+ Caveman (terse responses)

$8.77

−66%

224M

−80%

What you get

Our harness cuts your cost by 66%.

Same model, same long-horizon tasks from AgencyBench V2. DistilCode Premium Harness finishes them for $8.77 where the stock opencode harness spends $25.68, and burns 224M tokens instead of 1,128M.

See the published runs →

What you get

Our harness removes noise from your context.

It distils your context down to what the task needs.

Base harnessDistilCodeLossy, irreversible compactionLossless, reversible compression

Turns in a long session · illustrative · hover to inspect

How we save tokens

14 different ways, all to save your tokens.

01 / 07

AutoJev model selectionPRM

Match the starting model to the task instead of using the same class for every session.

Inspect the first prompt. Up to 4,000 characters of the first normal prompt are sent to the AutoJev classifier.
Choose a class. Jev returns a model class suited to the requested work.
Use the configured model. DistilCode switches to the model assigned to that class before sending the prompt.

How it works →

DistilCode — ~/projects/store

New session. AutoJev picks the model for your first prompt.

Build+0-0auto·AutoJev

~/projects/storeenter sendctrl+p commands

Demo uses keywords. The real classifier reads only the first prompt, with no repository files attached.

1.0× → 1.0× cost per token vs Sol for everythingno change

Noise reduction

Bash output trimmer

Stop installs, builds and container pulls from flooding the context with lines nobody reads.

Rewrite the command. Known commands run through rtk, which prints a filtered version of the same output.
Keep what matters. Errors, warnings and summary lines stay; progress bars and repeats collapse.
Bound the rest. Anything still oversized is capped by lines and bytes, with the full output saved for targeted reads.

How it works →

$ npm install

4,838 → 4,838 tokens · 252 linesno change

Noise reduction

JSON compressor

Turn large API responses and data files into a compact view without dropping keys or anomalies.

Lossless first. Arrays of objects become a column table before any row is sampled.
Keep the anchors. First/last rows, errors, outliers and rows matching the request survive sampling.
Always recoverable. Editable JSON stays exact, and viewraw returns any compressed original.

How it works →

Extend cachePRM

Pay cached prices after a pause instead of rewriting the whole prompt prefix at full price.

Watch the window. The heater tracks when the cache was last written and refreshes just inside the five-minute TTL.
Refresh only when useful. Refreshes run while you type, answer a prompt, or wait on a tool, never on an abandoned session.
Cap the upkeep. Each session has a refresh limit, so a long absence still ends in one rewrite.

How it works →

Cold cache reformatterPRM

Use the moment the cache goes cold to rewrite a leaner prefix instead of the same bloated one.

Wait for cold. Collapsing only happens once the five-minute cache window has passed.
Find what went stale. Re-reads, edited files, repeated commands, volatile output and older images are flagged.
Leave a pointer. Each collapsed result becomes a placeholder that says why, and viewraw restores it.

How it works →

Smart compactionPRM

Replace stale transcript detail with an anchored work summary while retaining recent turns verbatim.

Protect recent work. A bounded recent tail stays in context so the immediate exchange is not reduced to a summary.
Anchor the hand-off. The summary schema explicitly captures objectives, important details, work state, next moves, and relevant files.
Retire stale bulk. Old tool output is truncated during summarisation and can be cleared after a successful compaction.

How it works →

3 more, running quietly

See remaining techniques

Those were the ones you can play with. These work in the background on every session.

Dynamic delegation of tasksBounded work can move into a specialist child session with an appropriate model class and tool permissions.Codebase graphing and traversalGet the relevant source, call paths, and change impact from a local repository graph.Task-specific tool adaptationGive each specialist the capabilities its job needs and match editing tools to the active model.

See remaining techniques →

Pricing

Start free. Go Premium for 66%.

You still pay your model provider directly, just less of it. Premium Harness adds the techniques that save the most on long sessions.

Harness

−31% cost · −38% tokens

31% lower cost and 38% fewer tokens than the stock harness
Delegation, terse tool output, codebase graph and the other core techniques
Ponytail mode and Caveman mode
5 AutoJev model selections
10 Smart compaction instances

Best value

Premium Harness

$5/month

−66% cost · −80% tokens

Everything in Harness, plus

Unlimited Extend cacheKeeps the prompt cache warm between turns
Unlimited Cold cache reformattingCollapses stale tool results before a cold cache rewrite
Unlimited Dynamic compactions a monthLossless, task-localised compression of your context
5,000 Smart compactions a monthHigher-quality compactions assisted by System One models
2,000 AutoJev selections a monthAutomatically chooses the right models every week. Saves 41% more than a fixed model.

Teams

Per seat/month or year

−66% cost · every seat

Premium Harness for every seat, plus

Org dashboard of tokens, cost and savings
Roles, teams, invites and company-domain join
Bring-your-own-key and free-mode policies
Annual contracts, PO and invoicing

Questions

Do I need a separate subscription?

No. DistilCode is a harness, a way to talk to the models you already pay for. If your provider supports third-party harnesses, it works with DistilCode.

How does DistilCode work for teams?

Create an organization, invite your developers or let anyone with a verified company email join, and assign seats. Admins get a shared dashboard of tokens, cost and savings, and can require bring-your-own-key so code never passes through our servers. Annual contracts with purchase orders and invoicing are available, see Trust or contact us.

What does Premium Harness add over Harness?

Higher limits: 5,000 Smart compaction instances and 2,000 AutoJev selections a month, against 10 and 5 in total on Harness. That takes savings from 31% to 66%. At $200 a month of AI spend, Premium Harness saves about $65 more each month than Harness, after its $5 price.

Where does my code go?

Prompts and code go directly to your chosen model provider, not DistilCode servers. Your data is never saved to DistilCode servers. See the Privacy and Security.

Can I check the numbers myself?

Yes. The savings study publishes every transcript, a CSV of all runs and a hash of the raw data. Anyone can rerun the benchmark with the released build.

Distil your contextR§e#m0xo~ve% //n¶o{i;s·e^<tool_result>…ERR+412{…}██node_modules/∷retry0b1101$»log:#%~[object Object]200 OK…+3,812 linesTODOwarn\n\n≈at Object.<anonymous>

What we do for you

What you get

Our harness cuts your cost by 66%.

What you get

Our harness removes noise from your context.

14 different ways, all to save your tokens.

AutoJev model selectionPRM

See remaining techniques

Start free. Go Premium for 66%.

Harness

Premium Harness

Teams

Questions

Distil your context