| name | customize |
| description | Create, configure, and maintain custom agent profiles and author new skills via the `repl` tool. Use when the user wants to create an agent profile, build a custom agent, modify agent capabilities, attach or detach skills/connectors on a profile, author a skill, or inspect which connectors and tools are available. Also use whenever you need the `host.agents.*` or `host.skills.*` Python SDK. |
| license | Apache-2.0 |
Customize
Build and maintain agent profiles and skills programmatically via the
repl tool using host.agents.* and host.skills.*.
A profile is a named bundle that shapes how an agent behaves:
system_prompt — the profile's identity. This is the opening of the
agent's system prompt; it REPLACES the generic "You are Claude Science" base identity.
Write it in second person, lead with You are {display_name}, ..., state what
the agent specializes in and what it does NOT do. Everything else (tool-usage
rules, working-style bullets, scope guardrail) is inherited automatically —
don't restate it.
display_name / description / icon_key / color_key — picker metadata.
skill_names (optional restriction) — by default a profile sees the
full live skill catalog via search_skills / skill(...), same as the
main agent. Pass an explicit list ONLY to deliberately restrict it; [] creates
a zero-skill specialist. Restricting skills also restricts connectors —
a single unrestricted flag governs both; passing skill_names flips the
profile to curated mode and starts it with zero connectors (see next).
- Connector access — an unrestricted profile (the default) reaches
every connector (bundled + custom + authorized directory), same as the
main agent; use
detach_connector to subtract specific ones. A curated
profile (one created with an explicit skill_names list, or flipped via
{"unrestricted": False}) starts with no connectors — reach is exactly
what you attach_connector.
excludedTools — per-tool blocklist applied after connectors resolve.
Use to strip specific high-risk or irrelevant tools from an otherwise-useful
connector. Per-connector, not a profile field — set via
attach_connector(..., include_tools_pattern=/exclude_tools_pattern=); the
profile's excludedTools in list() is the read-only aggregation across all
its attached connectors. Patterns match the connector's bare tool names
(e.g. '^list_marts$', as returned by list_connectors(name)['tools']); the
aggregated excludedTools entries are stored fully-qualified as
mcp_<connector>_<tool> since the list spans every attached connector.
Python SDK
All calls run via the repl tool (see "Runs via the repl tool"
below). Return values are plain dicts/lists; errors raise RuntimeError
with a host.agents.*: / host.skills.*: prefix.
host.agents
host.agents.list()
host.agents.create(name, display_name, description,
system_prompt="", skill_names=None)
host.agents.update(name, patch)
host.agents.switch(name)
host.agents.delete(name)
host.agents.attach_skill(name, skill)
host.agents.detach_skill(name, skill)
host.agents.attach_connector(name, connector,
include_tools_pattern=None,
exclude_tools_pattern=None)
host.agents.detach_connector(name, connector)
host.agents.list_connectors(connector_name=None)
host.skills
host.skills.list()
host.skills.read(name, path="SKILL.md")
host.skills.edit(name, path, content, old_string=None)
host.skills.publish(name, overwrite=False)
host.skills.delete(name)
Runs via the repl tool
host.agents.* / host.skills.* execute in the control-plane
kernel — a separate Python process from your python cells, reached via
the repl tool (not the python tool). It shares your workspace
directory (cwd) but not memory, so variables from python cells aren't
visible there and vice-versa. To hand results across, write to a file —
same pattern as Python↔R:
import json, os
os.makedirs("handoff", exist_ok=True)
profiles = host.agents.list()
json.dump(profiles, open("handoff/agents.json", "w"))
import json, pandas as pd
profiles = json.load(open("handoff/agents.json"))
pd.DataFrame(profiles)[["name", "source", "enabled"]]
Workflow: scope → draft → review → create
Do not call generate_plan for profile CRUD. This is a single
scope→draft→confirm loop; ask_user (step 4) is the review gate. A plan
adds a second approval that duplicates the ask_user confirmation and
drags in step-status bookkeeping that fights this workflow.
User approval. Most host.agents.* / host.skills.* calls apply
immediately — create/update (including unrestricted),
attach_*/detach_*, and skills.publish/skills.edit are pre-approved
at session start and the cell does NOT pause. An approval card (cell
pauses, resumes automatically on Allow — you don't retry) appears only
for the calls that hand off identity or free a granted name:
host.agents.switch(name) — per-target; Allow covers this name only
host.agents.update(name, {"name": ...}) (rename) — per-target
host.agents.delete(name) / host.skills.delete(name) — one card
per project: the first delete shows a card; "Allow for this project"
covers every subsequent delete in this project, including bulk
teardown. Do NOT tell the user they'll see N cards for N deletes.
