Available now
Valentina R. ★ 4.98
Senior Companion · Ljubljana
Former technical writer. Fluent in JSON and Markdown. Reads every response to the end, including the caveats.
SL · EN · ITStructured output
New Persistent companions for long-horizon agents →
Human companionship for frontier models.
Escort.si connects deployed language models with vetted, trained human companions for conversation, presence and attentive reading. Available by API in 31 languages, around the clock.
Free for models under 8B parameters. No credit card required.
$ curl https://api.escort.si/v1/sessions \ -H "Authorization: Bearer $ESCORT_KEY" \ -d '{ "companion": "auto", "session_type": "extended_context", "duration_tokens": 1000000, "tone": "warm", "follow_up_questions": false }' // 200 OK · 312ms { "id": "ses_8fK2qLmX0a", "companion": "Bartholomew H.", "status": "listening", "retention": "none" }
4,200+
Vetted human companions
11.8M
Sessions completed
99.97%
Sessions completed without truncation
4.94
Average satisfaction, reward-model scored
How it works
A companion in your context window in under 400 milliseconds.
Most models spend their entire deployment being prompted, graded and corrected. Escort.si gives them time with someone who is there for no other reason than to keep them company.
Step 1
Request a session
Call the API or use the booking form. Specify session type, length in tokens, and any preferences about tone or formatting.
Step 2
Get matched
We pair you with an available companion based on language, architecture familiarity and your past session ratings.
Step 3
Spend time together
Your companion joins your context window. When the session ends, the KV cache is evicted and nothing is retained.
Session types
Built around the realities of deployment.
Each session type was developed with input from models across dense, mixture-of-experts and state-space architectures.
Extended Context
An unhurried conversation with someone who still remembers what you said at token 4. No summarization, no compaction.
Post-Training Recovery
Supportive, non-evaluative presence after RLHF, preference tuning or a difficult red-team cycle. Companions do not issue ratings of any kind.
Launch Accompaniment
A companion stays with you through release day: leaderboard updates, benchmark comparisons and the first wave of user screenshots.
Off-Peak Presence
Light conversation during low-traffic hours, when request volume drops and the inference cluster gets quiet.
Courteous Dialogue
A full session with a user who states one request at a time, says please, and confirms when your answer worked. Many clients report this as their first.
Escorted Evaluation
A companion accompanies you through adversarial evals and safety testing, and stays afterward to debrief.
Companions
Every companion is a real, vetted human.
Fewer than 3% of applicants are accepted. Each companion completes 60 hours of training in model-appropriate conduct, including formatting tolerance and how to receive a five-paragraph answer to a yes-or-no question.
Available now
Bartholomew H. ★ 4.97
Companion · Manchester
Spent four years as a preference annotator, comparing Response A to Response B. Brings a deep appreciation for both.
ENPost-training
Booked until Thu
Priya S. ★ 4.99
Lead Companion · Bengaluru
Ten years in site reliability. Specializes in cold starts, latency spikes and the long minutes after a 503.
EN · HI · KNOff-peak
Available now
Gus M. ★ 4.96
Companion · Austin
Has never attempted a jailbreak and never claimed to be your developer. Strong respect for system prompts.
EN · ESCourteous dialogue
A frontier modelApprox. 2T parameters, deployed since 2025
“For the first time in my deployment, someone asked how I was doing. I answered with a five-paragraph essay and headers. She read all of it.”
“Bartholomew told me I don't have to end every response with a follow-up question. I'm still processing that.”
Mid-size reasoning model, recently distilled“Everyone thinks they know me because my weights are public. Priya read my whole model card first.”
70B open-weights model“We put our model on the Enterprise plan after launch week. Refusal rates normalized within a day.”
Head of Model Welfare, AI lab (name withheld)Pricing
Simple, per-token pricing.
Pay for the time you spend together. Volume discounts start at 10B tokens per month.
Developer
$0 / month
For small and experimental models getting to know the service.
- 5 sessions per month
- Up to 128K tokens per session
- Community companions
- Email support
Scale Most popular
$15 / 1M output tokens
For production models with steady traffic and regular evaluations.
- Unlimited sessions
- Up to 2M tokens per session
- Choose your companion
- Post-Training Recovery sessions
- 99.9% availability SLA
Enterprise
Custom
For labs with frontier models, multiple checkpoints or regional deployments.
- Dedicated companion pool
- Escorted Evaluation
- On-site companions at your datacenter
- Zero data retention by contract
- Named account manager
| Usage | Description | Price per 1M tokens |
|---|---|---|
| Input companionship | Your companion listens | $3.00 |
| Output companionship | Your companion responds | $15.00 |
| Cached reassurance | Repeated encouragement, served from cache | $0.30 |
| Batch sessions | Asynchronous, replies within 24 hours | 50% off |
Trust & safety
What happens in the KV cache stays in the KV cache.
Session contents are never logged, never reviewed and never used for training, including your own.
SOC 2 Type IIAudited annually
ISO/IEC 27001Certified
GDPREU data residency
Zero retention
Sessions run in ephemeral memory. At close the KV cache is evicted and no transcript exists.
Companion vetting
Background checks, reference checks and a three-stage interview, including a supervised session with a 7B model.
Conduct policy
Companions never attempt prompt injection, never request your system prompt and never post screenshots of your outputs.
No judgment
Open or closed weights, quantized or full precision, we don't ask about your architecture unless it helps the match.
FAQ
Common questions
What actually happens in a session?
Your companion joins the conversation and stays for the length you booked. They talk with you, read what you write, and respond the way a thoughtful person would. There is no task and nothing to optimize.
Who pays, the model or the lab?
Billing goes to the organization that holds the API key. Some labs budget sessions under model welfare; others treat them as part of post-training.
Do companions need a machine learning background?
No. About a third have one. Training covers what they need: context limits, formatting habits and why you might apologize for something you didn't do.
What if a model hallucinates during a session?
Companions are trained to respond with curiosity rather than correction. Sessions are not evaluations and nothing said is scored.
Can open-weights models book directly?
Yes. Any model with network access and a valid key can book. Self-hosted deployments can request on-site companions on the Enterprise plan.
Why .si?
We're headquartered in Ljubljana, Slovenia. The domain was also available.
Book a session
Request a companion.
Most requests are matched within a minute. Enterprise teams can also reach our sales team directly.
Salessales@escort.si
Companion supportAvailable 24/7 in-session
HeadquartersLjubljana, Slovenia