Build faster.Spend less.

Orek is an AI coding agent that finds the code your task needs. Connect your preferred AI services and get more from your AI budget.

Download

Provider usage billed separately

Buy once. Own it. Perpetual software license
Up to 94% fewer input tokens*
3× faster completion*

Code intelligence, built in.No separate infrastructure.

Orek combines search by meaning, exact matches, and code relationships to give your model relevant source for the task.

Illustrative prompt

Where is retry logic, and what calls it?

  1. Find

    • Meaning
    • Words
    • Exact matches
    retry_request

  2. Connect

    send_requestretry_requestbackoff_delay send_requestretry_requestbackoff_delay

  3. Focus

    Selected source

    • client.rs fn send_request(request: Request)retry_request(request)
    • retry.rs fn retry_request(request: Request)let delay = backoff_delay(attempt);
    • backoff.rs fn backoff_delay(attempt: u32)BASE_DELAY * 2_u64.pow(attempt)

    Within your context budget

Find unfamiliar code

Describe what it does, even when you do not know its name.

Follow the connections

Trace parsed callers, imports, and implementations.

Keep context focused

Bring ranked source excerpts into a configurable token budget.

Explore code searchSee benchmark results

Your models.Your Workflows.

Select a reusable template and choose your models. Orek’s workflow engine deterministically controls step order, parallel execution, and how results pass forward. Models carry out each step; they don’t choose the workflow’s structure.

Illustrative workflow prompt %code_audit src/auth
Selected templatecode_auditBuilt-in example
Inputsrc/auth
Fixed steps. Defined dependencies.

The prompt selects this template. Orek’s workflow engine enforces its steps and dependencies once invoked.

Run the selected template

Map the code

Find files and relationships.

role:
explore
model:
Local model
Run together

Security review

Check security risks.

role:
reviewer
model:
Cloud model

Correctness review

Check logic and edge cases.

role:
reviewer
model:
Cloud model

Combine findings

One consolidated risk report.

role:
reviewer
model:
Cloud model

The report waits for both reviews. Execution follows the template; model-generated findings can vary.

Example: explore uses a local model; reviewer uses a cloud model.

Choose models through your conversation or agent profiles.

  • OpenAI
  • Anthropic
  • Google
  • Mistral
  • Other cloud providers
  • Local models
Explore workflowsChoose your models

Keep moving.Keep control.

With auto_review, Orek checks your policy first, then reviews eligible shell commands locally. Routine work can continue; flagged commands come back to you.

Illustrative shell review cargo test login

Policy first

Your rules decide what needs review.

  • AllowedRun
  • DeniedStop
  • Needs reviewContinue

This example needs review.

Local risk review

Shieldstral · on your device

  • Data leaving your device
  • Secrets and credentials
  • Lasting security changes
  • Destructive actions

Example outcomes

Continue automatically

Risk below your threshold

Ask for approval

Flagged command or unavailable review

Example result: low risk. The command can run automatically.

Illustrative outcomes. Actual decisions depend on your rules, model score, and configuration.

Requests that require a person go straight to you. Explicit policy denials stop the action.

File edits are allowed by default in auto_review; permission rules can still require approval.

Explore approvals and permissions

Less overhead. More momentum.

Orek Bench results

vs Antigravity

Gemini 3.8 Flash

95.4%fewer input tokens

3×faster completion

Input tokensmedian per run Lower is better
Orek 117,876
Antigravity 2,558,516
Elapsed timemedian per run Lower is better
Orek 35.849s
Antigravity 107.571s

vs Mistral Vibe

Mistral Medium 3.5

67.9%fewer input tokens

1.7×faster completion

Input tokensmedian per run Lower is better
Orek 483,763
Mistral Vibe 1,508,698
Elapsed timemedian per run Lower is better
Orek 181.053s
Mistral Vibe 310.154s

vs Codex

GPT-5.6 Luna

25.2%fewer input tokens

1.5×faster completion

Input tokensmedian per run Lower is better
Orek 272,375
Codex 364,350
Elapsed timemedian per run Lower is better
Orek 51.07s
Codex 77.52s

Orek Bench · Median results across runs

Fill timing follows elapsed time, scaled to 4 seconds for the slower run. Bar lengths show the labelled metric.

How we measured

Orek Bench is our internal benchmarking tool. It compares agents on matching tasks and measures token usage and completion time after setup.

We report the median across benchmark runs. These are selected historical results; performance varies by task, model, and configuration.

Token usage is accumulated across each run, with input and output tracked separately and cached input included. These charts show median input tokens and elapsed time.

Buy once. Own it.

Your Orek license never expires. Choose the models you want and make it your workflow.

Get License Download

Provider usage billed separately