Send less context.
Keep control.

TrimLayer returns the smallest safe context for your application. You choose if, when and where that context is sent next; TrimLayer does not call a model provider on your behalf.

Choose an integration

TrimLayer is an optimize-only service: it returns a verified context, while your application makes any later model request directly.

Create a project key

Open Console → API keys, create a key and store the displayed secret immediately. It begins with cl_ and is shown only once. Choose an expiry that matches the environment.

TRIMLAYER_KEY=cl_your_project_key
TRIMLAYER_BASE_URL=https://api.trimlayer.com

Compiler quickstart

The compiler returns context for your application to use elsewhere. A task or query improves focused selection but is optional. Use domain: auto unless your application has a reliable domain declaration.

Raw HTTP

curl https://api.trimlayer.com/v1/context/optimize \
  -H "X-TrimLayer-Key: $TRIMLAYER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "content": "Your long document, prompt or retrieved context...",
    "query": "Which cancellation conditions apply?",
    "domain": "auto"
  }'

TypeScript

import { TrimLayerCompiler } from "@trimlayer/openai";

const compiler = new TrimLayerCompiler({
  baseURL: "https://api.trimlayer.com",
  trimLayerKey: process.env.TRIMLAYER_KEY!
});

const result = await compiler.optimize(longDocument, {
  query: "Which cancellation conditions apply?", // optional
  domain: "auto"
});

// Always use this field. It contains the accepted candidate or,
// when protected, the byte-identical original.
const contextForYourLLM = result.optimized_context;

Python

from trimlayer_openai import TrimLayerCompiler

compiler = TrimLayerCompiler(
    base_url="https://api.trimlayer.com",
    trimlayer_key=os.environ["TRIMLAYER_KEY"],
)

result = compiler.optimize(
    long_document,
    query="Which cancellation conditions apply?", # optional
    domain="auto",
)

context_for_your_llm = result["optimized_context"]

Understand the response

Use optimized_context as the downstream context. Other fields explain the decision and measure savings without storing prompt content.

{
  "decision": "optimized",
  "optimized_context": "The accepted, compact context...",
  "original_tokens": 9341,
  "optimized_tokens": 3641,
  "avoided_tokens": 5700,
  "reduction_percent": 61.0,
  "counting_method": "o200k_base",
  "recovery_available": true,
  "recovery_id": "clctx_0123456789abcdef0123456789abcdef",
  "latency_ms": 24.8,
  "verification_level": "balanced",
  "proof": {
    "exact_recoverable": true,
    "format": "text",
    "query_parts": 4,
    "query_parts_covered": 4
  }
}

Focused questions

A focused query tells the compiler which evidence is necessary. Keep the source in content and the task in query; do not concatenate them when your integration can keep them separate.

{
  "content": contractText,
  "query": "What notice and cure periods apply?",
  "domain": "auto"
}

{
  "content": contractText + "\n\nSummarize everything",
  "domain": "auto"
}

Leave query out for generic compaction. Expect a more conservative result because material may matter to an unknown future task.

Improve savings without losing the answer

TrimLayer reduces the context that is unnecessary for a specific task. The clearest way to improve a result is to state the decision, fact, exception or code path the model needs—not to relax safety settings.

Content behavior

TrimLayer does not apply one lossy rewrite to every input. It identifies the content shape, protects structures that must remain exact and returns the original whenever an eligible candidate cannot be proven useful.

Exact recovery

When a result contains a recovery ID, your application can request the exact original source. The recovered bytes are not a summary or regenerated text.

const exactOriginal = await compiler.recover(result.recovery_id!);

// Your application decides whether to send it in a follow-up turn.

Usage and quota

Completed requests produce content-free measurement metadata. The dashboard shows original tokens, dispatched tokens, avoided tokens, decision, latency, recovery cost and effective reduction.

Every optimize-only API or MCP call reserves quota using the measured original source tokens, including protected results. Free includes 200K input tokens/month at 6 requests or 30K input tokens per minute; Starter includes 5M at 20 requests or 200K input tokens per minute; Growth includes 50M at 60 requests or 600K input tokens per minute. The dashboard’s authenticated verification flow is reported separately as Test traffic; project keys cannot select that lane.

Errors and safe outcomes

HTTP errors mean processing failed. A 200 response with decision: protected means processing succeeded and the original is safe to use.

Security boundaries

• Keep secrets server-side and redact sensitive headers from logs.
• Use separate keys and environments for development, staging and production.
• Scope project keys to the smallest required environment and project.
• Use application-owned validation for consequential legal, health and financial outputs.
• Rotate a key immediately when exposure is suspected.

Production checklist

MCP integration

Use the local trimlayer-mcp server in Codex, Claude Code, Claude Desktop, Cursor or any stdio MCP host. It prepares context before model use; it does not call a model provider and never needs an OpenAI, Anthropic or Google key.

Codex: one shared setup

Add this to ~/.codex/config.toml for your account, or .codex/config.toml for one trusted repository. Codex Desktop, CLI and the IDE extension share this MCP configuration.

[mcp_servers.trimlayer]
command = "/ABSOLUTE/PATH/ai-optimize/bin/trimlayer-mcp"
startup_timeout_sec = 20
tool_timeout_sec = 120

[mcp_servers.trimlayer.env]
TRIMLAYER_BASE_URL = "https://api.trimlayer.com"
TRIMLAYER_API_KEY = "cl_REPLACE_WITH_PROJECT_KEY"
TRIMLAYER_MCP_TIMEOUT = "2m"
TRIMLAYER_ALLOWED_ROOTS = "/ABSOLUTE/PATH/TO/WORKSPACE:/ABSOLUTE/PATH/TO/DOCUMENTS"

Restart Codex, then use /mcp in the terminal UI to verify that trimlayer is connected.

Ask the host to use it first

Before reading the source, call trimlayer_optimize_file for:
/absolute/path/to/repository/internal/gateway/optimize.go

Task: Explain validation, quota reservation and recovery behavior.
Use only optimized_context for the first answer. If decision is protected,
use optimized_context unchanged. Retrieve only a named missing detail.

Claude, Cursor and other MCP hosts

Choose a local stdio MCP server in the client’s settings, set the command to bin/trimlayer-mcp, then copy the same environment variables. Keep TRIMLAYER_ALLOWED_ROOTS limited to folders the tool may read. On macOS and Linux, separate multiple roots with :; Windows uses ;.

For large context, optimize before reading it into the model.
Prefer trimlayer_optimize_file when a local path exists.
Pass the user's task as query and leave domain=auto unless known.
Use mode=balanced by default. Treat optimized_context as the source of truth.
If decision=protected, do not retry with weaker settings.
Use trimlayer_retrieve only for a specific omitted detail.

make build-mcp

TRIMLAYER_BASE_URL=http://127.0.0.1:8080 \
TRIMLAYER_API_KEY=$TRIMLAYER_KEY \
TRIMLAYER_ALLOWED_ROOTS=/absolute/path/to/documents \
./bin/trimlayer-mcp

Available tools: trimlayer_optimize_text, trimlayer_optimize_file, trimlayer_retrieve and trimlayer_gateway_status.