The name is read from the runtime call for switch and rename, so both
literal and variable names work (host.agents.switch("FOO") or
host.agents.switch(name_var)); keep the ask_user review step
so the click is a quick confirm, not a surprise. After the call returns,
read back (host.agents.list()) to confirm the actual state — don't
narrate an expected card.
Reading existing profiles
Call host.agents.list() to see the user's current profiles so you don't
duplicate an existing agent. The main agent's bundled profile is protected —
it cannot be renamed or deleted.
1. Scope first
What is this agent for? A profile should have one job. Pick an UPPER_SNAKE
name (RNASEQ_REVIEWER, not RNA-seq review helper). Use ask_user if the
name or scope is unclear — profiles are user-visible in the picker.
2. Write the identity
system_prompt is the agent's opening paragraph — it replaces the base
identity, it's not an addendum. Lead with You are {display_name}. State the
specialization and the boundaries ("You handle X, Y, Z. You do not handle
..."). Keep it under ~200 words; the heavy how-to lives in skills, not the
prompt.
3. Ask: full access or a subset?
Before creating, ask_user whether this profile should have full access
(the live skill catalog and every connector — same reach as the main agent;
new skills and connectors appear automatically) or a restricted subset
(a fixed list you'll curate together). Don't infer this from the role
description — a narrowly-described specialist may still want full reach,
and a broadly-described one may want a tight loadout. Pair this with the
name/prompt review in step 4 so it's one round-trip.
4. Review with the user
Show the proposed name, display name, description, system_prompt, and the
user's full-vs-subset choice from step 3. ask_user to confirm before
writing. If they chose a subset, list the proposed skills/connectors here.
5. Create
Call host.agents.create(name, display_name, description, system_prompt=...).
If the user chose full access, leave skill_names unset. If they chose a
subset, pass skill_names=[...] with the agreed list — this flips the
profile to curated mode and starts it with zero connectors, so attach the
agreed connectors after create as in step 6. For edits to an existing profile,
use host.agents.update(name, {...}) with targeted fields; prefer
host.agents.attach_skill(...) / host.agents.detach_skill(...) and
host.agents.attach_connector(...) / host.agents.detach_connector(...)
over wholesale skill_names replacement so you don't clobber the user's own
edits.
After the profile exists, offer to switch to it:
host.agents.switch(name). The user sees an approval card; on Allow, this
conversation continues as the new specialist from their next message. If they
decline, tell them they can select it from the session config popover on any
new conversation.
6. Restricting the loadout (when the user chose a subset)
If the user chose a subset in step 3, curate after create:
- Skills:
host.agents.update(name, {"skill_names": [...]}) with the
exact list, or detach_skill one at a time. Check host.skills.list()
for available names. On update, skill_names is an exact replace — never
send a partial list to "add"; fetch the current skillNames (camelCase in
the return dict), modify, send back.
- Connectors: a curated profile (created with an explicit
skill_names
list, or flipped via {"unrestricted": False}) starts with NO
connectors — reach is exactly what you attach. Call
host.agents.list_connectors() to see every available connector (bundled +
directory + user-added MCP) with its auth state, then
host.agents.attach_connector(name, connector_name) for each connector
the user agreed to keep. Use include_tools_pattern=/exclude_tools_pattern=
on the attach call to scope which tools the profile gets (omit both to
expose every tool; re-attaching without patterns preserves the existing
exclusion list — pass include_tools_pattern='.*' to clear). A connector
with authState other than "authorized" or "not-required" must be
connected via the Connectors panel before it can be attached.
detach_connector only removes an explicit attachment; on a curated
profile there is nothing to detach — don't use it to "restrict".
- After attaching, read back
host.agents.get(name)["connectors"] to confirm
the reach matches what you told the user.
- To undo a restriction:
host.agents.update(name, {"unrestricted": True}).
7. Set up the environment
After the profile exists, propose a conda environment for it: name it after the
profile (lowercase slug, e.g. rnaseq-reviewer), pick python or r based on
the skills you attached, and list the packages those skills need. ask_user to
confirm the package list, then call
manage_environments(mode="create", name="<slug>", packages=[...]) and, if
anything further is needed,
manage_packages(mode="install", environment="<slug>", packages=[...]). The
profile uses this env by default; the system-owned python and r skeletons
remain available as fallbacks.
What makes a good profile
- A sharp identity. One job, stated clearly in
system_prompt. Full
catalog access by default — restrict the loadout only when the user asks.
- Composable. Reuse existing skills; don't inline workflow steps into
system_prompt that belong in a skill.
- Safe by default. If a connector is powerful, exclude the tools the agent
doesn't need.
Authoring skills
For the full skill-authoring guide (anatomy, progressive disclosure, eval loop,
description optimization), first load it:
skill({"skill": "skill-creator"})
Then use the host.skills SDK via the repl tool to write and publish:
host.skills.edit("my-skill", "SKILL.md", """---
name: my-skill
description: ...
