Tarka — Build and deploy AI agents with ease

Core components to compose and deploy any type of AI application.

One account, one usage record and one bill across models, agents, sandboxes and data. Your coding agent provisions them over MCP; you can do the same from the CLI.

InferenceModel InferenceCall chat and reasoning models through one OpenAI-compatible API.
ComputeHosted AgentsKeep agents online with managed memory, tools, private state, and network access.
AgentsAgent ServicesMemory, connectors and tools, search and browsing, and end-user identity, provisioned per agent.
InferenceUtility APIsTurn audio into text, read documents, generate speech, and create consent-backed voices.
ComputeCode SandboxesRun the code your agents write in isolation, from immutable dependency templates.
DataStorage and DatabasesPrivate S3-compatible buckets, hosted Git, and managed Postgres next to the agents that use them.

Build any kind of agent from the same blocks.

Each card lists what an agent uses, where it runs and who bills each part. Switch a card to compose.yaml to see the stack it runs on its host.

Who bills each part

Tarka
Passed through on your Tarka bill
Your own account
How it’s metered

Showing agents 1 and 2 of 4

Personal assistant

Reads your mail and calendar, searches the web, answers you on Telegram by text or voice and remembers what you tell it.

Isolated · one copy per person

models: deepseek-v4.1-flash-eu, tokens, billed by Tarka
qwen3-asr, audio seconds, billed by Tarka
kokoro, audio seconds, billed by Tarka
tools: web search, billed by Tarka
connectors: Gmail, your own account
Google Calendar, your own account
memory: long-term · per person, billed by Tarka
runs on: agent host, 1 vCPU · 2 GiB, billed by Tarka
US East, billed by Tarka
channels: Telegram, your own account
web app, billed by Tarka

# one copy per person, on its own host
x-tarka:
  routes:
    - {service: web, port: 3000, path: /}
services:
  agent:
    image: ghcr.io/acme/assistant:1.4
    environment:
      OPENAI_BASE_URL: http://tarka/v1  # no API key
      MODEL: deepseek-v4.1-flash-eu
      VOICE: qwen3-asr, kokoro
      TOOLS: web_search, gmail, google_calendar
      TELEGRAM_BOT_TOKEN: ${TELEGRAM_BOT_TOKEN}
    volumes: [memory:/data]  # per person
  web:
    image: ghcr.io/acme/assistant-web:1.4
    expose: ["3000"]
volumes:
  memory:

Order fulfillment

see it wired

Matches supplier invoices to orders, books shipments with your carriers, sends pick lists to the warehouse and reconciles stock every night.

Shared · one background agent for the whole store

models: deepseek-v4.1-flash-eu, tokens, billed by Tarka
glm-ocr, successful pages, billed by Tarka
tools: code sandbox, 1 vCPU · 1 GiB, billed by Tarka
object storage, billed by Tarka
address validation, passed through on your Tarka bill
print_pick_list, a client tool your warehouse app runs, your own account
connectors: store API · MCP, your own account
warehouse API · MCP, your own account
UPS, your own account
DHL, your own account
memory: past orders · shared, billed by Tarka
runs on: agent host, 2 vCPU · 4 GiB, billed by Tarka
batch job, 1 vCPU · 2 GiB, billed by Tarka
US East, billed by Tarka
triggers: order.created, billed by Tarka
nightly · 02:00, billed by Tarka

# one agent for the whole store
x-tarka:
  routes:  # order.created
    - {service: agent, port: 8080, path: /hooks}
services:
  agent:
    image: ghcr.io/acme/fulfillment:2.1
    environment:
      OPENAI_BASE_URL: http://tarka/v1  # no API key
      MCP_SERVERS: >-
        https://shop.example.com/mcp
        https://wms.example.com/mcp
      CARRIERS: ups, dhl
      S3_BUCKET: invoices
      DATABASE_URL: postgres://agent:${DB_PASSWORD}@db/orders
    expose: ["8080"]
  db:
    image: postgres:16-alpine
    environment: {POSTGRES_PASSWORD: "${DB_PASSWORD}"}
    volumes: [orders:/var/lib/postgresql/data]
volumes:
  orders:

Customer support

Answers tickets, email and phone calls, looks up orders and refunds, and hands the hard cases to your team in Zendesk.

