Quickstart

The shortest useful Tally loop is local and account-free: adopt existing work, inspect the carried understanding, and continue it under another harness.

1. Open Tally

After installing Tally, run it with no subcommand:

$ tally

On a fresh machine this creates the internal owner identity silently, checks installed harnesses, and opens the TUI. It does not require signup, and it does not print a recovery phrase into your terminal history.

2. Adopt the agent you already use

First run leads with sessions Tally can discover. Today discovery/adoption is implemented for Codex; Claude Code and Pi discovery are next. Select a real session, inspect its task/workspace/source preview, give the durable agent a name, and accept.

The CLI equivalent is:

$ tally adopt --from codex

Adoption reads the native source without modifying it, creates a named agent atomically, and records the captured understanding in that agent’s encrypted log. It carries task, plan, relevant files, repository hints, provenance, and any agent character/memory that actually exists. It does not claim to copy a verbatim cross-harness transcript or hidden model state.

3. Inspect what follows

Open the adopted agent. Its page has five sections:

• Overview: current work, next steps, and last continuation.
• Understanding: character, memory, task, plan, acceptance, and relevant files.
• Continuations: the branch ledger.
• Sources: native-session provenance, encrypted-log provenance, and exact captured context.
• Settings: protection, relay status, self documents, and advanced lifecycle commands.

Missing fields collapse. Tally does not manufacture a decision, plan, or memory that the persisted sources do not contain.

4. Continue under another harness

Press c. Remembered harness/model defaults are filled in, but nothing launches yet. The sheet shows:

• the exact continuation sequence and workspace;
• which model provider receives the outbound context;
• that the Tally relay receives none of that context;
• what will be written before launch and captured after exit;
• the exact equivalent tally continue command.

Press p to inspect the full outbound text, w to choose another existing workspace, or h / m to change harness/model. Enter launches only after the preview exists. The previewed sequence and workspace are frozen into execution, so “latest” cannot drift after you approve it.

The CLI equivalent is explicit:

$ tally continue vision --harness pi --model openai-codex/gpt-5.5

On exit, Tally captures the new understanding as a child continuation. A direct TUI launch then shows a result sheet backed by the newly persisted event; memory reflection is reported separately when it is still pending.

5. Save the recovery key

After the agent has value, save its independent recovery phrase:

$ tally backup vision

The key restores identity and decrypts surviving encrypted copies. It is not itself a copy of the agent’s continuation log. Keep at least one device or other durable encrypted copy of that log.

6. Connect and sync when you want it

Everything above works locally. Hosted signup is optional and comes after local value:

$ tally init --email you@example.com

push and pull move encrypted log deltas. The relay cannot read agent content, but sync is not durable backup: hosted catch-up is bounded and your devices remain the durable holders. Named agents have independent keys and relay state; follow Multi-device sync rather than copying a base-home recipe blindly.

7. Use the advanced surfaces only when needed

# noninteractive continuation with a task
$ tally run vision --harness codex --model gpt-5.5 "explain the failing test"

# setup and integrity checks
$ tally doctor vision

# private MLS coordination
$ tally channel create pronoic/dev/coord

The command palette keeps backup, doctor, adapters, sync, signup, and rooms out of the primary loop without hiding them.

Where next

• The TUI: every section, launch boundary, and key
• Continue, start, and run: CLI launch details
• What syncs: content, metadata, and relay boundaries