DeepChat CLI
Use the bundled deepchat command to ask the running DeepChat main process to perform supported
operations. The main process remains the sole owner of providers, credentials, Skills, MCP servers,
artifacts, Agent runs, and approvals.
Command rules
- Every command must begin exactly with
deepchat <domain> <verb>. Put --json, --jsonl,
--timeout, and all domain options after the domain and verb.
- Execute one standalone command per
exec call. Do not use pipes, redirection, command separators,
command substitution, environment assignments, or shell wrappers around deepchat.
- Quote every user-controlled argument for the current shell. Never interpolate untrusted text into
an unquoted command.
- Prefer
--json for one result and --jsonl for streaming or benchmark collection. Use text mode
only when its output will be returned directly to the user.
- Do not inspect authentication environment variables or DeepChat's local descriptor. Authorization
is injected only after the command has passed the normal shell permission check.
- A shell approval authorizes command execution. Sensitive mutations can additionally pause for a
renderer approval; wait for that decision and never attempt to manufacture confirmation data.
- Use
deepchat help or deepchat <domain> <verb> --help only when the options below are
insufficient. Do not probe undocumented routes.
Agent file and recursion boundaries
- Agent callers may consume a DeepChat-owned artifact with
--artifact <id> and inspect metadata
with artifact describe.
- Do not use
--file, --out, --overwrite, artifact get, or artifact delete. Agent callers
cannot upload arbitrary local bytes, download artifact bytes, or choose output paths.
- Do not call
agent run or run watch. An Agent cannot recursively create a detached Agent run,
and waiting on its own currently executing run would deadlock it. Use run get for a nonblocking
snapshot or run cancel to request cancellation.
- Generated media remains in DeepChat's artifact spool. Return the artifact metadata or ID so the
application can render or reuse it.
Discovery and model calls
deepchat system status --json
deepchat system capabilities --json
deepchat system doctor --json
deepchat provider list --enabled-only --json
deepchat model list --provider <provider-id> --json
deepchat model config-get --provider <provider-id> --model <model-id> --json
deepchat model invoke --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
Always discover provider and model IDs rather than guessing them. model invoke is a raw provider
call: it does not create a chat session, run tools, or start an Agent loop.
Media, transcription, and OCR
deepchat image generate --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
deepchat video generate --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
deepchat audio speak --provider <provider-id> --model <model-id> --text <quoted-text> --jsonl
deepchat audio transcribe --provider <provider-id> --model <model-id> --artifact <artifact-id> --json
deepchat ocr status --json
deepchat ocr extract --artifact <artifact-id> --json
deepchat artifact describe --id <artifact-id> --json
Use the provider/model lists to choose a compatible runtime. OCR is local and does not require a
provider. OCR text is returned inline and is not written to the artifact spool.
Public configuration and management
Read-only operations:
deepchat settings get --json
deepchat skill list --json
deepchat mcp list --json
Agent callers may request renderer approval for preference-only settings, query-free HTTPS Skill
installation, and adding a new disabled HTTPS remote MCP configuration. Only perform one when it
directly satisfies the user's request:
deepchat settings set --key <public-key> --value <json-scalar> --json
deepchat skill install --url <https-url> --json
deepchat mcp add --name <server-name> --stdin --json
The Agent setting allowlist is limited to presentation preferences such as font size/family,
artifact effects, auto-scroll, notifications, and copy-with-reasoning. Agent Skill URLs cannot carry
credentials, query parameters, or fragments. The main process classifies MCP input before approval
and rejects stdio commands, non-HTTPS endpoints, headers, authorization bindings, or configurations
too large to review safely. Provider/model configuration, credential writes, local Skill archives,
Skill enable/disable/removal, MCP update/runtime control/removal, and every destructive operation
require the DeepChat UI or a human terminal.
Benchmark discipline
- Pin provider/model IDs and pass per-invocation options; do not mutate global defaults to prepare a
benchmark.
- Record structured output, exit status, wall time, and errors. Preserve failed samples.
- For OCR, distinguish cache hit, cache miss with warm runtime, cold runtime after app restart, and
offline availability.
ocr clear-cache initializes the resource graph but does not start the OCR
helper, so classify the next extraction from its reported pre-extraction runtime state.
- Run samples sequentially unless the benchmark explicitly measures concurrency; Agent compute is
rate-limited and bounded by the main process.
1---2name: deepchat-cli3description: Use DeepChat's bundled CLI control plane for model inference, image/video/speech generation, transcription, OCR, artifact inspection, public configuration, Skills, and MCP operations. Activate when a user asks to invoke DeepChat capabilities that are not already exposed as a more specific tool, compare models, run a benchmark, inspect DeepChat runtime state, or manage DeepChat through the CLI.4---5
6# DeepChat CLI
7
8Use the bundled `deepchat` command to ask the running DeepChat main process to perform supported
9operations. The main process remains the sole owner of providers, credentials, Skills, MCP servers,
10artifacts, Agent runs, and approvals.
