MCP Agent Operations
The MCP Agent Operations server provides deterministic agent-development operations, including repository claim management with worktree isolation and resource locks (acquire, extend, heartbeat, release, journal maintenance, reporting), YAML and local Markdown link verification, full Agent Skill lifecycle management (list, find, read, load, validate, refresh, and load supporting resources), reference data aggregation, technology skill detection via a registry, and hierarchical HTML rendering for durable, nested plans and data.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@MCP Agent Operations list installed agent skills"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP Agent Operations
mcp-agent-ops is a local stdio MCP server for deterministic agent-development operations that otherwise cause repeated shell and generated Python calls.
The service owns six capability groups:
repository claims, worktree isolation, event journaling, archival, and contention reporting;
reusable YAML and checkpoint-scoped Markdown verification operations;
recursive reference-data aggregation across authorized project and user folders;
snapshot-based discovery and extension-aware batched loading of installed Agent Skills;
Agent Skill validation; and
evidence-based technology-skill detection.
The domain packages are independent of FastMCP. The installed mcp-agent-ops command starts the FastMCP stdio server, while mcp-agent-ops-claims preserves the direct claim CLI contract. Claim callers can own work-item activity, explicit or broad file scope, or one policy-bounded runtime resource with deadlines. Every checkout resolves disposable claim state to the primary worktree's ignored .agent-ops/resource-claim directory.
The claim engine, technology detector, and Agent Skill validator began as copies of the accepted dev-methodology implementations. See docs/reference/copied-scripts.md for exact provenance, intentional adapter changes, and compatibility coverage.
See docs/reference/mcp-tools.md for the complete small-call tool and resource surface.
Supported platforms
The stdio server, direct claim CLI, repository locking, reference and skill catalogs, verification, and technology detection run natively on:
macOS with Python 3.11, 3.12, or 3.13;
Linux with Python 3.11, 3.12, or 3.13; and
Windows with Python 3.11, 3.12, or 3.13 in releases newer than
v0.4.0.
Published release v0.4.0 predates native Windows claim locking. macOS and Linux remain supported by every published release. The optional shared multi-process evaluation audit uses POSIX file locking and remains limited to macOS and Linux; this does not affect ordinary MCP or claim operation on Windows.
Related MCP server: ai-mcp
Install the latest release
The supported distribution is the wheel attached to the latest GitHub Release. Do not install the generated source archive when only the runtime server is needed; the wheel excludes tests, documentation, and development dependencies.
On Windows, first confirm that the latest release is newer than v0.4.0. Then open PowerShell and install uv plus the GitHub CLI:
winget install --id astral-sh.uv -e
winget install --id GitHub.cli --source winget
gh auth loginOpen a new terminal after WinGet changes PATH, then download and verify the latest release:
$releaseDir = Join-Path ([System.IO.Path]::GetTempPath()) "mcp-agent-ops-$([guid]::NewGuid())"
New-Item -ItemType Directory -Path $releaseDir | Out-Null
gh release download `
--repo martinbechard/mcp-agent-ops `
--pattern '*' `
--dir $releaseDir
Push-Location $releaseDir
Get-Content .\SHA256SUMS | ForEach-Object {
$expected, $file = $_ -split '\s+', 2
$file = $file.TrimStart('*')
$actual = (Get-FileHash -Algorithm SHA256 $file).Hash.ToLowerInvariant()
if ($actual -ne $expected.ToLowerInvariant()) {
throw "Checksum mismatch: $file"
}
}
Pop-LocationInstall the wheel with its tested, locked runtime dependencies:
$wheels = @(Get-ChildItem $releaseDir -Filter 'mcp_agent_ops-*.whl')
if ($wheels.Count -ne 1) {
throw "Expected exactly one mcp-agent-ops wheel."
}
uv tool install `
--python 3.11 `
--with-requirements (Join-Path $releaseDir 'runtime-requirements.txt') `
$wheels[0].FullName
mcp-agent-ops --version
mcp-agent-ops --identity-json
uv tool dir --binOn macOS or Linux, install uv and an authenticated GitHub CLI, then download and verify the latest release assets:
release_dir="$(mktemp -d)"
gh release download \
--repo martinbechard/mcp-agent-ops \
--pattern '*' \
--dir "$release_dir"
(
cd "$release_dir"
if command -v sha256sum >/dev/null; then
sha256sum -c SHA256SUMS
else
shasum -a 256 -c SHA256SUMS
fi
)Install the wheel and its tested, locked runtime dependencies into an isolated tool environment:
uv tool install \
--python 3.11 \
--with-requirements "$release_dir/runtime-requirements.txt" \
"$release_dir"/mcp_agent_ops-*.whl
rm -rf "$release_dir"Running the same procedure after a newer release is published replaces the installed tool version. uv tool list reports the installed version and commands.
