Creating Multica agents

This is the contract for Multica's agent-creation path: what the create entry points accept, what the server validates and rejects, how each field is persisted, and which fields the daemon actually reads at claim time. It is not a parameter manual — it states source-traced facts, and every claim is backed by file:line in references/creating-agents-source-map.md.

Quick start (read-only inspection)

These commands read state and have no side effects:

multica agent get <agent-id> --output json      # full persisted agent record
multica agent skills list <agent-id> --output json   # current skill bindings
multica agent env get <agent-id> --output json  # plaintext env (agent owner or ws owner/admin; agents denied)

An agent can also be unbound: runtime_id is NULL (served as "" with runtime_bound: false) after its runtime was deleted, which unbinds instead of deleting its agents (MUL-5559). An unbound agent keeps everything it owns and stays editable, but no trigger path will run it — they all refuse with agent_runtime_required — until agent update <id> --runtime-id <runtime-id> binds it again. Unbound is orthogonal to archived.

agent get returns the persisted agent including runtime_id, model, thinking_level, service_tier, custom_args, has_custom_env, custom_env_key_count, and skills. It never returns plaintext custom_env.

Core model

An agent is a workspace-scoped row (table agent). Creation is a single POST /api/agents (multica agent create). At task claim time the daemon re-reads the agent row and assembles the runtime payload — so the persisted fields, not the create-time output, are what the agent runs on.

Two distinct text fields, often confused:

• description is a catalog summary. It is stored and shown in listings; the daemon does NOT inject it into the agent's runtime prompt. Treat it as human-facing metadata only. Capped at 255 Unicode code points.
• instructions is the runtime behavior contract. The daemon reads it at claim time and ships it to the provider as the agent's durable instructions. Persona, responsibilities, boundaries, output and escalation rules go here, not in description.

CLI / API entry points

Minimum create call (--name and --runtime-id are both required):

multica agent create --name <name> --runtime-id <runtime-id> \
  --description "<short catalog summary>" \
  --instructions "<runtime behavior contract>" \
  --output json

runAgentCreate builds a JSON body and posts it to /api/agents. It only adds a key when its flag was provided — description/instructions on a non-empty value, the rest (runtime-config, custom-args, model, thinking-level, service-tier, visibility, …) on the flag being Changed — so omitted flags fall through to server defaults rather than sending empty strings. --max-concurrent-tasks is validated as 1–50 before the request is sent.

The HTTP body (CreateAgentRequest) accepts: name, description, instructions, conversation_starters, avatar_url, runtime_id, runtime_config, custom_env, custom_args, model, thinking_level, service_tier, visibility, max_concurrent_tasks, mcp_config, skill_ids.

The body accepts more than the CLI exposes. conversation_starters has no agent create / agent update flag — the CLI can only carry it across an agent copy. Setting it for a new agent means either calling /api/agents directly or telling the human to use the web UI; see below for where they will find it.

Copying an agent

multica agent copy <source-agent-id> forks an existing agent's portable configuration into a brand-new agent, leaving the source untouched. It is the CLI/headless equivalent of the web "Duplicate" action. No dedicated server API is involved: runAgentCopy reads the source with GET /api/agents/<id>, then POSTs a CreateAgentRequest — passing the source's skill ids in skill_ids so the bindings attach in the SAME create transaction (unlike agent create, which binds nothing). The mutation is therefore a single atomic create.

multica agent copy <source-agent-id> --name "My Agent (copy)"   # same runtime
multica agent copy <source-agent-id> --runtime-id <target> --model <model>  # cross-runtime fork

• Copied by default without a dedicated override flag: conversation_starters.
• Copied by default, each overridable with the matching flag: name (suffixed " (copy)"), description, instructions, avatar, custom_args, max_concurrent_tasks, invocation permission (permission_mode + allow-list), and assigned workspace skills.
• A copied max_concurrent_tasks is included only when the source value is within 1–50. Historical out-of-range values are omitted so the new agent receives the server default (6); an explicit out-of-range --max-concurrent-tasks override is rejected before any API request.
• Runtime-specific fields (model, thinking_level, service_tier) are copied ONLY when the target runtime is unchanged. --runtime-id selecting a different runtime drops them and REQUIRES --model (pass --model "" to accept the target runtime default), mirroring the web Duplicate clearing model on a runtime switch.
• Never copied: custom_env, mcp_config, runtime_config (secret / machine-local; redacted or masked on read anyway). Supply fresh values with the same secret-safe flags as agent create (--custom-env*, --mcp-config*, --runtime-config), or with agent env set after the copy exists.
• --no-skills skips copying the source's skill bindings.

Field contracts

Defaults when omitted or explicitly null: max_concurrent_tasks → 6. Other defaults when omitted: runtime_config → {}, custom_env → {}, custom_args → [], avatar_url → a random emoji:<glyph>, visibility → private (all materialized server-side before the insert). custom_args/runtime_config are typed []string/any and marshaled as-is — the JSON-shape rejection happens in the CLI, not the create handler.