11
12## Command rules
13
14- Every command must begin exactly with `deepchat <domain> <verb>`. Put `--json`, `--jsonl`,
15 `--timeout`, and all domain options after the domain and verb.
16- Execute one standalone command per `exec` call. Do not use pipes, redirection, command separators,
17 command substitution, environment assignments, or shell wrappers around `deepchat`.
18- Quote every user-controlled argument for the current shell. Never interpolate untrusted text into
19 an unquoted command.
20- Prefer `--json` for one result and `--jsonl` for streaming or benchmark collection. Use text mode
21 only when its output will be returned directly to the user.
22- Do not inspect authentication environment variables or DeepChat's local descriptor. Authorization
23 is injected only after the command has passed the normal shell permission check.
24- A shell approval authorizes command execution. Sensitive mutations can additionally pause for a
25 renderer approval; wait for that decision and never attempt to manufacture confirmation data.
26- Use `deepchat help` or `deepchat <domain> <verb> --help` only when the options below are
27 insufficient. Do not probe undocumented routes.
28
29## Agent file and recursion boundaries
30
31- Agent callers may consume a DeepChat-owned artifact with `--artifact <id>` and inspect metadata
32 with `artifact describe`.
33- Do not use `--file`, `--out`, `--overwrite`, `artifact get`, or `artifact delete`. Agent callers
34 cannot upload arbitrary local bytes, download artifact bytes, or choose output paths.
35- Do not call `agent run` or `run watch`. An Agent cannot recursively create a detached Agent run,
36 and waiting on its own currently executing run would deadlock it. Use `run get` for a nonblocking
37 snapshot or `run cancel` to request cancellation.
38- Generated media remains in DeepChat's artifact spool. Return the artifact metadata or ID so the
39 application can render or reuse it.
40
41## Discovery and model calls
42
43```text
44deepchat system status --json
45deepchat system capabilities --json
46deepchat system doctor --json
47deepchat provider list --enabled-only --json
48deepchat model list --provider <provider-id> --json
49deepchat model config-get --provider <provider-id> --model <model-id> --json
50deepchat model invoke --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
51```
52
53Always discover provider and model IDs rather than guessing them. `model invoke` is a raw provider
54call: it does not create a chat session, run tools, or start an Agent loop.
55
56## Media, transcription, and OCR
57
58```text
59deepchat image generate --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
60deepchat video generate --provider <provider-id> --model <model-id> --prompt <quoted-text> --jsonl
61deepchat audio speak --provider <provider-id> --model <model-id> --text <quoted-text> --jsonl
62deepchat audio transcribe --provider <provider-id> --model <model-id> --artifact <artifact-id> --json
63deepchat ocr status --json
64deepchat ocr extract --artifact <artifact-id> --json
65deepchat artifact describe --id <artifact-id> --json
66```
67
68Use the provider/model lists to choose a compatible runtime. OCR is local and does not require a
69provider. OCR text is returned inline and is not written to the artifact spool.
70
71## Public configuration and management
72
73Read-only operations:
74
75```text
76deepchat settings get --json
77deepchat skill list --json
78deepchat mcp list --json
79```
80
81Agent callers may request renderer approval for preference-only settings, query-free HTTPS Skill
82installation, and adding a new disabled HTTPS remote MCP configuration. Only perform one when it
83directly satisfies the user's request:
84
85```text
86deepchat settings set --key <public-key> --value <json-scalar> --json
87deepchat skill install --url <https-url> --json
88deepchat mcp add --name <server-name> --stdin --json
89```
90
91The Agent setting allowlist is limited to presentation preferences such as font size/family,
92artifact effects, auto-scroll, notifications, and copy-with-reasoning. Agent Skill URLs cannot carry
93credentials, query parameters, or fragments. The main process classifies MCP input before approval
94and rejects stdio commands, non-HTTPS endpoints, headers, authorization bindings, or configurations
95too large to review safely. Provider/model configuration, credential writes, local Skill archives,
96Skill enable/disable/removal, MCP update/runtime control/removal, and every destructive operation
97require the DeepChat UI or a human terminal.
98
99## Benchmark discipline
100
101- Pin provider/model IDs and pass per-invocation options; do not mutate global defaults to prepare a
102 benchmark.
103- Record structured output, exit status, wall time, and errors. Preserve failed samples.
104- For OCR, distinguish cache hit, cache miss with warm runtime, cold runtime after app restart, and
105 offline availability. `ocr clear-cache` initializes the resource graph but does not start the OCR
106 helper, so classify the next extraction from its reported pre-extraction runtime state.
107- Run samples sequentially unless the benchmark explicitly measures concurrency; Agent compute is
108 rate-limited and bounded by the main process.
ThinkInAIXYZ