Verify the installed server package without starting stdio:
mcp-agent-ops --version
mcp-agent-ops --identity-jsonThe JSON identity includes a location-independent SHA-256 digest over installed runtime resources. Evaluation runners can pin both that runtime digest and the launcher executable digest instead of treating a small console-script wrapper as the server implementation.
Development
Python 3.11 or newer and uv are required for development.
uv sync
uv run pytest
uv run ruff check .
uv run mypy srcHierarchical HTML rendering and durable plans
The package and MCP server expose render_hierarchy_html for turning nested mappings and sequences
into responsive, self-contained HTML trees. The operation accepts in-memory data, JSON or YAML
text, or an existing .json, .yaml, or .yml file.
from mcp_agent_ops.hierarchy import render_hierarchy_html
plan = {
"Delivery": {
"steps": [
{"name": "Discover", "complete": True},
{"name": "Implement", "complete": False},
]
}
}
html = render_hierarchy_html(plan, title="Delivery plan")For an agent-managed plan, create a durable JSON source beside the rendered HTML. The creation function returns the JSON path that every later mutation accepts:
from mcp_agent_ops.hierarchy import create_hierarchy_plan, update_hierarchy_plan
plan_path = create_hierarchy_plan(
{"Delivery plan": ["Discover", "Implement", "Release"]},
output_filename="delivery-plan.html",
output_folder="reports",
)
update_hierarchy_plan(plan_path, "2", add_child="Write focused tests")
result = update_hierarchy_plan(plan_path, "2.1", completed=True)
if result.next_task is not None:
print(result.next_task.identifier, result.next_task.label)The MCP server publishes the same three names: render_hierarchy_html,
create_hierarchy_plan, and update_hierarchy_plan. MCP file and folder paths must be absolute
and resolve beneath MCP_AGENT_OPS_WORKSPACE_ROOTS or Codex's conventional
~/.codex/visualizations subtree. Creation returns canonical JSON unless an explicit or configured
destination is available, in which case it returns the resolved JSON plan path. Update returns
structured success, that persisted path, any automatically completed ancestors, and the next
incomplete executable leaf with its parent context.
See the complete hierarchical HTML renderer reference for rendering, durable plan creation, exact item targeting, mutations, numbering, read-only browser markers, themes, validation, errors, regeneration behavior, and implementation links.
A runnable hierarchy gallery includes structured examples
and a reviewable Markdown document outline. The examples use the packaged default, outline, and
midnight themes plus a deliberately distinctive caller-supplied blueprint theme. Generate the
live previews from the repository root:
uv run python examples/hierarchy-gallery/generate_gallery.pyLocal MCP configuration
Configure an MCP host to run:
mcp-agent-opsThe server uses stdio by default. Configure these boundaries before exposing it to an agent:
MCP_AGENT_OPS_SKILL_ROOTScontains precedence-ordered readable skill roots, separated by the operating system path separator. A root may contain child skill directories or may be one exact skill directory containingSKILL.md.MCP_AGENT_OPS_REFERENCE_ROOTScontains ordered readable user folders (conventionally~/.agents/referencethen~/.codex/reference), separated by the operating system path separator. Every UTF-8 file beneath a configured folder is available by relative path.MCP_AGENT_OPS_DETECTION_REGISTRYidentifies the trusted methodology-owned technology registry.MCP_AGENT_OPS_WORKSPACE_ROOTScontains allowed project and worktree roots, separated by the operating system path separator.MCP_AGENT_OPS_HIERARCHY_OUTPUT_FOLDERoptionally selects the default authorized destination for hierarchy HTML and JSON files when a tool call omitsoutput_folder.
Hierarchy source, theme, output, and plan paths additionally accept Codex's conventional
~/.codex/visualizations subtree. This hierarchy-only default does not authorize repository,
claim, verification, skill, or reference operations in that location. If neither an explicit
output_folder nor MCP_AGENT_OPS_HIERARCHY_OUTPUT_FOLDER is available, render calls return HTML
and plan-creation calls return canonical JSON without creating files.
