harnessapi 0.2.1

Python framework to build streaming APIs and MCP tools from skill folders

Python Skill Framework for MCP Tools and Streaming APIs

Write a skill. Get an API. Get an MCP tool. Ship.

---

harnessapi is a Python framework that turns a skill folder into a streaming HTTP API and a Model Context Protocol (MCP) tool simultaneously — no routes, no decorators, no separate MCP server to maintain.

skills/summarize/
├── models.py    ← define input & output
├── handler.py   ← write your logic
└── skill.toml   ← name, description, tags, timeout

Drop the folder. Run the server. Your skill is live as an HTTP endpoint, an MCP tool, and in Swagger docs.

---

Contents

• Why harnessapi
• Quick start
• Try it instantly with uvx
• Streaming
• Every skill is an MCP tool
• Works with
• Scaffold in one command
• Example: streaming factorial
• Skill folder reference
• Hot-swap handlers at runtime
• Multi-tenancy & per-user skill variants
• Features
• Philosophy
• See also

---

Why harnessapi

Use harnessapi when you are:

• Building tools for Claude Desktop, Cursor, Copilot, or any MCP client
• Exposing Python functions as streaming API endpoints (Server-Sent Events)
• Converting an agentskills.io skill folder into a production API
• Shipping an LLM-powered microservice without FastAPI boilerplate
• Wrapping any Python function as an MCP tool in under a minute

---

Quick start

uv add harnessapi

skills/summarize/models.py

from harnessapi import SkillInput, SkillOutput

class Input(SkillInput):
    text: str
    max_length: int = 200

class Output(SkillOutput):
    summary: str

skills/summarize/handler.py

from .models import Input, Output

async def handle(input: Input) -> Output:
    return Output(summary=input.text[:input.max_length])

skills/summarize/skill.toml

[skill]
description  = "Summarize text to a target length"
is_mcp       = true
tags         = ["text"]
timeout_secs = 30

main.py

from pathlib import Path
from harnessapi import HarnessAPI

app = HarnessAPI(skills_dir=Path(__file__).parent / "skills")

harnessapi run

Your skill is live at three places simultaneously:

Endpoint	Details
POST /skills/summarize	HTTP endpoint — SSE streaming by default
GET /docs	Interactive OpenAPI / Swagger UI
http://localhost:8000/mcp	MCP server — ready for Claude, Cursor, Copilot

---

Try it instantly with uvx

No install needed. uvx runs harnessapi in an isolated environment:

# Scaffold a new project
uvx harnessapi init my-project

# Enter and run it
cd my-project
uvx harnessapi run

Then call your skill:

# Streaming (SSE — default)
curl -X POST http://localhost:8000/skills/greet \
  -H "Content-Type: application/json" \
  -d '{"name": "world"}'

# Plain JSON
curl -X POST http://localhost:8000/skills/greet \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{"name": "world"}'

{"message": "Hello, world! Welcome to harnessapi.", "length": 36}

---

Streaming — just use yield

Return a value for a single response. Use yield to stream chunks as they're produced. Same endpoint, same URL, no extra config.

# Non-streaming — return a value
async def handle(input: Input) -> Output:
    return Output(result=compute(input))

# Streaming — yield chunks
async def handle(input: Input):
    async for token in llm.stream(input.prompt):
        yield token

Clients receive standard Server-Sent Events (SSE):

event: chunk
data: The answer is

event: chunk
data: 42.

event: done
data:

Need plain JSON? Add Accept: application/json — harnessapi collects all chunks and returns them together. Same handler, zero changes.

---

Every skill is an MCP tool

Every skill folder is automatically registered as a Model Context Protocol (MCP) tool. No extra code required.

{
  "mcpServers": {
    "my-skills": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Add a skill folder → restart the server → it appears as an MCP tool. No registration. No schema maintenance.

---

Works with

Client	How to connect
Claude Desktop	Add http://localhost:8000/mcp as an MCP server in settings
Cursor	Add under MCP Servers in Cursor settings
Copilot / VS Code	Any MCP-compatible client works
agentskills.io	Drop-in compatible — existing skill folders work as-is
Any HTTP client	POST /skills/{name} — curl, httpx, fetch

---

Scaffold in one command

# New project with a sample greet skill
harnessapi init my-project

# Add API + MCP layer to an existing agentskills.io skill folder
harnessapi init --skill .agents/skills/summarize

# Convert an entire skills directory at once
harnessapi init --skills-dir .agents/skills

# Wrap a plain Python function as a skill
harnessapi init --function utils/compute.py --output skills

harnessapi is a compatible superset of the agentskills.io standard — existing skill folders with a SKILL.md are detected automatically.

