Governance & Security
Simon treats LLMs as workers and applies deterministic rules: budgets halt execution, commands are scoped, and evidence gates prevent hallucinated deliveries.
Deterministic control plane
Simon governs every episode, so AI agents stay factual, budgeted, and verifiable.
Local-first infrastructure that enforces goals, safe command usage, and evidence collection. Slack, fuzziness, and hallucinations are handled by the runtime, not by hopeful prompting.
Journey spotlight
The full experience, not just logs
Plan, guard, verify, archive, and share—the recording mirrors the entire journey so the site communicates how Simon governs AI sessions.
Architecture
Built in Go with Cobra + Bubble Tea, storage on SQLite + the local filesystem, and a plugin-ready guard/coach layer.
Deployment
The README outlines GitHub Pages for marketing and optionally a Homebrew tap for distribution.
Recordings
Every run is captured end-to-end, from the first plan to the final evidence check, so reviewers can replay the exact execution trail.
Governance stack
The README highlights Coach, Guard, and Runtime as the fundamental pillars, with MCP proxy and Memory providing context and safety.
Coach
Defines a goal, definition of done, and evidence list before the agent starts making API calls.
Guard
Applies hard budgets, command scoping, and verification gates so every tool call stays within policy.
Runtime
Episodes advance with rolling summaries so the session never collapses even during long interactions.
MCP Proxy
Digest tool outputs internally to reduce noise and protect secrets before the agent sees them.
Memory
Vectorized experience recall archives completed sessions for future reference.
CLI first
Local-first Go binary with a Bubble Tea TUI keeps you in control on every host.
Session lifecycle
Every run progresses through planning, iteration, verification, and archiving. There are no loose ends once the definition of done is satisfied and the evidence list is confirmed.
Plan the mission
Write the goal, definition of done, and evidence list in the YAML task file.
Run under guard
The CLI enforces budgets, intercepts tool calls, and streams status updates at each iteration.
Verify completion
Every piece of evidence is checked before the task is marked done and memory is archived.
Record the impact
Asciinema captures the terminal with chapters so the run can be replayed and reviewed with full context.
Best-in-class recordings
Recordings that prove the work
Simon doesn't just log output; it captures the journey in a replayable narrative. Reviewers can understand intent, guardrails, and evidence without hunting through terminal history.
Full-fidelity capture
The CLI session is recorded as it runs, preserving every command, guard decision, and runtime status update.
Episode-aware playback
Recordings are anchored to the same episode cadence as the runtime, so reviewers can jump to meaningful checkpoints fast.
Shareable by default
Asciinema cast files play anywhere and keep stakeholders aligned without shipping raw logs or screen recordings.
Recording performance
Designed to keep reviewers in flow while preserving the full execution trail.
Typical runtime
20-60s
Fast iterations with real AI providers, depending on task complexity.
Playback speed
1.1x
Optimized for clarity while preserving real execution timing.
Stakeholder handoff
Cast + notes
Share the recording alongside the outcome.
Journey timeline
Each stage of the session is visible on the marketing site so visitors understand how Simon shepherds work from start to finish.
Plan
Define the goal, definition of done, and evidence list in the YAML task so the session starts with clarity.
Execute
Simon runs the iterations under guard, enforcing budgets and summarizing progress every episode.
Verify
Evidence is collected, checked, and only then does the runtime mark the task complete.
Archive
Session memory is vectorized so future runs learn from the exact same journey.
Share
The cast recording pairs with the evidence summary so stakeholders replay the full journey, not just logs.
How to get started
Install Simon via Homebrew or build from source, configure your preferred provider, then run a task with the CLI.
# Install via Homebrew (recommended) brew install felixgeelhaar/tap/simon # Or build from source git clone https://github.com/felixgeelhaar/simon.git cd simon && go build -o simon cmd/simon/main.go # Configure and run simon config set openai.api_key your-api-key simon run task.yaml --provider openai -i
The -i flag starts the interactive TUI for real-time execution visibility.
Session playback
The recording runs against real AI providers (OpenAI, Anthropic, Gemini, or Ollama), demonstrating actual latency, tool calls, and the full guard and verification workflow in production conditions.
Recording checklist
- Goal + definition of done snapshot
- Guard decisions and budget checkpoints
- Evidence validation captured in-line
Built for power developers
Control the execution, not just the prompt
Simon is local-first, configurable, and deterministic. The documentation and recording show exactly what the CLI outputs— no vague marketing fluff.