When the server starts with its working directory beneath a configured workspace root, it automatically overlays recursively discovered skills from <cwd>/.agents/skills and <cwd>/.codex/skills ahead of the configured user roots. The .agents project root has precedence over the .codex project root. Nested project skill directories are supported; duplicate skill names inside either one project root are rejected as ambiguous. skill_refresh rescans both project and configured roots.
If a discovered skill is a symlink whose SKILL.md resolves outside every configured root, catalog
operations report the discovered manifest, resolved target, current roots, and an exact
MCP_AGENT_OPS_SKILL_ROOTS value to authorize the target. Correct the setting and restart the MCP
server before retrying. Other invalid catalog entries report the underlying manifest or read error
and the containing root that must be corrected or removed.
skill_read and skill_load accept the optional include_extensions switch. Its default is
false. When enabled for a base skill such as python, the loader searches the same precedence
order for python.extension and appends that complete skill when found. The base and extension
resolve independently, so either one can come from a project or configured user root.
When the working directory is beneath an allowed workspace, the server recursively publishes files beneath <cwd>/.agents/reference and <cwd>/.codex/reference before files beneath the configured user reference folders. The server aggregates every matching relative path in search order with one newline between sources. Traversal and symlinks that resolve outside their selected folder are not published. reference_refresh rescans all reference folders. Project .agents/reference precedes .codex/reference; all project sources precede configured user sources. No filename allowlist is used, and neither the project root nor user home is implicitly published. Configure user folders explicitly, as with skill roots.
Repository, project, verification, worktree, and validation paths supplied through tools must be absolute and resolve beneath their configured boundary. Name-based skill validation uses the same catalog lookup as skill loading. Explicit skill-validation paths may also target unpublished skills anywhere beneath the authorized working project, without adding those paths to catalog discovery. Catalog discovery, skill validation, and technology detection recheck every nested manifest, metadata file, source file, and supporting resource before reading it. The server rejects missing boundary configuration, traversal, and symlink escape rather than granting ambient filesystem access.
Verify Markdown changed by one operation
Use a repository checkpoint to verify only the Markdown files changed by one bounded operation:
capture_repository_state(repository_root)
perform the bounded operation
verify_markdown_links(
repository_root,
scope="changed_since_checkpoint",
checkpoint_id="..."
)The server derives added, modified, renamed, and deleted paths. It also checks current Markdown files that refer to deleted or renamed targets. Changes present before checkpoint capture are not selected. Checkpoints remain in one MCP server process and expire when that process exits.
Use scope="git_changed" when verification must include all current staged, unstaged, renamed,
deleted, and untracked changes. Use the default patterns scope for explicit paths or globs. See
docs/reference/mcp-tools.md for result fields, missing
path behavior, unmatched-glob behavior, and rename limitations.
Checkpoint capture and verification are read-only. They do not change repository files,
permissions, HEAD, or the Git index.
The reference and skill catalogs are built lazily and reused for the life of the server process. reference_refresh and skill_refresh atomically publish new snapshots after source files change. Technology registry configuration is also cached and takes effect after restarting the server. Claim state remains disk-authoritative and coordinates across server processes.
Junie on macOS
Junie reads MCP configuration from ~/.junie/mcp/mcp.json for user scope or .junie/mcp/mcp.json beneath one project. In the IDE, open Settings | Tools | Junie | MCP Settings.
Use the absolute executable directory reported by uv tool dir --bin. Replace YOUR_NAME and the workspace path with existing absolute paths:
{
"mcpServers": {
"mcp-agent-ops": {
"command": "/Users/YOUR_NAME/.local/bin/mcp-agent-ops",
"args": [],
"env": {
"MCP_AGENT_OPS_SKILL_ROOTS": "/Users/YOUR_NAME/.agents/skills:/Users/YOUR_NAME/.codex/skills",
"MCP_AGENT_OPS_REFERENCE_ROOTS": "/Users/YOUR_NAME/.agents:/Users/YOUR_NAME/.codex/skills",
"MCP_AGENT_OPS_DETECTION_REGISTRY": "/Users/YOUR_NAME/.agents/skills/detect-technology-skills/references/technology-skill-detection-registry.yaml",
"MCP_AGENT_OPS_WORKSPACE_ROOTS": "/Users/YOUR_NAME/dev",
"MCP_AGENT_OPS_HIERARCHY_OUTPUT_FOLDER": "/Users/YOUR_NAME/.codex/visualizations"
}
}
}
}macOS and Linux path lists use colons. On Linux, use the same configuration with Linux paths such as /home/YOUR_NAME.
