Watchlight — Developer Edition

Govern an AI agent in five minutes. One install, zero infrastructure, same API as production.

Watchlight is an Agent Runtime Governance Control Plane — it puts a policy decision point in front of every action your AI agents take, authorizing tool calls and recording a tamper-evident, value-free audit trail.

The Developer Edition is the free, open front door to it. It runs the real authorization engine in-process, so you can add a governed ALLOW / DENY to your agent on your own laptop — no server, no database, no signup. It's for evaluating the model and shipping governed agents; the code you write here is the code you run in production — going to production is pointing the same code at the running control plane, not a rewrite.

For enterprises, Watchlight provides the wider Agent Runtime Governance Control Plane: signed, tamper-evident lineage; multi-tenant isolation; drift & anomaly detection → automatic quarantine; and fleet-wide revocation across every agent and environment. → watchlight.ai

How the pieces fit together

             your code · agents · tools
  ┌─────────────┬──────────────────┬──────────────┬─────────────┐
  │ @govern.tool│  LangGraph /     │  MCP client  │  custom app │
  │ (decorator) │  Pydantic AI /   │  (any tool)  │             │
  │             │  Claude Agent    │              │             │
  └──────┬──────┴────────┬─────────┴──────┬───────┴──────┬──────┘
         │               │                │              │
     watchlight   watchlight.<fw>   watchlight-mcp    watchlight
      (govern)   .governed_plugin   (PEP · MCP spec)     SDK
         └───────────────┴───────┬────────┴──────────────┘
                                 ▼
                 ┌───────────────────────────────┐
                 │  Watchlight engine · Cedar     │   the REAL engine,
                 │  in-process · zero infra       │   a compiled wheel
                 └───────────────┬───────────────┘
                PERMIT ─ forward │ ─ DENY  (blocked before execution)
                                 ▼
              value-free audit → .watchlight/audit.jsonl
                                 ▼
                 watchlight dev  ·  http://localhost:7000

  Production = the SAME code, pointed at the governed control plane
  (signed audit · multi-tenant · drift→quarantine · fleet revocation).

Runnable, self-contained examples for every one of these are in examples/ — start with examples/governed_research_agent.py.

Quickstart

The govern decorator, watchlight dev, and the framework + MCP integrations below all work today. Full guide: Developer Edition docs.

pip install watchlight

from watchlight import govern

@govern.tool(intent="research")
def web_search(query: str) -> str:
    ...

@govern.tool(intent="transfer")           # governed, but no policy permits it
def transfer_funds(to: str, amount: int) -> str:
    ...

$ python agent.py
watchlight: governing 'my-agent' (dev mode, in-process engine)
watchlight: ALLOW  read     tool/web_search
watchlight: DENY   execute  tool/transfer_funds     no matching policy

That DENY line — in your own terminal, in under five minutes, with no account — is the product.

Already using a framework? Govern it in-process

Bring your existing LangGraph, Pydantic AI, or Claude Agent SDK agent under governance with zero infrastructure — the same plugin you ship to production, wired to the in-process engine:

pip install 'watchlight[langgraph]'   # or [pydantic-ai], [claude-agent]

from watchlight.langgraph import governed_plugin   # .pydantic_ai / .claude_agent

plugin = governed_plugin("watchlight.policy.json")   # in-process, zero infra

async with await plugin.start_run("research-agent") as handle:
    if not await handle.authorize_action("read", "tool/web_search"):
        raise PermissionError("denied before it executed")
    ...  # your tool runs, every action governed + recorded to .watchlight/audit.jsonl

Going to production is one environment variable, not a rewrite — set WATCHLIGHT_APDP_URL and the identical code authorizes against a running policy service. Runnable examples for all three frameworks are in examples/.

Watch every decision live — watchlight dev

A zero-dependency local dashboard that tails your value-free audit trail and shows every governance decision as it happens — the ALLOWs, and the DENYs that stopped a tool before it ran.

watchlight dev            # → http://127.0.0.1:7000

Run your governed agent in another terminal and watch the decisions stream in. It shows only this process — fleet-wide lineage, signed audit, and drift→quarantine are the governed control plane (Enterprise).

