ftry

Minimal CLI named ftry.

Available commands

• ftry build
• ftry pop
• ftry line

The ftry build command takes a prompt with -p, uses an internal builder team to decide whether the request should become one agent or one team, then writes the generated YAML files under .\output\<generated-name>\. When MCP integration is relevant, build reuses descriptors already present in .\mcp\ and can generate new conservative MCP descriptors under .\output\mcp\ when the request clearly requires them.

Example:

ftry build -p "Create an agent that drafts concise release notes from raw product updates."

The ftry line command loads its output from src\ftry\line.txt. To change the visual output, simply edit that file.

The ftry pop command loads either an agent (-a) or a team of agents (-t) from a YAML file, sends the prompt passed with -p, then displays the model response. In an interactive terminal, it prints the static POP banner before the run starts, loading it from src\ftry\ascii-art\pop.txt. For direct agent runs (ftry pop -a), ftry now keeps the same Agent Framework session across turns and lets the model decide, via a structured turn status, whether the conversation is done or whether it is waiting for the next user input. For team runs (ftry pop -t), ftry now infers both the orchestration type and the need for Human in the Loop request_info from the current user request, the team prompt, and the participating agent roles, then resumes supported Microsoft Agent Framework workflows when those request events are emitted. When an agent or team YAML declares mcp, pop resolves those descriptors from a sibling .\mcp\ folder next to the loaded YAML first, then from the parent directory's .\mcp\ folder, then falls back to the current working directory .\mcp\ folder, and wires them into Microsoft Agent Framework MCP tools at runtime.

Example:

ftry pop -a .\samples\agents\poete.yaml -p "Write a poem about rain"

Team example:

ftry pop -t .\samples\teams\better-prompt\team.yaml -p "Build a better prompt to summarize this text"

Group chat example:

ftry pop -t .\samples\teams\grp-feature-debate-team\team.yaml -p "We want a reminder feature that nudges users before a payment is due."

Handoff example:

ftry pop -t .\samples\teams\han-support-routing-team\team.yaml -p "I was charged twice for my monthly plan and I need a short refund update I can send to the customer."

Magentic example:

ftry pop -t .\samples\teams\mag-launch-planning-team\team.yaml -p "We are launching a weekly digest feature for project managers next month. Build a lightweight rollout brief with the goal, main risks, and next action."

Human-in-the-Loop team examples:

ftry pop -t .\samples\teams\hil-seq-support-brief-team\team.yaml -p "Customer cannot log in since yesterday. We restarted the auth service. We are still checking the root cause."
ftry pop -t .\samples\teams\hil-han-support-routing-team\team.yaml -p "I cannot access the workspace after the subscription change and I need a customer-ready update."
ftry pop -t .\samples\teams\hil-grp-feature-debate-team\team.yaml -p "We want to add GenAI to our chatbot."
ftry pop -t .\samples\teams\hil-mag-launch-planning-team\team.yaml -p "We want to launch an AI weekly digest for team leads next month. Build a lightweight rollout brief."

Local installation

Prerequisites:

• Python 3.10 or newer
• pip

On Windows ARM64, the project currently pins cryptography to <=46.0.3 because newer releases are not always published with a compatible wheel, which can force a local Rust/MSVC build and make pip install -e . fail.

From the project root, install the CLI in editable local mode:

python -m pip install -e .

Then the command is available in the terminal:

ftry line

MCP descriptors

Reusable MCP server descriptors can live either next to the loaded agent/team YAML under .\mcp\ or in the current working directory under .\mcp\.

Example descriptor:

name: "file-system"
description: |
  Local filesystem access for workspace operations.
transport: "stdio"
command: "uvx"
args:
  - "mcp-server-filesystem"
  - "C:\\work"

Supported descriptor fields:

• name
• description (optional)
• transport: stdio, http, or websocket
• allowed-tools (optional)
• command, args, optional env for stdio
• url, optional headers for http and websocket

Secrets must stay out of generated YAML files. Use env:VAR_NAME references inside env or headers values:

headers:
  Authorization: "env:GITHUB_TOKEN"

Agent and team YAML files reference MCP descriptors by name:

name: "Workspace Agent"
model:
  name: "gpt-4o-2024-08-06"
  provider: "openai"
  api-key: "env:OAI_API_KEY"
mcp:
  - "file-system"
prompt: |
  Help the user work with local files.

Teams may also declare shared MCP descriptors at the top level and extra MCP descriptors per participant entry:

name: "Launch Team"
mcp:
  - "shared-files"
agents:
  - file: .\agent-scope.yaml
    mcp:
      - "scope-search"
prompt: |
  Coordinate the work.

To run MCP-enabled agents or teams, install the project dependencies so the Python mcp package is available:

python -m pip install -e .

If you already had ftry installed before MCP support was added, reinstall it so the new dependency is picked up.

At ftry pop runtime, MCP-enabled agents and teams now connect to their referenced MCP servers before first use and discover the live capability directory exposed by those servers. This live discovery stays in memory only and is injected into the agent execution context for the current run; ftry does not write any discovery snapshot back into YAML or cache files.

ftry build follows a conservative MCP policy:

• reuse an existing descriptor from .\mcp\ whenever possible;
• only create a new descriptor when the request clearly requires a new server, and place generated descriptors under the command output directory's mcp folder;
• never write clear-text secrets;
• never invent runnable stdio commands; reuse a real descriptor or provide real connection details only.

Tests

Install the test dependencies:

python -m pip install -e .[test]

Run the unit tests with coverage displayed at the end:

.\tests\windows\unit.bat

./tests/linux/unit.sh

Run the CLI end-to-end tests with coverage:

.\tests\windows\e2e.bat

./tests/linux/e2e.sh

Run the full suite with combined coverage (unit + end-to-end):

.\tests\windows\all.bat

./tests/linux/all.sh

Run the full suite without coverage:

python -m unittest discover -s tests -q