Junie on Windows
Junie reads MCP configuration from %USERPROFILE%\.junie\mcp\mcp.json for user scope or .junie\mcp\mcp.json beneath one project. In the IDE, the same configuration is available under Settings | Tools | Junie | MCP Settings.
Use the absolute executable directory reported by uv tool dir --bin. Replace the example user and workspace paths with existing absolute paths:
{
"mcpServers": {
"mcp-agent-ops": {
"command": "C:\\Users\\YOUR_NAME\\.local\\bin\\mcp-agent-ops.exe",
"args": [],
"env": {
"MCP_AGENT_OPS_SKILL_ROOTS": "C:\\Users\\YOUR_NAME\\.agents\\skills;C:\\Users\\YOUR_NAME\\.codex\\skills",
"MCP_AGENT_OPS_REFERENCE_ROOTS": "C:\\Users\\YOUR_NAME\\.agents;C:\\Users\\YOUR_NAME\\.codex\\skills",
"MCP_AGENT_OPS_DETECTION_REGISTRY": "C:\\Users\\YOUR_NAME\\.agents\\skills\\detect-technology-skills\\references\\technology-skill-detection-registry.yaml",
"MCP_AGENT_OPS_WORKSPACE_ROOTS": "C:\\Users\\YOUR_NAME\\dev",
"MCP_AGENT_OPS_HIERARCHY_OUTPUT_FOLDER": "C:\\Users\\YOUR_NAME\\.codex\\visualizations"
}
}
}
}Windows path lists use semicolons. Restart Junie after saving the configuration, then confirm that mcp-agent-ops is active and exposes its tools in MCP Settings.
Evaluation runners may configure MCP_AGENT_OPS_AUDIT_LOG plus MCP_AGENT_OPS_AUDIT_ROOTS to create one exclusive digest-only JSON Lines tool-call trace. When a harness starts inherited MCP servers for a parent and subagent, set MCP_AGENT_OPS_AUDIT_SHARED=true plus a 32-character lowercase hexadecimal MCP_AGENT_OPS_AUDIT_SESSION_ID; each process then writes a separate random stream identity and process-local sequence into the same owner-only file under a POSIX file lock. Both modes record only canonical tool name, lifecycle status, call identity, sequence, and argument or result digests. Shared version-two records also carry the session and process stream identities, and their terminal records carry bounded canonical outcomes for supported deterministic operations. Reference loads record LOADED or REJECTED, and reference refresh records CATALOG or EMPTY. skill_validate records VALID or FINDINGS. These labels reveal no reference content, skill content, names, paths, or validation findings. The trace never stores arguments, returned content, prompts, or configured paths. Do not configure this trace for ordinary sessions that do not need evaluator-owned call evidence.
An evaluator can also set MCP_AGENT_OPS_REQUIRED_RUNTIME_DIGEST to the pinned value returned by --identity-json. The server checks it before importing FastMCP or starting stdio and fails closed when the installed runtime has drifted.
Create a release
Releases use semantic versions. The Git tag must be v followed by the exact project.version value in pyproject.toml; the release workflow rejects a mismatch.
Choose the next version and update
project.versioninpyproject.toml.Refresh the lockfile and run the complete local verification:
uv lock uv sync --locked uv run pytest uv run ruff check . uv run mypy srcCommit the version and lockfile, push
main, and wait for its CI run to succeed:VERSION="$(sed -n 's/^version = "\([^"]*\)"/\1/p' pyproject.toml)" git add pyproject.toml uv.lock git commit -m "Prepare release v${VERSION}" git push origin main gh run list --workflow CI --branch main --limit 1Tag the verified commit and push the tag:
git tag -a "v${VERSION}" -m "Release v${VERSION}" git push origin "v${VERSION}"Confirm that the Release workflow passed and the release is available:
gh run list --workflow Release --limit 1 gh release view "v${VERSION}"
The tag-triggered workflow reruns tests on Python 3.11, 3.12, and 3.13 before publishing the wheel, runtime-requirements.txt, and SHA256SUMS. Do not reuse or move a published version tag; increment the package version for the next release.