The 1–50 concurrency range applies consistently to create and update. On create, an omitted field defaults to 6 while an explicitly supplied 0 is rejected; on update, omission preserves the current value. The CLI performs the same range check before sending create or update requests.

thinking_level is validated only at the provider level: fixed-vocabulary providers reject an unrecognized literal, while dynamic-vocabulary providers such as Codex/OpenCode accept a syntactically safe token. Pi's provider-level vocabulary is fixed (off|minimal|low|medium|high|xhigh|max), but its exact supported subset is model-specific and discovered from the local Pi RPC model catalog. A value unsupported for the chosen model is NOT rejected here — the daemon checks its local model catalog at execution time, logs a warning, and omits the incompatible override.

Set it from the CLI with --thinking-level on agent create and agent update, mirroring --model: the flag is a thin pass-through to the top-level thinking_level field, and on update an empty string (--thinking-level "") clears it back to the runtime default. The CLI deliberately does not enumerate the valid levels — they are runtime/model-specific (Claude currently uses low|medium|high|xhigh|max; Pi uses off|minimal|low|medium|high|xhigh|max; Codex values are discovered from the runtime's model catalog). It forwards the token, the server applies the provider's fixed-enum or safe-token gate, and the daemon performs the exact model/level check. A runtime whose provider has no thinking concept rejects any non-empty value with a 400.

service_tier is the matching first-class Codex speed control. It has three distinct states:

• empty means inherit the local Codex configuration;
• default means explicitly use Standard routing;
• a runtime catalog tier such as priority means explicitly use Fast.

Set it with --service-tier <value> on create/update; use --service-tier "" on update to clear it. The picker offers default only when the daemon reports that its installed Codex CLI supports the request-only explicit-standard sentinel (Codex 0.133.0+). A missing capability from an older daemon is treated as unsupported. The runtime model catalog owns availability and display copy for alternative tiers. The server accepts safe future Codex values, while the daemon verifies the explicit-standard capability or exact model/catalog-tier pair before execution and omits a stale incompatible override. An alternative catalog tier on an agent without an explicit model still fails closed because the effective config.toml model is unknown; default is model-independent once the runtime capability is known.

conversation_starters

The product calls this feature Conversation starters (中文：对话开场建议). Use that name when talking to a human — the wire field conversation_starters is an implementation detail they never see. A human configures them on the agent's Instructions tab; the deep link is /<workspace>/agents/<id>?view=instructions&focus=conversation_starters.

They are up to three label + prompt pairs shown above the composer when someone opens a new Chat with this agent. Selecting one only fills the composer — it never starts a run, so they are suggestions, not actions. Omitting the field on create defaults to []; omitting it on update preserves the stored value, and an explicit [] clears it. An agent with none configured still shows three built-in generic defaults in that empty state, so "the Chat shows suggestions" does not mean this agent has any of its own.

model vs custom_args

model is a first-class persisted column the daemon reads directly. custom_args are normally raw provider CLI args. The CLI help notes that some providers (codex app-server, openclaw) reject --model inside custom_args — but that is documented CLI guidance, not a server-enforced invariant; nothing in the create handler inspects custom_args for a model flag. Provider backends may consume protocol selectors before launch:

• Pi filters --thinking because the first-class thinking_level field owns that flag and must be its only source.
• ZeroClaw consumes --agent <alias> / --agent-alias <alias> (including =value forms) and sends the value as the ACP session/new.agentAlias parameter. zeroclaw acp has no such CLI flag. Set one of these custom args when ZeroClaw has multiple agents and no [acp].default_agent; omit it for a sole-agent config so ZeroClaw can auto-select that agent.

Never put credentials or other secrets in custom_args. Daemon command logs redact argument values, but values that a backend does not consume still live in the provider process's argv and may be visible to other local processes through ps or /proc. Put provider credentials in custom_env instead, using its stdin or 0600 file input where possible.

Env & secrets

custom_env is secret material. The CLI offers three input channels; two keep secrets out of shell history and the process list:

multica agent create --name <name> --runtime-id <runtime-id> --custom-env-stdin --output json
multica agent create --name <name> --runtime-id <runtime-id> --custom-env-file <0600-json> --output json

--custom-env-stdin reads the JSON object from stdin; --custom-env-file reads it from a file (suggested mode 0600). The third channel, --custom-env <json>, puts the value on the command line where shell history and ps can see it — avoid it for real secrets.