Govern an MCP server

Put a policy decision point in front of any MCP server (spec 2026-07-28). Every tools/call is authorized in-process before it reaches the server — a denied call never executes.

pip install watchlight-mcp

import watchlight_mcp

watchlight_mcp.serve(
    listen_addr="127.0.0.1:9700",
    upstream_url="http://localhost:3000/mcp",   # the MCP server you're governing
    upstream_server="github",
    policy_files=["examples/mcp.policy.json"],
    audit_path=".watchlight/audit.jsonl",       # ← the file `watchlight dev` tails
)

Point your MCP client at http://127.0.0.1:9700/mcp instead of the server. A self-contained, self-demonstrating example (it fires an allowed and a denied call and proves the denied one never ran) is in examples/governed_mcp_server.py.

Watch the MCP decisions live

watchlight dev tails .watchlight/audit.jsonl — so give the PEP the same audit_path (above) and run the console beside it, from the same directory:

# terminal 1 — the governed MCP server, auditing to the file the console tails
python examples/governed_mcp_server.py

# terminal 2 — the live console
watchlight dev                # → http://127.0.0.1:7000

Every tools/call decision streams in — the tool, the upstream it fronts, and the reason a call was denied. (The example prints the exact watchlight dev --audit … command for its own audit file.)

What runs locally

Everything the Developer Edition removes is infrastructure, never a guarantee. Fail-closed semantics, engine-side attenuation, explicit scopes, and value-free audit are identical in every mode.

Progressive disclosure

Each level is one environment variable away from the next. Nothing is rewritten between levels.

• Level 0 — pip install watchlight. In-process engine, audit to stdout.
• Level 1 — watchlight dev. Adds a local dashboard (decisions, denials, scope tree, execution lineage).
• Level 2 — docker compose up. Real policy service + database; policies still from your local file.
• Level 3 — Production. The full governed platform.

Deploying to production? We're glad to help you get there — email sales@watchlight.ai and we'll help you plan the rollout.

Developer Edition vs Enterprise

The Developer Edition is the real engine — free, open, and running in-process so you can evaluate the entire authorization model on your laptop with zero infrastructure. Enterprise is the same code pointed at the governed control plane; it doesn't replace anything, it adds what a fleet in production needs:

• Signed, tamper-evident lineage & audit — every decision and lineage event cryptographically signed (KMS-backed), so the trail is court-defensible.
• Multi-tenant isolation + roll-up administration — tenant hierarchy, scoped admins, and a cross-tenant authorization matrix.
• Drift & anomaly detection → automatic quarantine — behavioural, goal-drift, and argument-shape detectors that quarantine a misbehaving agent before the next action.
• Full enforcement-effect taxonomy — beyond allow/deny: block, terminate, quarantine, sever-subtree, and revoke, enforced at runtime across the plane.
• Fleet-wide revocation & cross-environment governance — revoke authority across every agent at once, and govern dev, staging, and prod under one authority model (including sovereign / air-gapped deployment).
• Content / PII guardrails service and global execution-graph lineage with the full operator console.
• SSO / RBAC / enterprise audit, high availability, support, and SLAs.

You've outgrown the Developer Edition when…

• Compliance asks "prove who authorized this in production" → you need signed, tamper-evident lineage.
• You're governing more than one agent, or more than one environment → central policy lifecycle + the global execution graph.
• Security wants a misbehaving agent stopped before its next action → drift/anomaly detection → automatic quarantine.
• You need to revoke authority fleet-wide, not process-by-process.
• Procurement needs SSO, RBAC, HA, SLAs, or sovereign/air-gapped deployment.

Each of these is a governance guarantee a single in-process engine structurally cannot provide — it needs the control plane.

Migrating is one environment variable — never a rewrite. The tools you decorate, the policies you write, and the guarantees you rely on (fail-closed, engine-side attenuation, explicit scopes, value-free audit) are identical in every mode. Enterprise simply points the same code at a running plane.

The engine ships as a compiled wheel and the Developer Edition is a deliberate subset of the platform — the governed control plane (signing, multi-tenant, guardrails, drift, execution-graph) is the enterprise product, never bundled here.

→ Talk to us about Enterprise when you're ready for production.

License

The Developer-Edition SDK, the framework plugins, this repository, and the watchlight dev dashboard are Apache-2.0 — use, fork, and ship them freely. The authorization engine (watchlight-engine) and the MCP runtime (watchlight-mcp) ship as compiled wheels under the Watchlight Developer Edition license; they are free to use, including in production.

Want the engine source, an air-gapped build, or to govern a fleet in production? That's the Enterprise plane — email sales@watchlight.ai.