State ownership
Claim registries and event journals live beneath each target repository's primary-worktree .agent-ops/resource-claim directory, which projects ignore with the exact anchored /.agent-ops/resource-claim/ rule. Read-only status and reporting do not create absent state. A first mutating operation migrates only empty state from the rejected .codex/agent-claim root, preserves its history, writes the version-two canonical state marker, and installs incompatible markers at the old paths. Live legacy claims are drain-only through exact release. Once the complete marker is durable, normal claim operations require no access to .git or .codex. Reference and skill files remain authoritative in their configured scopes. Process-local catalog snapshots are read versions identified by digests, not independent state stores; publishing or restarting replaces them from disk.
Available Tools
26 toolscapture_repository_stateARead-only
Capture an opaque process-local checkpoint without changing the repository.
Inputs: repository_root (required). Returns: Return the opaque identity of one process-local repository checkpoint.
claim_acquireB
Atomically acquire one work item, file domain, or timed resource.
Work-item scope uses work_item_id with activity work or update. Named resources require complete timing evidence resolved against PROJECT.yaml. Project-files excludes backlog and supports caller-requested isolation only beneath the primary worktree's canonical .worktrees/ root, with backlog omitted through worktree-specific sparse checkout. Schema-version-2 results preserve the package-owned claim engine's current outcomes, rejection fields, and exit codes.
Inputs: task (required), agent (required), claim_id (required), repository (required), +19 more. Returns: Return one package claim-engine exit code and structured JSON result.
claim_extendA
Atomically add same-domain scope without weakening existing ownership.
A work-item claim cannot add operational scope. A named resource extension requires complete timing evidence. Primary-only scope from a linked claim returns SHARED_CHECKOUT_REQUIRED; invalid mixed scope leaves the registry unchanged.
Inputs: claim_id (required), repository (required), files, trees, +10 more. Returns: Return one package claim-engine exit code and structured JSON result.
claim_extend_deadlineC
Extend one configured resource hard stop with bounded evidence.
Inputs: claim_id (required), repository (required), extension_evidence (required), requested_hard_stop_duration_seconds (required… Returns: Return one package claim-engine exit code and structured JSON result.
claim_heartbeatC
Refresh one active claim heartbeat in the repository-global registry.
Inputs: claim_id (required), repository (required). Returns: Return one package claim-engine exit code and structured JSON result.
claim_maintain_journalB
Archive complete UTC claim-event days while retaining the configured hot window.
Inputs: repository (required), hot_days. Returns: Return one package claim-engine exit code and structured JSON result.
claim_releaseC
Release one exact claim, including a disposition for work-item ownership.
Inputs: claim_id (required), repository (required), disposition, blocker_reference. Returns: Return one package claim-engine exit code and structured JSON result.
claim_reportA
Report canonical contention metrics without creating or migrating absent state.
Inputs: repository (required), since. Returns: Return one package claim-engine exit code and structured JSON result.
claim_resetB
Replace the live registry with an empty claim list under its exact lock.
Inputs: repository (required). Returns: Return one package claim-engine exit code and structured JSON result.
claim_statusC
Return primary-root claim state or a structured migration stop without creating absent state.
Inputs: repository (required). Returns: Return one package claim-engine exit code and structured JSON result.
create_hierarchy_planA
Create a hierarchy plan as durable files or canonical inline JSON.
Source, theme, and output paths must be absolute beneath the hierarchy roots. The
output folder defaults to MCP_AGENT_OPS_HIERARCHY_OUTPUT_FOLDER. When neither
destination is supplied, the tool returns JSON without writing files.
Inputs: source (required), output_filename (required), theme, title, +3 more. Returns: result
detect_technology_skillsC
Detect required technology skills for project scopes using the configured registry.
Inputs: scopes (required), project_root (required). Returns: Return the copied detector exit code and structured loadout result.
reference_loadCRead-onlyIdempotent
Load recursive text references aggregated across every matching allowed folder.
Inputs: names (required). Returns: Return ordered aggregated references or an error-only all-or-nothing result.
reference_refreshB
Atomically refresh project and configured user reference scopes.
Inputs: no named parameters. Returns: Describe the active host-path-free snapshot and its available relative paths.
render_hierarchy_htmlA
Render safe self-contained hierarchy HTML and optionally save it.
Inline mappings, sequences, JSON, and YAML require no filesystem access. An explicit source path, custom theme folder, or output folder must be absolute and resolve beneath the hierarchy roots. A filename without an explicit or configured output folder returns inline HTML without writing a file.
Inputs: source (required), theme, title, numbering, +5 more. Returns: result
skill_findA
Return the precedence-resolved absolute SKILL.md path for one skill name.