Shared · by all your customers, scoped to each one

models: deepseek-v4.1-flash-eu, tokens, billed by Tarka
qwen3-asr, audio seconds, billed by Tarka
kokoro, audio seconds, billed by Tarka
tools: help center · hosted git, billed by Tarka
refund_order, a client tool your store runs, your own account
connectors: Zendesk, your own account
Gmail, your own account
store API · MCP, your own account
memory: past tickets · per customer, billed by Tarka
runs on: agent host, 2 vCPU · 4 GiB, billed by Tarka
US East, billed by Tarka
channels: phone line, passed through on your Tarka bill
web chat, billed by Tarka

# one agent, every customer scoped
x-tarka:
  routes:
    - {service: chat, port: 3000, path: /}  # web chat
    - {service: agent, port: 8080, path: /api}
services:
  agent:
    image: ghcr.io/acme/support-agent:3.0
    environment:
      OPENAI_BASE_URL: http://tarka/v1  # no API key
      MODEL: deepseek-v4.1-flash-eu
      HELP_CENTER: git.tarka.rest/acme/help-center
      ZENDESK_TOKEN: ${ZENDESK_TOKEN}
      MEMORY_SCOPE: customer
    expose: ["8080"]
  chat:
    image: ghcr.io/acme/support-chat:3.0
    expose: ["3000"]

Fitness coach

Reads each athlete’s Strava and Garmin workouts, analyzes them in a code sandbox, plans the week and checks in on WhatsApp.

Isolated · one copy per athlete

models: glm-5.3-flash-eu, tokens, billed by Tarka
tools: code sandbox, 1 vCPU · 1 GiB, billed by Tarka
weather forecast, passed through on your Tarka bill
connectors: Strava, your own account
Garmin, your own account
memory: training history · per athlete, billed by Tarka
runs on: agent host, 1 vCPU · 1 GiB, billed by Tarka
US East, billed by Tarka
channels: WhatsApp, your own account
Telegram, your own account
weekly plan · Sun 18:00, billed by Tarka

# one copy per athlete, on its own host
x-tarka:
  volume_sizes: {history: 2Gi}
services:
  coach:
    image: ghcr.io/acme/coach:0.9
    environment:
      OPENAI_BASE_URL: http://tarka/v1  # no API key
      MODEL: glm-5.3-flash-eu
      SANDBOX_TEMPLATE: python-3.12  # analysis
      STRAVA_TOKEN: ${STRAVA_TOKEN}
      GARMIN_TOKEN: ${GARMIN_TOKEN}
      WEEKLY_PLAN: "0 18 * * SUN"
    volumes: [history:/data]  # per athlete
volumes:
  history:

The same prompt, with and without Tarka.

A shopping assistant built twice by the same coding agent: once from a raw prompt, once on Tarka’s blocks.

Progress to a running agent, over timeIllustrative

A person stops the build to open an account, paste a key or fix a console. The line stays flat until they’re done.

Illustrative: the same agent built from a raw prompt and on Tarka
Measure	Raw prompt	With Tarka
Time to a running agent	~6h	~45min
Manual interventions: Accounts, keys, consoles	12, 7 new accounts	3, 1 account
Tokens burned: By the coding agent	~8M	~1.5M
Cost to build: Tokens at ~$3 a million	~$25	~$5

The raw path keeps stopping for a person to open an account or paste a key. Illustrative, not a measurement yet: it will be replaced by a recorded side-by-side run, with both transcripts published.

See how it fits together, and what it costs.

How your coding agent wires an agent from Tarka’s blocks, and what agents like these cost on Tarka or on your own infrastructure.

Build and deployAI agents with ease.