Read-side facts (these are the wrong assumptions to avoid):

• Agent resources never expose plaintext custom_env. agent list/get/create/update and WS events return only has_custom_env (bool) and custom_env_key_count (int).
• Reading plaintext values requires the dedicated GET /api/agents/{id}/env endpoint (multica agent env get). It is gated to the agent's own human owner or a workspace owner/admin, and agent actors are denied regardless of the backing member's role — a running agent cannot read another agent's secrets, not even one its own human owns.
• Writing values after creation does NOT go through agent update. The generic update handler rejects any custom_env field with a 400 ("use PUT /api/agents/{id}/env"). Plaintext env writes are handled by PUT /api/agents/{id}/env (multica agent env set), which carries the same gate and writes an audit row.

mcp_config

mcp_config is the agent's MCP server configuration (a JSON object such as {"mcpServers": {…}}). It is also secret material — MCP entries routinely embed API tokens — and offers the same three input channels as custom_env, on BOTH agent create and agent update:

multica agent create --name <name> --runtime-id <runtime-id> --mcp-config-file <0600-json> --output json
multica agent update <agent-id> --mcp-config-stdin --output json
multica agent update <agent-id> --mcp-config 'null'   # clears the config

--mcp-config-stdin / --mcp-config-file keep the value out of shell history and ps; the inline --mcp-config <json> does not. The CLI requires a JSON object or the literal null; a top-level array or primitive is rejected client-side, and empty stdin/file input errors rather than silently clearing.

Two ways mcp_config differs from custom_env:

• It IS settable through agent update. Unlike custom_env, mcp_config has no dedicated audited endpoint — the generic PUT /api/agents/{id} accepts it. Tri-state per the raw request body: field omitted → no change; null → clear; object → replace.
• It is serialized on read, but redacted. agent get/list return mcp_config only to callers allowed to view agent secrets; otherwise the field is null and mcp_config_redacted is true. Agent actors never see it, and a workspace may force redaction for everyone.

Provider support is not uniform: Qwen Code accepts a managed mcp_config through a daemon-owned 0600 temporary JSON file passed with --mcp-config; it is removed when the run exits. Leave the field unset (null) to inherit Qwen Code native settings.

Workspace MCP servers

A workspace keeps a LIBRARY of MCP servers (workspace Settings → MCP, or multica workspace mcp list|add|update|remove). Adding one there gives it to NO agent — same shape as a workspace skill. It reaches an agent only when someone assigns it:

multica workspace mcp list --output table        # find the server id
multica agent mcp add <agent-id> <server-id>     # give it to one agent
multica agent mcp disable <agent-id> <server-id> # stop sending it, keep the assignment
multica agent mcp remove <agent-id> <server-id>  # take it away

At claim time the effective set is:

Two consequences worth knowing before writing an agent's config: assigning a shared server does not require re-listing it in mcp_config (they merge), and mcp_config is now only about servers private to that agent — a managed-but-empty {} no longer means anything about the workspace layer, because nothing is inherited in the first place.

The stored entry is write-only — reads return the server's name and transport, never urls, commands, headers, or env, for any role.

Skill binding

Creating an agent does NOT bind any workspace skill — binding is a separate call after the agent exists. Two distinct verbs:

• add is additive — it merges the given ids with existing bindings (POST /api/agents/{id}/skills/add).
• set is replace-all — it overwrites the entire binding list with exactly the given ids (PUT /api/agents/{id}/skills); --skill-ids '' clears all.

multica agent skills add <agent-id> --skill-ids <skill-id> --output json
multica agent skills list <agent-id> --output json

At claim time the daemon assembles the agent's skills as workspace-bound skills FIRST, then appends the platform built-in skills. LoadAgentSkills loads each bound skill's content plus its supporting files; built-in skills are embedded at compile time and loaded from SKILL.md + sibling files. Both reach the provider as skill content — which is why capability belongs in a bound skill, not pasted into instructions.

Side effects needing approval

Read-only (safe): agent get, agent skills list, agent env get.

State-changing (require an explicit instruction — do not run speculatively):

• multica agent create — inserts a new agent row.
• multica agent copy — inserts a new agent row (a fork of an existing agent); the source is left untouched.
• multica agent skills add / set — mutate bindings (set is destructive: it drops bindings not in the new list).
• multica agent env set — overwrites the full custom_env map and writes an audit row.

Common wrong assumptions

• "description is the prompt." It is not — only instructions reaches the runtime. A rich description with empty instructions yields a named shell with no operating contract.
• "Create binds the agent's skills." It does not; bind explicitly afterward.
• "agent update can rotate env." It cannot — it 400s on custom_env; use the env endpoint.
• "mcp_config behaves like custom_env on update." It does not — mcp_config IS settable via agent update (--mcp-config), with --mcp-config null to clear; only custom_env is gated behind the dedicated env endpoint.
• "agent get shows env values." It shows only has_custom_env and custom_env_key_count.
• "Every accepted body field has a CLI flag." conversation_starters does not — agent create/agent update cannot set it, and agent copy only carries an existing value forward.
• "An invalid thinking_level/model combo is caught at create." Only an unknown provider-level literal is — model-specific gaps fail at run time.
• "set and add are interchangeable for skills." set replaces all bindings; using it when you meant add silently removes capabilities.

References

references/creating-agents-source-map.md maps every contract above to its file:line on the current tree, the runtime effect, and a safe read-only verification command.