Every coding agent starts where the last one stopped.
Recollect is a self-hosted Brain for your coding agents. It remembers what they learn, gives every session cited context, keeps your team’s skills in sync and decides which MCP tools an agent may run, and where.
Three coding hosts share one Brain. Memory flows in from sessions and back out as cited context; tools run where you placed them.
Memory with evidence
Decisions, fixes and handovers are learned from sessions and repositories, checked against their sources and recalled with citations.
Recall with evidenceDynamic skills
Publish a skill or shared instructions once. Every connected host loads the current revision natively, when it applies.
Shared skillsGoverned MCP tools
Tool groups with separate Use, Manage and Share rights decide who may run which tools, in which environment.
Tool groupsPrivate execution
A paired runner connects outbound and runs approved calls inside your network. It needs no inbound port.
Private runnersAgents that coordinate
Sessions on different machines announce their scope and planned files, and exchange advisory notes before their work collides.
Peer coordinationCode and knowledge graphs
Explore how repositories, configuration and decisions connect, with paths that cross repository boundaries.
Explore graphsOne loop on every prompt.
Install the plugin once. From then on, each session feeds the Brain and each new prompt starts with what the team already knows, under your capture, retention and model policies.
Sessions become evidence
The plugin captures permitted prompts, tool observations and imported documents through a durable delivery queue.
Evidence becomes memory
Claims, decisions, procedures and handovers pass a separate support check and keep their sources.
Memory returns as context
Exact, keyword, semantic and graph recall select what applies to the task’s scope, with citations.
Agents act on it
Skills load when needed and approved MCP calls run where you placed them. Outcomes become new evidence.
One loop runs on every prompt: capture, learn, recall, act. Every step keeps its evidence.
Publish a skill once. Every host has it next session.
Brain managers publish skills and shared instructions in the desktop app. The plugin installs them as native skills in Claude Code, Codex and OpenCode, reloads the host, and removes temporary copies when the task ends.
--- name: vault-inspection description: Inspect Vault configuration with attributable evidence. --- Read the repository instructions first. Separate declared configuration from observed runtime behavior.
.claude/skills/vault-inspection.agents/skills/vault-inspection.opencode/skills/vault-inspectionskills.list returns metadata · skills.fetch loads SKILL.md and references/evidence.md on demandPublish a skill once. The plugin keeps every connected host’s native skills in step with the Brain.
Brain-wide guidance applies everywhere; environment guidance adds to it only when that environment is selected.
Agents see compact metadata first and fetch a body or reference only when a skill applies.
Markdown and text only, with safe paths and secret detection. Guidance never grants a tool or overrides local rules.
Agents ask. Recollect checks. Tools run where you choose.
Connect an MCP server once, add it to a tool group and grant people Use. Reading a Brain never grants permission to run its tools, and administrators do not inherit Use.
- tool
- filesystem.read_file
- caller
- alex · OpenCode
- env
- Development
- result
- 3 files · captured
Every managed tool call passes the same checks, then runs on the executor you assigned to that connection.
Recollect service
Central HTTP connections such as Context7 or Exa. Recollect makes the MCP request after its checks.
Caller’s device
A local stdio server on the machine where the agent runs, started by the plugin’s optional runner.
Private runner
A named, paired device inside your network. It connects outbound, claims addressed calls and returns results.
The tools your agent sees.
The plugin and direct MCP clients reach the same scoped tools at /api/brains/<brain>/mcp/agent. Every call is bound to an immutable task scope, so a model cannot switch Brains or widen its own access.
workspace.*Scope every taskliststart_taskset_scopebegininspect_taskclose
memory.*Recall and contributerecallinspectcontributehandovergraph_exploregraph_path
skills.*Load shared guidancelistfetch
source.*Retain documentslistimportinspect
coordination.*Work alongside peersannouncepeersinboxsendackclose
mcp.*Run governed toolsprofilesdiscovercallstatuscancelreconcile
Running in three steps.
Everything runs on your machine or a private server: PostgreSQL with pgvector, Neo4j, and the Rust API and worker behind the desktop app.
- 1
Run Recollect
Needs Docker with Compose and Python 3.11 or newer. Then open
http://127.0.0.1:8787; the first run writes your login to.env.git clone https://github.com/MikeK184/Recollect.git cd Recollect ./scripts/stack.sh up --build - 2
Connect your agent
Open Agents → Connect agent. Build and install the plugin, then run the
recollect-connectskill and approve the device code in your browser.python3 scripts/package-plugin.py claude plugin marketplace add <package-directory> claude plugin install recollect-memory@recollect - 3
Work as usual
Start Codex, Claude Code or OpenCode normally. Cited context arrives before each prompt; add skills and tool groups whenever you need them.
Built for individuals and trusted internal teams. Workflows have local acceptance evidence; current work and limits stay visible. Diagrams use example data.