CLI

Use gligabao from the command line, for agents: remember, query, dataset, research, api, status and stop on a local daemon.

gligabao is also a client: remember, query, dataset, research, api, status and stop let an agent (or a person) use gligabao from the command line, without writing HTTP calls. The commands are thin: each one builds the JSON body of the matching REST endpoint, sends it and prints the response, so fields, validation and errors are exactly those of the API (remember, query, datasets). The commands live in the public cli package (cli/commands.go), so a program that embeds gligabao with cli.New gets them too (its own serve, with its hooks, is what the CLI starts as the local daemon).

The local daemon

The commands always talk to a server. By default that is a local daemon: gligabao serve --host 127.0.0.1 of this same binary, which the first command starts in the background (its own session, so it outlives the terminal) when nothing answers on the port. If a gligabao server already answers there, the command uses it and starts nothing. So there is one process on the database, however many agents call the CLI at once, and the remember manager still runs one remember at a time.

• Port: port of the config (default 7878). If another program holds it, the command fails and tells so.
• No authentication: the daemon listens on 127.0.0.1 only (serve --host, or host in the config), and the CLI sends no API key to it.
• Keys: the daemon is a normal server, so it needs ANTHROPIC_API_KEY and OPENAI_API_KEY (environment or config). If it cannot start, the command exits with code 5 and prints the end of the daemon’s log.
• Files: the log is daemon-<port>.log and the pid daemon-<port>.pid, in ~/.gligabao (GLIGABAO_HOME changes the directory). gligabao stop shuts the daemon down (the next command starts it again); gligabao status shows whether it runs, without starting it.

Where the config and the data are

• If the current directory has a gligabao.toml, or --config is given, the daemon runs as gligabao serve would: that config and the default paths of serve (build/gligabao.db, relative to the directory).
• Otherwise it uses a stable directory, so an agent gets the same memory whatever directory it runs in: ~/.gligabao/gligabao.toml if it exists (else the defaults) and the database ~/.gligabao/gligabao.db (the vectors and images next to it).

Another server

--server URL (or GLIGABAO_URL) sends the commands to that server instead; nothing is started, and GLIGABAO_API_KEY is sent as Authorization: Bearer. (Remote servers as targets are not finished: how the CLI authenticates to them is still to be designed.)

Commands

gligabao remember [TEXT | -] [-f FILE] [-d DATASET] [--kind episode] [--at TIME] [--session S]
                  [--web-search] [--series FILE.json] [--image FILE] [--model STAGE=NAME] [--reasoning STAGE=LEVEL] [--json BODY]
gligabao query QUESTION [-d DATASET] [--session S] [--evidence-only] [--geo-nearby] [--geo-radius KM]
                  [--model STAGE=NAME] [--reasoning STAGE=LEVEL] [--raw] [--json BODY]
gligabao dataset ls | create NAME [--description D] [--tag T] | rm NAME | graph NAME
gligabao research start TOPIC [--max-steps N] [--model M] [-d DATASET] [--wait [--raw]] [--json BODY]
gligabao research ls [--active] [--status running|completed|failed|stopped] [--limit N]
gligabao research get ID [--raw]
gligabao api METHOD PATH [JSON | @file | -]      # any other endpoint (timeseries, ontologies, ...)
gligabao status                                  # the target, and whether it runs
gligabao stop                                    # stop the local daemon

Global flags: --user (env GLIGABAO_USER, default local_user: the user_id of every request), --server, --config.

• remember takes the text as an argument, from --file, or from stdin with -; --image remembers an image instead. The flags map to the request fields (--at is reference_time, --series is timeseries.series, --model and --reasoning are models and reasoning, repeatable as STAGE=VALUE).
• --json (on remember and query) is the whole request body as JSON text, @file or - (stdin), so the exact REST body can be passed and a new request field never needs a new flag. The flags that are given override its fields; user_id defaults to --user only if the body has none.
• query --evidence-only is a retrieval-only call, see Query. query --raw prints just the answer, as text.
• research starts and follows research agents: start prints the run right away (-d picks the dataset of your knowledge base the agent queries and remembers into, main by default) (the agent goes on in the background, it costs money), --wait waits for it to end and prints the finished run (exit code 1 if it failed; --raw prints only the report), ls --active lists the agents that are working, and get ID shows a run with every state it went through (--raw: only the report).
• api sends the request as it is (it adds no user_id), for what has no command.

Output and exit codes

The response of the server is printed to stdout as indented JSON. An error is printed to stderr as {"error": "...", "status": 404} (status is the HTTP status, if there was one), and the exit code says what kind it is:

Code	Meaning
0	ok
1	any other error
2	bad usage, or the server answered 400
3	not found (404)
4	conflict (409)
5	not available: a 5xx (no LLM key, a failed LLM call, …), the server cannot be reached, or the daemon could not start

Not covered

Training, evals, the sidecar and the admin API are not commands (they need the sidecar and long-running event streams); use api for the plain request/response ones, or the server’s endpoints. Forgetting a single artifact or fact, and reading the status of a run, are not available yet.