Press Start runtime, then describe a task in plain language. Each message becomes a governed run: you see every intent, the runtime's verdict, and each commit, with replay proof. The model can propose anything; only an authorized proposal ever runs, and risky steps pause for your approval. What it may touch is set once on the left and saved to this model.
Mind — reasoning in 3D
A pure neural map of the run — every node is a real runtime event (intent, proposal, grant, execution, commit, error) plus the context the run actually used (provider, tools, replay). It's the Intent → Proposal → Commit spine made visible; click any node to see exactly what it was, click a legend type to filter it.
The graph streams live as the agent works; the side panel reads the current turn top-to-bottom. Click a node (or a panel row) for details · drag to pan · scroll to zoom · double-click empty space to reset · click a legend type to filter.
Runs
Every message you send is a governed run — a trajectory of intents → proposals → commits recorded on the ledger. Each row shows what the agent did and how it ended; open one to see its full audit trail and replay it.
Providers
Available providers
Hook up any LLM. Pick a name above (or type any preset id) and Save — the runtime restarts using it. Three shapes:
- Claude: provider
anthropic+ your key. - Hosted (OpenAI, Groq, OpenRouter, DeepSeek, Gemini, …): pick the preset + paste its key. The model field is optional.
- Local / self-hosted (Ollama, LM Studio, vLLM, a gateway):
provider
ollama(oropenaifor a custom endpoint), set Base URL (Ollama defaults tohttp://localhost:11434/v1), model = the local model name, key optional.
Keys are read server-side only — they're injected into the local
runtime at startup, never sent to the model and never leave your
machine. Cognition gains no execution authority. With no working key the
runtime answers with the deterministic mock (real governance, no
real model). The CLI mirrors this: thymos providers lists
every preset, thymos use <name> sets one,
thymos doctor verifies the wiring.
Skills
A skill never grants authority — it narrows a run: effective
tools = requested ∩ allow-list, ceiling = AND, budget = min. Editing a
skill bumps its version, minting a fresh content-addressed id; each run
records the bound skill so replay can verify it. Bind one from the
Chat composer's dropdown. Mirrors the CLI's thymos skill.
Tools
These are the runtime's registered tool contracts, each tagged with the effect ceiling the governor enforces. Every call is checked against the writ (Pure ≤ Read ≤ Write ≤ External ≤ Irreversible) before it runs — a tool can never exceed its class.
＋ Add a custom tool — declare a governed tool (no code)
Audit
The receipts. Every message you send becomes a recorded run; this is the tamper-proof history of what actually happened — what the agent tried, what the runtime allowed or blocked, and what it changed. Search to find anything; leave it blank to browse all activity by day; click any row to open that run's full story.
Backups & state
A live snapshot of your local runtime and the ledger it writes to — the single hash-chained file that holds every governed action. Copy it anywhere to back up; integrity is re-verifiable by replay.
The ledger is a single content-addressed, hash-chained SQLite file.
A backup is a copy of that file; integrity is re-verifiable with
thymos replay <run> --verify or the runtime's
verify_integrity.
—
Copy that file anywhere to back it up; restore by pointing the
runtime at the copy. One-click backup/restore + a chain-head
integrity report is tracked in docs/rfcs/desktop-app.md §3.
Advanced — runtime inspection
The deeper controls and live truth of the runtime under this app. Normal chat never needs this tab; it's here when you want to see — or steer — exactly what the runtime is doing.
Runtime
Model & routing
Applies to new chat messages. Provider, key, and base URL are managed in the Providers tab.
Authority
Grants are the writ scope for new chats (left rail chips). Tools at
high risk (e.g. shell) always pause for your
approval — tune with THYMOS_APPROVE_RISK.
Working folder
The folder the agent may read, search, and edit. Without one, its file tools have nowhere to work. Changing it applies on the next runtime restart.
— none chosen —
Responses
Off by default. When on, replies appear as they're generated instead of all at once. Falls back automatically for providers that don't stream.
Ledger & debug
—
The single hash-chained file holding every governed action. Inspect any run in Audit (now with the full narrative: retries, rejections, grants, errors); verify with replay.