Bring your own harness

An agent can run on its own harness instead of Lagrange, Claude Code or Codex. The harness is a command that runs in the agent's sandbox and talks to the platform over stdin and stdout, one JSON object per line. Any language works.

# lgr.yaml
name: support-bot
harness:
  command: python3
  args: [harness/main.py]
files: [harness]
setup: pip install --user -r harness/requirements.txt
secrets: [ANTHROPIC_API_KEY]

lgr deploy uploads harness/ with the rest of the agent's files. Each run then unpacks them, runs setup, and starts the command in the working directory. Runs show custom as their harness.

#Environment

The process gets the run's environment: HOME, PATH, the agent's secrets, and the egress proxy variables when the platform runs one. It does not get the platform's model keys: to call a model, set your own key as a project secret (ANTHROPIC_API_KEY, OPENAI_API_KEY, ...) and list it in secrets. Runs on your own harness are billed for sandbox time only.

#Messages from the platform (stdin)

The first line is the run:

{
  "type": "run",
  "harness": "custom",
  "approvals": false,
  "request": {
    "task": "Summarise the open tickets",
    "instructions": "Answer from the ticket history.",
    "workdir": "/workspace",
    "stateDir": "/lgr/state",
    "model": "claude-sonnet-4-6",
    "tools": ["Bash", "Read"],
    "mcpServers": [{ "name": "linear", "url": "https://mcp.linear.app/mcp", "token": "..." }],
    "skillsDir": "/home/lgr/.lgr/agent/skills",
    "limits": { "max_minutes": 30 },
    "resumeId": "the resumeId your harness returned last time"
  }
}

Only task and workdir are always there. stateDir persists on instances and profiles: keep conversation state there and return its id as resumeId to support continue. skillsDir holds the agent's skills as <name>/SKILL.md. Local MCP servers come with command, args and env (secrets filled in) instead of url and token.

Later lines are:

• {"type": "approval", "id": "a1", "approved": true}: the answer to an approval you asked for.
• {"type": "cancel"}: stop and send a cancelled result. The process is killed 10 seconds later.

Stdin closes once your result is received.

#Messages to the platform (stdout)

Write one JSON object per line. Anything else on stdout is ignored; use stderr for logs (its last lines become the error if the process exits without a result).

Events appear on the run's timeline and trace:

{"type": "event", "event": {"type": "log", "data": {"level": "info", "message": "Loaded 12 tickets"}}}
{"type": "event", "event": {"type": "model.call.started", "span_id": "m1", "data": {"model": "claude-sonnet-4-6", "input": "..."}}}
{"type": "event", "event": {"type": "model.call.completed", "span_id": "m1", "data": {"output": "...", "tokens_in": 1200, "tokens_out": 80}}}
{"type": "event", "event": {"type": "tool.call.started", "span_id": "t1", "parent_span_id": "m1", "data": {"tool": "search", "input": {"q": "refund"}}}}
{"type": "event", "event": {"type": "tool.call.completed", "span_id": "t1", "data": {"tool": "search", "output": "3 results"}}}

span_id pairs a started event with its completed one; parent_span_id nests spans in the trace. Other event types (run lifecycle, approvals, artifacts) belong to the platform and are dropped.

To ask a person before a tool call (only when the run message has "approvals": true):

{"type": "approval", "id": "a1", "request": {"tool": "send_email", "input": {"to": "..."}}}

Finish with exactly one result:

{"type": "result", "result": {"status": "succeeded", "result": "3 tickets need a reply", "model": "claude-sonnet-4-6", "usage": {"tokens_in": 5200, "tokens_out": 640}, "resumeId": "conv_42"}}

status is succeeded, failed (with error) or cancelled. Files the harness writes in the working directory become the run's artifacts.

#A minimal harness

import json, sys

def send(msg):
    print(json.dumps(msg), flush=True)

run = json.loads(sys.stdin.readline())["request"]
send({"type": "event", "event": {"type": "log", "data": {"message": f"working on: {run['task']}"}}})
# ... call your model, use tools, write files ...
send({"type": "result", "result": {"status": "succeeded", "result": "done", "usage": {}}})