NEW One spend cap. Eleven model APIs. Meet your stack ↗
LockpinStart building ↗
Built for agents that keep going ↗

Let your agents work.
Keep spend in bounds.

Give every task a budget that is checked before each call. Cut stale context, catch repeated work, and stay in control.

Set up Lockpin ↗▷ Watch a real run
Your models. Your machine. No extra AI calls.
Lockpin / Playground RECORDED RUN
WORKSPACE Runs on your machine

More progress. Same budget.

One log-reading task, with context trimming off and on.

step 30 of 30
TASK BUDGET$0.20Enforced cap
STEPS COMPLETED30 / 30With trimming
TOTAL SPEND$0.1491Under budget
Context tokens per step
Trimming offTrimming on
Spend against the $0.20 capDashed line: cap · Hatched area: next call’s worst case
Trimming offStopped by the cap
24 / 30 steps$0.1458
Trimming onFinished
30 / 30 steps$0.1491

Without trimming, the next call’s worst-case cost exceeded the remaining budget. Lockpin blocked it before it reached the provider.

CONTEXT: 10,367 TOKENS
00 — task startsDrag to inspect any step30 — complete

All 30 steps finished under the same $0.20 cap.

Recorded on Muse Spark 1.3 (Standard), October 19, 2026. Same task. Same $0.20 cap.
Replay data supplied with this page; results vary by agent and task.

One layer of control. The models you already use.

OpenAIAnthropic✧ Gemini∞ MetaMistraldeepseek
Also supports xAI, Groq, Together, Fireworks, OpenRouter, and your own compatible server.
⌘ Your agent⊣ Lockpin✧ Your model
A little control goes a long way

Keep the progress.
Cut the runaway spend.

Between your agent and its model API, Lockpin keeps every task accountable to its budget and its work.

01 / Budget guard

Set a limit. Mean it.

Before a call goes out, Lockpin reserves its worst-case cost. If it won’t fit, it won’t run. Even when requests arrive at once.

02 / Context trimming

Less history. More headroom.

Swap old tool output for a small, recoverable note. Keep recent results, and trim only when savings outweigh the prompt cache you break.

03 / Work history

Catch the work
your agent does twice.

See deleted files, undone changes, and checked-off plans with no matching work. Every finding connects to a step and its cost.

Project activityRecorded replay
STEP 07
login.go recreatedDeleted at step 6. Written again at step 7.
STEP 10
config.go revertedStep 9 changes were undone.
$ lockpin restore -step 4 validate.go

Your team. Their own limits.

Give each developer or agent a key with daily, weekly, or monthly budgets.

A heads-up before the cap.

Group spend with tags. Send Slack or webhook alerts at 80% and when a call is blocked.

Every task, in view.

Follow spend, cache hits, trimmed tokens, and added latency in the live meter.

The difference in one run.

Explore the replay ↗
30 / 30Steps finished with trimming.
24 finished without it.
20% lessSpent across the 24 steps
both agents completed.
658kTokens of old tool output
kept out of requests.
ZeroExtra model calls
made by Lockpin.

First three figures describe the single supplied run, not an average or a guarantee. Exact tokens omitted: 658,854.

Made for your existing workflow

Same SDK.
A smarter route.

Point your agent at Lockpin, name the task, and set its budget. Your provider’s SDK, streaming, and tool calls stay familiar.

  1. Run Lockpin locally or on your serverOne small program. No database to set up. Download it, run lockpin init, and it tells you what to set.
  2. Change the URL. Add two headers.Track a task and give it a maximum spend.
  3. Keep an eye on the live meterSee where your budget goes, call by call.
from openai import OpenAI

client = OpenAI(
    # was https://api.meta.ai/v1
    base_url="http://127.0.0.1:8787/v1",
    api_key=MODEL_API_KEY,
    default_headers={
        "X-Lockpin-Task": "fix-login-bug",
        "X-Lockpin-Budget-Usd": "0.50",
    },
)
Edit the task and budget to personalize these examples. Meta-compatible endpoint; other providers use their own routes.
Local by design

Your code isn’t
our business.

Lockpin runs where your agent runs and connects directly to your model provider. Your code and conversations never pass through us.

  • Your prompts and API keys stay with youSent directly to the provider you choose.
  • Your project snapshots stay localSeparate from your Git history.
  • No hidden model callsJust arithmetic and file snapshots.
A few things worth knowing

Good questions.
Clear answers.

Which models does it work with?

Meta, OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, Together, Fireworks and OpenRouter, with prices for 56 models built in, including cached input, cache writes and long-context rates. Add any other model, or an OpenAI-compatible server of your own, with one small file.

Does it change what the model sees?

The spend cap never does. Trimming and project notes do, so both are off unless you turn them on. Trimming can also be switched per request.

Does it slow my agent down?

Hardly. Far from the cap, Lockpin sizes each call itself and sends it straight on. Only near the cap does it ask the provider for an exact input count, where the provider offers one. The live meter shows how many milliseconds Lockpin adds to each call.

What happens when a task hits its budget?

The call that would go over is stopped before it reaches the provider. Your agent gets HTTP 402 with the numbers, so it can save its work and stop cleanly. You can raise the budget with one header.

Does Lockpin use AI itself?

No. All of its logic is plain arithmetic and file snapshots, so it adds nothing to your bill.

Build with a little more certainty

Big ideas.
Bounded budgets.

Let your agents do their thing.
Give every task a limit you can live with.

Set up Lockpin ↗Explore the demo