autonomous execution runtime

deepstrain

modeldeepseek-r1 / v3
tools52 built-in
binary~4 MB (nuitka)
licenseHMAC-signed · BYOK
runtimepython ≥ 3.11
platformwin · linux · macos
get deepstrain docs
deepstrain — autonomous agentlive
deepstrain demo
$ pip install deepstrain
✓ deepstrain installed (4.1 MB compiled)
$ deepstrain eval "run tests, fix failures, repeat until green"
▸ reading codebase (37 files)...
▸ running pytest — 4 failures found
▸ patching src/auth/session.py
▸ all 47 tests pass · 3 turns
█
52
tools
~4 MB
binary size
3
commands to start
BYOK
deepseek api
// toolset
52 total
12
filesystem
9
execution
8
web
7
git
6
reasoning
5
system
4
memory

execution model

→ tools run in parallel where safe
→ autonomous loop until task complete
→ HMAC-signed execution graph
→ deterministic replay for auditability
→ keyring-backed secret storage
// architecture
compiled binary

Nuitka-compiled .pyd/.so via PyPI. No source exposure, no interpreter overhead.

local execution

Agents run on your machine. Your code and data never leave your hardware.

cryptographic licensing

HMAC-SHA256 signed activation. Edge revocation. Works offline after first verify.

zero framework

stdlib + requests only. No LangChain, no orchestration overhead, no abstractions.

deepseek byok

Bring your own API key. Requests go direct to DeepSeek — never proxied.

setup

01pip install deepstrain
02deepstrain chat
03deepstrain eval "ship this"
// the engine
// two separate LLMs — don't confuse them
① deepstrain's engine
always DeepSeek — cloud or local
This is what executes tasks: reads files,
runs tools, loops until done. You provide
the DeepSeek key (BYOK). Non-negotiable.
② your chat LLM
anything — Claude, GPT, Ollama, etc.
This is what you talk to. It understands
your intent, then calls deepstrain as an
MCP tool. deepstrain does the actual work.
you → [Claude / GPT / Ollama / any] ──MCP──▶ deepstrain ──DeepSeek──▶ your codebase

DeepSeek: cloud or local.

deepstrain's execution engine is always DeepSeek — either the cloud API (BYOK, ~$0.009/task) or a local DeepSeek model running via Ollama. Other models are not supported as the deepstrain engine.

cloud (default)platform.deepseek.com

API key in deepstrain configure · ~$0.009/task · no GPU needed

local deepseekOllama · deepseek-r1:7b/14b/32b

GPU on your machine · $0/task · fully air-gapped

local deepseek models (ollama)
deepseek-r1:1.5b~1.5 GB
deepseek-r1:7b~4.7 GB
deepseek-r1:14b~9 GB
deepseek-r1:32b~20 GB
cloud deepseek (default)
$ deepstrain configure
api_key: sk-•••••••••••••••• # your deepseek key
model: deepseek-v3
base_url: platform.deepseek.com/v1
$ strain chat "fix all test failures"
▸ all 47 tests pass · 3 turns · $0.009
local deepseek via ollama
$ ollama pull deepseek-r1:7b
✓ deepseek-r1:7b (4.7 GB) ready
$ deepstrain configure
api_key: none # no key for local
model: deepseek-r1:7b
base_url: http://localhost:11434/v1
$ strain chat "fix all test failures"
▸ engine: deepseek-r1:7b (local) · 0 external calls
▸ all 47 tests pass · 5 turns · $0.000
note: local deepseek = same architecture, more turns for same quality. cloud deepseek-v3 is faster and more capable for complex tasks.
// mcp — your chat LLM
use any frontend

talk to Claude.
it talks to deepstrain.

deepstrain exposes an MCP server. Any LLM that supports MCP tools — Claude, GPT-4, a local Ollama model, Continue.dev — can call deepstrain and have it execute tasks using DeepSeek internally. Your chat interface is your choice. The execution engine is always DeepSeek.

Claude DesktopMCP native — add deepstrain in settings.json
Ollama + Open WebUIlocal, free chat → deepstrain MCP tools
Continue.devIDE plugin → deepstrain as provider
GPT-4 / APIfunction calling → deepstrain tools
direct CLIstrain chat — standalone, no external LLM
cost breakdown (local chat example)
chat LLM (Ollama)$0.000
deepstrain subscription$9/mo
DeepSeek API (local deepseek-r1)$0.000
total / month$9 flat
Claude Desktop — mcp config
~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"deepstrain": {
"command": "deepstrain",
"args": ["mcp"]
}
}
}
result — Claude calls deepstrain
[Claude] "run the tests and fix any failures"
→ deepstrain_eval("run pytest, fix failures")
▸ engine: deepseek-v3 · reading codebase...
▸ running pytest — 4 failures found
▸ patching src/auth/session.py
▸ all 47 tests pass · 3 turns · $0.009
[Claude] "Done. Fixed 4 test failures — deepstrain
completed in 3 turns (~$0.009 via DeepSeek)."
start MCP server
$ deepstrain mcp
MCP server ready — stdio transport
51 tools registered
engine: deepseek-v3 (platform.deepseek.com)
waiting for tool calls...
// pricing

professional

$9/mo

cancel anytime

unlimited agent runs
byok deepseek api
52 tools
offline-capable
cloud activation + revocation
keyring secret storage
gui settings panel
get deepstrain

also from us

ATLASdeterministic code intelligence

Offline AST analysis. No LLM, no API calls, no telemetry. Dependency graphs, security scanning, hotspot detection.

bundle — save 20% →atlas →