Local Agent ecosystem
Local Agent is the conversation, model, permission, and foreground-goal host. The surrounding projects remain focused tools with explicit interfaces. They are not all bundled integrations.
The shared transcript vocabulary keeps those ownership boundaries legible:
| Tool | Product role | Evidence meaning |
|---|---|---|
| MCPHub | Gateway and lazy catalog | Routing only |
| Cortex | Coordination, criteria, and evidence closure | Preserves claims and proof relationships |
| Vecgrep | Semantic discovery | Candidate |
| Codemap | Code structure and impact | Supported, or stale when its index reports drift |
| Glyphrun | CLI and TUI verification | Verified or contradicted from a versioned run |
| Cairntrace | Browser verification | Verified or contradicted from a versioned run |
| Vidtrace | Video evidence | Artifact-backed support |
| file.cheap | Artifact storage | Preserves evidence; does not create truth by itself |
| TinyVault | Secret boundary | Security state, never evidence content |
| Monitor | Ambient and requested observability | Operational support |
| Bob | Repository contract and convergence | No evidence upgrade; clean, drift, conflict, blocked, or failed domain state |
| VecLite | Vector storage | Internal capability, not an independent claim |
Local Agent shows these as semantic tool receipts in one chronological transcript. A green transport receipt is never substituted for a domain result or verification outcome.
Connected runtime
Cortex
An evidence-guided engineering kernel for durable task state, bounded changes, and verification. Local Agent can use Cortex as the optional semantic advisor for durable goals, directly or through MCPHub. Local acceptance IDs and statements cross that boundary as typed, immutable acceptanceCriteria; they are never flattened into goal prose.
Exact Cortex actions can become bounded next-step suggestions. Local Agent still owns tool resolution, schemas, effect classification, approval, budgets, loop prevention, scheduling, and cancellation; it does not reproduce Cortex's task lifecycle or evidence ledger.
MCPHub
A single configuration and gateway for discovering MCP servers and synchronizing them to agent harnesses. MCPHub is the recommended gateway for Local Agent; it does not replace Local Agent's final approval policy.
The MCPHub website and source repository document the gateway, synchronization, and deployment workflow.
Discovery and local data
Codemap
Structural code maps, references, impact, and index freshness. Local Agent treats fresh structural results as support, while missing or stale indexes remain visible instead of being interpreted as negative evidence.
Vecgrep
Local-first semantic code search with Ollama embeddings, hybrid search, and an MCP surface. It is MCP-capable and related to Local Agent; it is not a native embedded search engine inside Local Agent.
TinyVault
A local-first secrets manager and MCP server in a Go binary. It can be configured through MCP; it is not bundled.
file.cheap
A local-first CLI and MCP file vault for storing, finding, compressing, and comparing workflow artifacts. It can be configured through MCP; it is not Local Agent's session store.
VecLite
An embeddable Go vector database with single-file storage, HNSW, BM25, and hybrid search. It is related infrastructure; Local Agent's current ICE store does not embed VecLite.
Evidence and verification
Glyphrun
Black-box testing for CLI and TUI workflows over a real pseudo-terminal and deterministic emulator. Local Agent uses Glyphrun as test infrastructure for its terminal contracts.
Cairntrace
Local-first behavioral browser specs for coding agents, with repairable steps and evidence-backed outcomes. It is a related verifier, not bundled into Local Agent.
Vidtrace
A local-first Go CLI that turns bug recordings into frames, OCR, transcripts, and timestamped evidence timelines.
Operations and building
Monitor
A terminal system monitor for macOS and Linux with a TUI, JSON CLI, and MCP surface. It can be configured through MCP; it is not a built-in resource scheduler.
Bob
A deterministic, model-free repository factory that turns a bob.yaml contract into owned infrastructure and detects drift in CI. It complements Local Agent's coding workflow but is not embedded.
Bob is built to be driven by agents. Run bob learn --json once at the start of a session for a machine-readable onboarding brief: every command, the exit-code contract (0 success, 2 conflicts, 3 drift without conflicts, 4 invalid input; plan always exits 0), the stable data.error.code values, and the recipe catalog. Every --json failure envelope carries copy-pasteable corrective commands in next_actions, and every plan action carries a stable code field for branching without parsing prose.
Bob's files recipe lets an agent declare an arbitrary file tree in bob.yaml — paths, modes, contents, and literal ${vars.key} substitution — and scaffold or reconcile any repository through Bob's plan/apply/lock conflict safety; bob recipe show files --json returns the schema with an example.
When Bob is reached through an exact trusted MCP route, Local Agent parses bob_context, bob_path, and bob_playbook through their supported versioned schemas. Context reports bounded repository and capability state; path reports ownership and human-edit effects; playbook reports availability, blockers, required inputs, scope, risk, and bounded steps. A lookalike route, malformed document, future schema, or unknown closed value fails closed.
Every Bob result carries EvidenceNone: no evidence assessment. Bob can prove whether repository files converge on its contract, but not whether application behavior is correct. A Bob clean state is therefore visually and semantically distinct from verified evidence or Cortex task completion.
Raw Bob structured output stays inside the parser boundary. Saved session state contains only a bounded semantic digest; absolute workspace paths, arbitrary reasons or commands, user values, manifests, previews, and file contents are not persisted. A smaller validated projection may be supplied transiently to the active model turn. Direct Bob calls and complete MCPHub stored results use the same parser after the gateway pages have been bounded and matched to the exact original call ID and route.
At session work time, a regular root bob.yaml is only a cheap candidate signal. If one unambiguous eligible registered bob_context read is available, Local Agent can suggest the compact read and cache its bounded digest. A unique direct route is preferred; otherwise exactly one pinned MCPHub route is required. With explicitly configured auto_read_only and an AUTO turn, it may perform the read only if the current host policy proves the target is read-only, non-destructive, idempotent, closed-world, workspace-bound, and not denied. The automatic continuation budget never exceeds two calls.
Bob and Cortex actions share a bounded host projection, not product authority. In the default suggestion mode, Local Agent exposes a valid next tool, named missing inputs, and blockers without executing it. Optional read-only auto-follow revalidates the route, current tool schema, workspace, effect, approval policy, staleness, and replay fingerprint before each dispatch. Shell, mutation, secreted, unresolved proxy, and prose-derived actions remain gated or unsupported.
The Bob website and source repository document its repository contracts and build workflow; the agent guide is the onboarding reference for coding agents.
Relationship rule
“MCP-capable” means a tool can be exposed through a configured MCP server. It does not mean Local Agent ships, authenticates, or silently trusts that tool. Every configured MCP call still crosses the Local Agent approval boundary.