---

Example: streaming factorial (SSE + MCP)

git clone https://github.com/edwinjosechittilappilly/harnessapi
cd harnessapi
uv sync
uv run uvicorn examples.factorial_app.main:app --reload

curl -X POST http://localhost:8000/skills/factorial \
  -H "Content-Type: application/json" \
  -d '{"n": 5}'

event: chunk
data: start: 1

event: chunk
data: 2: 2

event: chunk
data: 3: 6

event: chunk
data: 4: 24

event: chunk
data: 5: 120

event: done
data:

Or collect everything as JSON:

curl -X POST http://localhost:8000/skills/factorial \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{"n": 5}'

{"chunks": ["start: 1", "2: 2", "3: 6", "4: 24", "5: 120"]}

---

Skill folder reference

skills/
└── my_skill/
    ├── handler.py        ← required: your logic
    ├── models.py         ← required: Pydantic Input + Output
    ├── SKILL.md          ← optional: agentskills.io compatible metadata
    ├── skill.toml        ← optional: name, description, tags, timeout
    ├── defaults/
    │   └── input.json    ← optional: default values shown in /docs
    └── examples/
        └── 01.json       ← optional: {input, output} pairs for docs

skill.toml

[skill]
description  = "What this skill does"
is_mcp       = true      # set false to hide from MCP
tags         = ["nlp"]
timeout_secs = 30

---

Hot-swap handlers at runtime

Patch a running skill handler without restarting the server:

app = HarnessAPI(skills_dir="./skills", enable_edit_endpoints=True)

curl -X POST http://localhost:8000/skills/summarize/edit \
  -H "Content-Type: application/json" \
  -d '{"source_code": "async def handle(input):
    return Output(summary=input.text.upper())", "persist": true}'

Disabled by default. Add auth middleware before enabling in production.

---

Multi-tenancy & per-user skill variants

Add one parameter to HarnessAPI(...) and every skill gets per-user variants — same schema, different handler implementation, isolated routing. No restarts, no per-tenant route tables.

from harnessapi import HarnessAPI
from harnessapi.multitenancy import TenantBackend, SQLiteStorageBackend

backend = TenantBackend(
    tenant_extractor=lambda req: req.headers.get("X-Tenant-ID"),
    storage=SQLiteStorageBackend(path="./variants.db"),
)
app = HarnessAPI(skills_dir="./skills", tenant_backend=backend)

All existing endpoints continue to work. A /tenants/* management API is added automatically.

Agent workflow — clone → customize → test → promote:

# 1. Clone base skill source as starting point
curl -X POST /tenants/user-a/skills/greet/clone

# 2. Submit customized handler (validated before stored)
curl -X POST /tenants/user-a/skills/greet/customize \
  -d '{"source_code": "async def handle(input):\n    return Output(message=f\"Howdy, {input.name}!\")"}'

# 3. Test variant in isolation
curl -X POST /tenants/user-a/skills/greet/variants/{id}/run -d '{"name": "Alice"}'
# → {"message": "Howdy, Alice!"}

# 3b. Optional: set as preview — routes real tenant traffic without hard-promoting
curl -X POST /tenants/user-a/skills/greet/variants/{id}/preview

# 4. Promote — variant becomes the permanent active handler
curl -X POST /tenants/user-a/skills/greet/variants/{id}/promote

# 5. Route to variant automatically
curl -X POST /skills/greet -H "X-Tenant-ID: user-a" -d '{"name": "Alice"}'
# → {"message": "Howdy, Alice!"}   ← variant handler

curl -X POST /skills/greet -H "X-Tenant-ID: user-b" -d '{"name": "Bob"}'
# → {"message": "Hello, Bob!"}     ← base handler

Optional: per-user sandboxes for full process isolation:

from harnessapi.multitenancy import SandboxRegistry

backend = TenantBackend(
    ...,
    sandbox_registry=SandboxRegistry(),
    sandbox_provider="local_subprocess",  # or "docker" / "kubernetes"
)

Optional: admin MCP server so agents manage variants as MCP tools:

app = HarnessAPI(
    skills_dir="./skills",
    tenant_backend=backend,
    enable_admin_mcp=True,
    admin_mcp_auth=require_api_key,  # protect /admin-mcp in production
)
# Claude Desktop / Claude Code: add http://localhost:8000/admin-mcp as an MCP server
# Tools: clone_skill, customize_skill, preview_variant, promote_variant, run_variant, provision_sandbox, ...

See the multi-tenancy guide for the full API reference, sandbox providers, storage backends, and TenantBackend configuration.

---

Features

Feature	Details
HTTP endpoint	POST /skills/{name} for every skill, automatically
Streaming	SSE by default · JSON via Accept: application/json
MCP server	/mcp · all skills auto-registered as MCP tools
OpenAPI docs	/docs · full Swagger UI, zero config
Pydantic validation	Invalid input rejected before your handler runs
Timeouts	Per-skill timeout_secs in skill.toml
Hot-swap	Runtime handler replacement via opt-in edit endpoint
Multi-tenancy	Per-user skill variants · promote/demote · SQLite or custom storage
Preview status	Sandbox-tested variants route real tenant traffic before full promotion
Per-user sandboxes	Local subprocess, Docker, or Kubernetes — pluggable SandboxProvider
Admin MCP server	/admin-mcp · manage variants as MCP tools from Claude Desktop / Claude Code
LocalFileStorageBackend	File-per-variant JSON storage, no DB deps
agentskills.io	Drop-in compatible — existing skill folders just work
CLI scaffold	uvx harnessapi init · --skill · --skills-dir · --function

---

Philosophy

Most frameworks start with routes. Most agent frameworks start with tools. harnessapi starts with skills — the capability itself. The HTTP API and the MCP tool are consequences, not configuration.

Write the thing. Everything else follows.

---

See also

• FastMCP — MCP server framework harnessapi builds on
• FastAPI — the HTTP layer underneath
• agentskills.io — skill folder standard harnessapi is compatible with
• Model Context Protocol — the open protocol for agent tools
• Pydantic — data validation for skill inputs and outputs

---

Built on FastAPI · FastMCP · Pydantic · uv