Inputs: name (required). Returns: Identify one catalog name and its precedence-selected absolute manifest path. C…
skill_listA
List path-free skill metadata, digests, resources, and shadowing counts.
Inputs: no named parameters. Returns: Return the model-facing catalog revision and path-free skill metadata.
skill_loadBRead-onlyIdempotent
Load skills and optionally append independently resolved extensions.
Inputs: names (required), include_extensions. Returns: Return an ordered all-or-nothing set of complete skill documents.
skill_readB
Read one precedence-resolved skill and optionally append its extension.
Inputs: name (required), include_extensions. Returns: Return one model-facing skill document without exposing host filesystem paths.…
skill_read_resourceA
Read one supporting skill resource without permitting directory traversal.
Inputs: name (required), resource_path (required). Returns: Return one safely resolved supporting resource in text or base64 form.
skill_refreshA
Atomically refresh project and configured user skill roots.
Inputs: no named parameters. Returns: Return the model-facing catalog revision and path-free skill metadata.
skill_resource_loadB
Load several supporting resources in one ordered all-or-nothing operation.
Inputs: requests (required). Returns: Return an ordered all-or-nothing set of supporting skill resources.
skill_validateB
Validate catalog names or explicit skill paths beneath the working project.
Inputs: paths (required). Returns: Return whether every requested Agent Skill passed copied validation.
update_hierarchy_planA
Apply exactly one targeted plan mutation and regenerate its HTML rendering.
The plan path must be absolute beneath the hierarchy roots. The target is
an exact dotted hierarchy number or an exact unique title. Supply exactly one of
completed, text, add_child, replace_children, or add_peer_after.
Inputs: target (required), plan_path (required), text, add_child, +3 more. Returns: Report a successful durable mutation and the next executable plan task. `automa…
verify_markdown_linksBRead-onlyIdempotent
Verify local Markdown links using explicit or server-derived file selection.
Inputs: repository_root (required), scope, patterns, checkpoint_id. Returns: Return the complete read-only result of one verification operation. `ok` is tru…
verify_yamlC
Validate exact YAML files, including duplicate keys, with structured diagnostics.
Inputs: paths (required), repository_root (required). Returns: Return the complete read-only result of one verification operation. `ok` is tru…
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 26 tools
The tool set is split into reasonably clear sub-domains (claim_*, skill_*, reference_*, verify_*, hierarchy), but several near-pairs invite misselection: skill_read vs skill_load, skill_read_resource vs skill_resource_load, and claim_extend vs claim_extend_deadline differ only in plurality or nuance. The long inline descriptions clarify the distinctions, but an agent must read them carefully to choose correctly.
Most tools use a domain-prefix convention (claim_acquire, skill_list, reference_load, verify_yaml), but the hierarchy and detection tools break it by leading with the verb (create_hierarchy_plan, render_hierarchy_html, detect_technology_skills) and capture_repository_state stands alone. The mixture of prefix-first and verb-first orderings means there is no single predictable pattern, though each name is still readable.
At 26 tools the surface is heavy, and it spans six loosely related sub-domains (claims, skills, references, hierarchy plans, verification, capture) rather than one coherent domain. Many tools are legitimate, but the aggregate exceeds what an agent can hold comfortably and pushes past the 25-tool threshold.
The claim lifecycle is essentially complete (acquire, extend, extend_deadline, heartbeat, release, reset, status, report, journal maintenance), and skills cover list/find/read/load/resources/refresh/validate/detect. Minor gaps remain: hierarchy plans can be created, updated, and rendered but never listed, read back, or deleted, and reference handling lacks a validate counterpart to reference_refresh.
Maintenance
Related MCP Connectors
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
MCP server for agentverse documentation, generated by doc2mcp.
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Related MCP Servers
- AlicenseBqualityDmaintenanceProduction-grade MCP server that gives AI agents safe access to your local dev environment: filesystem, databases, processes, and OpenAPI specs.1540 npm4MIT
- FlicenseAqualityDmaintenanceLightweight MCP server that exposes tools for system information and weather lookup, designed for agent integration via stdio.1-
- AlicenseNot gradedqualityDmaintenanceA lightweight, stdio-based MCP server enabling AI assistants to perform local file system operations like reading, writing, searching, and executing commands.5,198 npmMIT
- AlicenseNot gradedqualityCmaintenanceLocal-first MCP server that provides project context, verification gates, and structured tools for coding agents to discover knowledge, run diagnostics, and execute allowlisted commands within a repository.40 npmMIT