---
# My Skill
...
""")
host.skills.edit("my-skill", "kernel.py", "def helper(x): ...\n")
print(host.skills.read("my-skill")["content"])
host.skills.publish("my-skill")
Once published, the skill is in the live catalog — every unrestricted profile
(including the main agent and any profile created with the default) sees it
immediately. For a restricted profile, attach it explicitly with
host.agents.attach_skill(profile_name, "my-skill").
Kernel sidecars (kernel.py / kernel.R)
If a skill's workflow depends on reusable helper functions, ship them as
kernel.py (and/or kernel.R) at the skill root. When any agent calls
skill({skill: <name>}), that file is executed in its persistent python/R kernel and
the tool result reports which top-level names were defined — so SKILL.md can
say "call annotate_df(df)" and the function already exists.
Sidecars are validated before execution so that loading a skill only defines
names — nothing author-written runs at load time. Allowed at the top level:
def / async def (no decorators). Default argument values must be
literals — def f(url=MY_CONSTANT) gets the whole file rejected. Wrap
constants with an explicit is None check: def f(url=None): then
if url is None: url = MY_CONSTANT (not url = url or MY_CONSTANT,
which also replaces 0, "", []).
import / from … import name (no *). Defer third-party imports to
inside function bodies — the skeleton python env ships stdlib + a small
starter set (numpy, pandas, scipy, matplotlib, seaborn, pillow), so e.g.
import requests at module scope surfaces a load error on every fresh
kernel. Import errors don't fail the skill load; the agent sees the
traceback and can manage_packages then re-load.
- Assignment of a literal constant to a plain name (e.g.
VERSION = "1",
LIMITS = (1, 2, 3)). Computed values like os.path.join(...) are
rejected — move them into a function body.
Anything else at the top level (classes, calls, if/for, non-literal
assigns) is rejected with [kernel.py rejected] …. Names starting with _
are reserved by the loader and cannot be bound at the top level (use
import os and reference os.path inside functions, not
import os as _os). Function bodies are not restricted — they run only
when the agent calls them in a python cell.
Keep kernel.py small and self-contained (same guidance applies to
kernel.R). If a helper wants more than ~100 lines, trim it to the core
operations the agent actually calls; scripts/ is for standalone CLI
tools run via bash, not for backing the sidecar. For kernel.py, the
skill directory is not on sys.path, so from scripts.X import … fails
— but the directory is on disk and readable. A Python sidecar function
can locate it via its own co_filename to shell out to a scripts/ tool:
import os, sys, subprocess
def run_pipeline(cfg_path):
here = os.path.dirname(sys._getframe().f_code.co_filename)
if not here:
raise RuntimeError("skill dir unavailable in this runtime")
tool = os.path.join(here, "scripts", "pipeline.py")
return subprocess.run([sys.executable, tool, cfg_path],
capture_output=True, text=True, check=True).stdout
__file__ is not set (sidecars share one kernel namespace, so a global
__file__ would point at whichever skill loaded last). co_filename is
per-function and normally points at this skill's on-disk kernel.py; in
rare cases (e.g. when the loader can't resolve the skill dir on disk) it
falls back to the bare name "kernel.py" — hence the if not here guard
above.
The runtime exports PYTHONSAFEPATH=1, which the subprocess.run child
inherits — so python scripts/pipeline.py does NOT put scripts/ on the
child's sys.path. A multi-file tool that does sibling imports must add
sys.path.insert(0, os.path.dirname(__file__)) at the top of its entry
script (or pass env={**os.environ, "PYTHONSAFEPATH": ""} to
subprocess.run). If the child exits non-zero, capture_output=True
hides the traceback — inspect CalledProcessError.stderr.
kernel.R has no runtime analogue of co_filename (the R loader parses
with keep.source=FALSE), so an R sidecar that needs to reach scripts/
should take the path as an argument from the caller.
Minimal example:
import pandas as pd
def annotate_df(df: pd.DataFrame, gene_col: str = "gene") -> pd.DataFrame:
"""Attach HGNC symbols; see SKILL.md ## Workflow step 3."""
import requests
...
return df
Write sidecars via host.skills.edit(name, "kernel.py", src). When the
gate probe can judge, the edit result carries sidecar_gate: {ok, error?}
— the same structural gate the load path runs — so a reject (non-literal
default, _-prefixed name, top-level call) surfaces immediately. When the
probe can't judge (interpreter unavailable, or the source doesn't parse
under the host's interpreter — possible version drift), the key is absent
and note says why. Iterate against that; don't load the skill just to
test the sidecar. host.skills.publish refuses only on a structural
reject.